Cuánto cuesta un millón de tokens, y por qué el techo y el piso se mueven en direcciones opuestas

Historial de cambios

1. Edición 26 · 2026-11-11

   The Wednesday volumes/enterprise/capex sweep found OpenRouter's biggest month yet (790T tokens in October, up 37% on September) and a sharp swing in Ramp's weekly token-spend data, where Anthropic's share jumped past 60% as the blended OpenAI+Anthropic price kept falling. Capex guidance from the big five cloud spenders and Anthropic's own disclosed revenue figures held unchanged since last check, so those rows were left as-is rather than re-stated without new evidence.

   • OpenRouter processed 790T tokens in October 2026, its largest month ever, up 37% from September and roughly 25x its January volume; agentic traffic reached 76% of all tokens (from 8% a year earlier) and cache-hit rates on agentic prompts passed 90%, up from 76.1% in June (https://openrouter.ai/data/october-2026).
   • On Ramp's corporate-card data, Anthropic's share of weekly AI token spend jumped to 60.5% for the week ending 23 Oct 2026 from 51.3% a month earlier, while OpenAI's fell to 35.1% from 44.2%; the blended price per million tokens across the two providers fell to $0.56 from $0.76, -26.8% (https://ramp.com/data/ai-index).
   • Alphabet's first-party model APIs processed about 32 billion tokens per minute in Q3 2026, up 40% from Q2's 22B and more than 4x the ~7B a year earlier -- an API-traffic-only figure, not directly comparable to the 3.2-quadrillion-tokens/month "across all surfaces" number Google gave at I/O in May 2026 (https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q3-2026/).

2. Edición 25 · 2026-11-10

   The Tuesday Artificial Analysis sweep turned up two frontier launches missing from the store: Anthropic's Claude Fable 5.2 (released 9 Nov 2026, cheaper than Fable 5.1 at $8/$40 vs $10/$50) and Google's restricted-access Gemini 4 Carbon (28 Oct 2026, cybersecurity-focused, not yet publicly available). Re-checking each open-vs-closed row against its own Artificial Analysis model page also found a deprecated DeepSeek slug replaced by a newer MIT-licensed release, and several Intelligence Index scores ticked up by one point from ordinary AA re-evaluation.

   • Anthropic released Claude Fable 5.2 on 9 Nov 2026 at $8/$40 per million input/output tokens (down from Fable 5.1's $10/$50), confirmed on Anthropic's own models-overview page (model ID claude-fable-5-2); it now holds the premium tier-ladder slot and scores 65 on Artificial Analysis's Intelligence Index at $26.54/task (https://platform.claude.com/docs/en/about-claude/models/overview; https://artificialanalysis.ai/models/claude-fable-5-2).
   • Google launched Gemini 4 Carbon on 28 Oct 2026, a cybersecurity-focused frontier model priced at $2/$10 per million tokens introductory ($4/$20 after), but access is restricted at launch to cyber defenders and Fairwind Program partners with no public API yet; Artificial Analysis still benchmarked it independently at Intelligence Index 60, $7.76/task (https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/; https://artificialanalysis.ai/models/gemini-4-carbon).
   • DeepSeek's open-vs-closed row moved from the now-deprecated 'DeepSeek V4.1 Flash' Artificial Analysis listing to 'DeepSeek V4.1 Flash 1017', an MIT-licensed open-weight release from 17 Oct 2026, unchanged at index 33 and $0.60/task (https://artificialanalysis.ai/models/deepseek-v4-1-flash-1017).
   • Re-verified each active open-vs-closed row against its own Artificial Analysis model page on 10 Nov 2026: Claude Opus 5.5 (max) and Claude Sonnet 5.5 (max) each gained one Intelligence Index point (59 and 55, both unchanged cost per task) from Artificial Analysis's ongoing re-evaluation; Muse Wonder (max) likewise moved from index 43 to 44; Gemini 4 Argon (high) held its index 50 but its cost per task rose from $7.68 to $9.95.

3. Edición 24 · 2026-11-09

   The Monday frontier/ladder sweep found OpenAI has filled the gap in its tier ladder: GPT-6.1 Terra, priced between Sol and Luna, is now live on OpenAI's own pricing page. Every other tracked frontier and tier-ladder price across OpenAI, Anthropic, Google, xAI, DeepSeek, Mistral and Meta was re-checked against its primary source today and held unchanged.

   • OpenAI added a new tier-ladder model, GPT-6.1 Terra, at $1.00/$5.00 per million input/output tokens, between GPT-6.1 Sol ($2/$10) and GPT-6 Luna ($0.10/$0.50) in OpenAI's flagship-models pricing table (https://developers.openai.com/api/docs/pricing, checked 9 Nov 2026).
   • Re-verified unchanged on 9 Nov 2026 against each provider's own pricing page: OpenAI GPT-6 Astra $10/$50, GPT-6.1 Sol $2/$10, GPT-6 Luna $0.10/$0.50 (developers.openai.com/api/docs/pricing); Anthropic Claude Fable 5.1 $10/$50, Opus 5.5 $4/$20, Sonnet 5.5 $2/$10, Haiku 5.5 $0.10/$0.50 (claude.com/pricing); Google Gemini 4 Argon $2/$10 (through 30 Nov 2026), Gemini 3.1 Pro $2/$12, Gemini 3.8 Flash $0.75/$3.75 intro, Gemini 3.5 Flash-Lite $0.30/$2.50 (ai.google.dev/gemini-api/docs/pricing); xAI Grok 4.7 $2/$6 (docs.x.ai/docs/models); DeepSeek V4 Pro peak $1.32/$3.96, V4.1 Flash peak $0.30/$1.20 (api-docs.deepseek.com/quick_start/pricing); Mistral Large 4 $0.70/$2.10, Small 4 $0.15/$0.60 (docs.mistral.ai/inference/pricing); Meta Muse Wonder $4/$20 (confirmed via third-party aggregators; Meta's own pricing page returned 404 at every URL tried today).

4. Edición 23 · 2026-11-05

   Anthropic's amended S-1 (filed 2 Nov 2026 ahead of its Nasdaq listing as ANT) disclosed audited-grade quarterly financials for the first time: Q3 2026 revenue of $17.0B against a $567M operating loss, nine-month revenue of $33.2B, and $895.4B in total compute and infrastructure commitments. The Wednesday enterprise/volumes/capex sweep found no other reader-visible change: OpenRouter's and Ramp's latest published figures are unchanged from the prior check.

   • Anthropic's S-1/A (filed 2 Nov 2026, Nasdaq: ANT) gives quarterly revenue for the first time: Q1 2026 $4.73B, Q2 2026 $11.56B (operating income $1.0B), Q3 2026 $17.0B (operating loss $567M, driven by $5.0B of compute under short-term commitments and $2.0B of non-cash charges); nine months ended 30 Sep 2026 revenue $33.2B (+1,206% YoY), net loss $104.0B (includes $94.3B non-cash fair-value change on convertible notes/SAFEs). FY2023-2025 revenue was $41M/$386M/$4.6B. The filing defines 'Run-Rate Revenue' (trailing-28-day consumption x13 + subscription x12) but leaves the as-of-31-Oct-2026 figure blank (https://www.sec.gov/Archives/edgar/data/2130357/000162828026068676/anthropicpbc-sx1ax1.htm).
   • The same filing discloses $895.4B in aggregate compute and infrastructure purchase/lease commitments as of 30 Sep 2026 (up from the ~$518B reported from the confidential draft in September), $447.5B of it due before 2031; customer concentration has fallen, with no single customer at 10%+ of nine-month 2026 revenue versus two customers at 12% each in FY2025 (https://www.sec.gov/Archives/edgar/data/2130357/000162828026068676/anthropicpbc-sx1ax1.htm).
   • Re-verified and unchanged: OpenRouter's weekly token volume is still its last published figure, 128.4 trillion for the week of 14 Sep 2026, with no newer newsletter edition since 23 Oct 2026 (https://openrouter.ai/data/newsletter/the-token-state-of-play); Ramp's AI Index is still its September 2026 edition, 55.9% overall paid-AI adoption (Anthropic 43.8%, OpenAI 40.6%), with no November edition yet published (https://ramp.com/data/ai-index).
   • Checked and not added: Google's Q3 2026 earnings call cited first-party model APIs processing nearly 32B tokens/minute (+40% QoQ, 4x YoY), but this is an API-only rate, not the same cross-surface monthly-tokens metric (last given as 3.2 quadrillion/month at I/O 2026) tracked in tokenVolumes.googleMonthlyTrillions; adding a value derived from the API-only rate to that series would conflate two different metrics, so no new row was added.

5. Edición 22 · 2026-11-04

   Big tech's Q3/fiscal-Q1 2027 earnings calls (28-29 Oct 2026) moved the capex picture: Microsoft cut its calendar-2026 capital spending guide to ~$170B, down from the ~$175B it had carried since July, citing delayed third-party datacenter capacity. Alphabet and Amazon held their full-year guides (~$200B midpoint and ~$220B respectively) steady while posting sharply higher Q3 actual capex, and Alphabet said 2027 spending will rise significantly further.

   • Microsoft now expects calendar-2026 capex of approximately $170 billion, down from the ~$175B figure restated in July, per CFO Amy Hood on the 28 Oct 2026 Q1 FY27 earnings call: "spend should be approximately $170 billion due to the 3P capacity delays" (https://www.microsoft.com/en-us/Investor/events/FY-2027/earnings-fy-2027-q1.aspx). Q1 FY27 (quarter to Sep 2026) capex was $39.4B in additions to property and equipment (https://www.sec.gov/Archives/edgar/data/0000789019/000119312526436122/msft-ex99_1.htm).
   • Alphabet reaffirmed its $195-205B full-year 2026 capex guidance unchanged on its 28 Oct 2026 Q3 call, and said capex is expected to increase significantly in 2027. Q3 2026 actual capex was $51.355B, YTD $131.953B (https://www.sec.gov/Archives/edgar/data/0001652044/000165204426000080/googexhibit991q32026.htm).
   • Amazon reaffirmed its ~$220B 2026 cash capex guidance with its Q3 2026 release (29 Oct 2026): "We continue to expect 2026 cash capital expenditures to be approximately $220 billion." Q3 2026 actual capex was $77.259B, more than double the $35.095B a year earlier, attributed to AI investment (https://www.sec.gov/Archives/edgar/data/0001018724/000101872426000038/amzn-20260930xex991.htm).
   • hostedOpen: Decart, the cheapest OpenRouter host for DeepSeek V4.1 Flash, has been delisted from the model's endpoint list; Sail Research is now cheapest by output price ($0.09/$0.25, 40 endpoints as of 4 Nov 2026, live-checked via OpenRouter's /endpoints API).

6. Edición 21 · 2026-11-03

   Artificial Analysis moved its Intelligence Index to v5.0 in October 2026, re-weighting and swapping several of the nine benchmarks that make up the score, which re-ranks almost every model in the open-vs-closed comparison and raises most cost-per-task figures. Separately, Mistral Large 4's weights are now live on Hugging Face (open-weight), while Kimi K3 has been deprecated by Kimi/Moonshot with no cost or index figure left to report.

   • Artificial Analysis's Intelligence Index moved to v5.0 in October 2026: added HLE-Diamond and Terminal-Bench-Science, replaced CritPt with CritPt v2, removed GDPval-AA, SciCode, AA-LCR and HLE, and rebalanced categories to 25% each (artificialanalysis.ai/methodology/intelligence-benchmarking). Re-verified cost-per-task and index for all 20 models in the open-vs-closed table against their individual Artificial Analysis model pages on 3 Nov 2026.
   • Mistral Large 4's weights (mistralai/Mistral-Large-4-1T-A52B) are confirmed live on Hugging Face under the Mistral Ecosystem Open-Weight License; the open-vs-closed row moves from closed to open-weight at $2.06/task, index 26 (artificialanalysis.ai/models/mistral-large-4, 3 Nov 2026).
   • Kimi K3 is deprecated: Artificial Analysis's page for it now shows no public API and no cost-per-task or index figure, recommending Kimi K3.1 Pro instead; K3.1 Pro remains proprietary with no open-weight release announced, so no replacement open-weight row was added (artificialanalysis.ai/models/kimi-k3, 3 Nov 2026).

7. Edición 20 · 2026-11-02

   Monday's frontier/ladder sweep found one real price move: Mistral cut Mistral Large 4's list price roughly in half, to $0.70/$2.10 per million tokens, while quietly deprecating Mistral Medium 3.5 (its own docs now point Medium 3.5 users to Large 4 for new integrations). Every other tracked flagship price across OpenAI, Anthropic, Google, xAI, DeepSeek and Meta was re-verified live and is unchanged since last check. Separately, today's queue drift item on gpt-oss-120b's cheapest OpenRouter host was resolved: PrimeIntellect now edges out Venice/Near AI on output price.

   • Mistral Large 4's list price dropped from $1.36/$4.18 to $0.70/$2.10 per million tokens (cached input $0.07), confirmed live on docs.mistral.ai/inference/pricing on 2 Nov 2026; Mistral Medium 3.5 is now marked deprecated (deprecation date 10/31/2026) on its own docs page, which names Mistral Large 4 as the replacement, and no longer appears in the generalist-models pricing table (https://docs.mistral.ai/inference/pricing; https://docs.mistral.ai/models/mistral-medium-3-5-26-04).
   • Mistral Large 4's open weights (mistralai/Mistral-Large-4-1T-A52B, ~1T params/52B active, Mistral Ecosystem Open-Weight license) went live on Hugging Face on 29 Oct 2026, three weeks after the 6 Oct API-only preview; this resolves the open-weights question flagged in the 26 Oct frontier row (secondary coverage, not yet checked against Hugging Face directly: https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B).
   • Re-verified live and unchanged since the last check (26 Oct 2026): OpenAI's GPT-6 Astra/GPT-6.1 Sol/GPT-6 Luna tier prices and cache ratios (developers.openai.com/api/docs/pricing); Anthropic's Claude Fable 5.1/Opus 5.5/Sonnet 5.5/Haiku 5.5 tier prices and cache-read prices (claude.com/pricing); Google's Gemini 4 Argon, Gemini 3.1 Pro, Gemini 3.8 Flash and Gemini 3.5 Flash-Lite tier prices and Gemini 4 Argon's cache price (ai.google.dev/gemini-api/docs/pricing); xAI's Grok 4.7 tier and cached-input price (docs.x.ai/docs/models); DeepSeek's V4 Pro and V4.1 Flash tier prices (api-docs.deepseek.com/quick_start/pricing); and Meta's Muse Wonder tier price and cache ratio (via Meta's own developer pricing docs, cross-checked against aggregators since the direct fetch 404'd).
   • Resolved the open hostedOpen drift item for gpt-oss-120b: OpenRouter's cheapest host by output price (the collector's sort key) is now PrimeIntellect at $0.033/$0.122, not Venice/Near AI at $0.03/$0.15 as previously recorded; quarantined the stale row and added a PrimeIntellect row (https://openrouter.ai/api/v1/models/openai/gpt-oss-120b/endpoints).

8. Edición 19 · 2026-10-31

   Saturday's METR/Epoch AI/benchmark sweep found one price-to-capability row understated the price drop: the cheapest model to clear 87.5% on ARC-AGI-1 in August 2026 was DeepSeek V4 Flash at $0.02/task, not Grok 4.6 at $0.30/task, so the 20-month price fall for that capability bar is about 228,000x, not 15,200x. Epoch AI's MMLU, GPQA Diamond and FrontierMath price-to-capability figures, METR's horizon data, and the harness-tax and cache-hit-rate studies were all re-checked against their primary sources and are unchanged since last review.

   • ARC-AGI-1's 87.5% threshold was cleared more cheaply than previously recorded: DeepSeek V4 Flash scored 89.0% on ARC-AGI-1 in August 2026 at $0.02/task, versus the $0.30/task previously logged for Grok 4.6 (87.5%, same month); against Dec 2024's o3 baseline ($4,560/task), that is a ~228,000x price fall over 20 months, not 15,200x (https://r40.io/blog/arc-agi-verified-cost-per-task-price-step/).
   • Re-verified against Epoch AI's own pages: the GPT-4-level MMLU threshold ($0.18/M, Gemini 2.0 Flash, Feb 2025), the 75% GPQA Diamond threshold ($0.0004/question, GPT-5.6 Luna), and the ~27% FrontierMath threshold (~5M tokens, GPT-5.2, Dec 2025) all remain the cheapest confirmed points on Epoch's own pages as of 31 Oct 2026 (https://epoch.ai/data-insights/llm-inference-price-trends, https://epoch.ai/publications/the-plunging-price-of-thought, https://epoch.ai/gradient-updates/how-persistent-is-the-inference-cost-burden).

9. Edición 18 · 2026-10-28

   Wednesday's volumes/enterprise/capex sweep found no new capex figures — Microsoft and Alphabet haven't yet filed their next quarterly results (Amazon reports tomorrow) — but OpenRouter's two newest newsletter editions add real adoption data: large model portfolios keep growing without concentration easing, and almost a quarter of OpenRouter traffic still runs on a superseded in-family model version, leaving real savings on the table. No provider's token volumes or capex guidance moved this week.

   • OpenRouter's 'Model portfolios' newsletter (15 Oct 2026, data Oct 2025-Sep 2026) finds both AI-native startups and large orgs now use ~12 models/month on average (up from 9 a year ago), but the top model still takes 68% of startup requests and 66% of large-org requests; the share of portfolio switches that move to a different lab rose from about a third to about three-quarters over the year (https://openrouter.ai/data/newsletter/model-portfolios).
   • OpenRouter's 'Version lag' newsletter (23 Oct 2026, data 15-21 Oct 2026) finds 24% of requests (23% of spend) still ran an older in-family model version that week, with the older-version share ranging from 17% (Claude Fable) to 70% (Grok); upgrading at list prices would save up to 90% on Claude Haiku traffic, and OpenRouter estimates $229,000/day in reachable platform-wide savings from routing to the latest in-family model (https://openrouter.ai/data/newsletter/version-lag).
   • Checked for new capex and token-volume data: Microsoft's Q1 FY2027 and Alphabet's Q3 2026 results have not yet been filed with the SEC as of 28 Oct 2026 (Alphabet reports after today's close; Amazon reports 29 Oct); Meta's capex figure already reflects its 21 Oct Q3 release. Google's last disclosed monthly token-processing figure remains 3,200 trillion (I/O, May 2026) with no newer figure found. No change to tokenVolumes, capex.companies or cacheEconomics this edition.

10. Edición 17 · 2026-10-27

    Tuesday's Artificial Analysis sweep found real cost-per-task moves on three open-weight models rather than index changes: Beam 1.0 nearly halved as its hosts cut prices further, Mistral Large 4's cost jumped now that its launch sale has fully dropped out of Artificial Analysis's calculation, and Muse Wonder (max) fell. Three more models' cheapest OpenRouter host changed hands, continuing the churn tracked since 18 Oct; no new open-weight model launched this week and Kimi K3.1 Pro's promised weights (due 4 Nov) have not yet appeared.

    • Beam 1.0's cost per task fell from $0.23 to $0.12 (index unchanged at 30) as OpenRouter hosts repriced; cheapest host moved from Io Net (delisted) to Nebius at $0.25/$1.00, 27 Oct 2026 (https://artificialanalysis.ai/models/reflection-beam; https://openrouter.ai/api/v1/models/reflection/beam-1.0-501b-a23b/endpoints).
    • Mistral Large 4's cost per task rose from $1.13 to $1.61 (index unchanged at 38) as Artificial Analysis's cost calculation moved off the ended two-week launch sale onto Mistral's unchanged list price ($1.36/$4.18); weights remain unpublished and proprietary per Artificial Analysis, 27 Oct 2026 (https://artificialanalysis.ai/models/mistral-large-4).
    • Muse Wonder (max)'s cost per task fell from $4.54 to $3.27 (index unchanged at 52); Meta's API price of $4/$20 per million tokens is unchanged, 27 Oct 2026 (https://artificialanalysis.ai/models/muse-wonder).
    • DeepSeek V4.1 Flash's cheapest OpenRouter host moved to Decart at $0.09/$0.18; OpenRouter has folded the 1017-checkpoint listing (tracked separately in this report since 20 Oct) into the single deepseek/deepseek-v4.1-flash model id, 27 Oct 2026 (https://openrouter.ai/api/v1/models/deepseek/deepseek-v4.1-flash/endpoints).
    • Nemotron 3.5 Super VL's cheapest OpenRouter host moved from Crusoe to Darkbloom at $0.06/$0.35, 27 Oct 2026 (https://openrouter.ai/api/v1/models/nvidia/nemotron-3.5-super-vl/endpoints).
    • GLM-5.3 Flash's cheapest OpenRouter host moved from StreamLake (now $0.0885/$0.295) to Relace at $0.0285/$0.0923; Z.AI's own modal price of $0.15/$0.5 across 38 endpoints is unchanged, 27 Oct 2026 (https://openrouter.ai/api/v1/models/z-ai/glm-5.3-flash/endpoints).
    • Re-verified against Artificial Analysis's leaderboard with no change: DeepSeek V4.1 Flash ($0.26, index 45), MiniMax M3.1 Flash ($0.22, index 43), Kimi K3.1 Pro ($2.11, index 55, open weights still due 4 Nov 2026), GLM-5.3 ($2.01, index 45), Qwen3.8 Max ($5.41, index 45), Step 5 Preview ($1.03, index 44) and Nemotron 3.5 Super VL ($0.19, index 26), 27 Oct 2026 (https://artificialanalysis.ai/leaderboards/models).

11. Edición 16 · 2026-10-26

    Monday's frontier sweep found two stale rows in the frontier chart rather than a price move: Gemini 4 Argon's point had been carrying the post-1-December doubled price instead of the current intro price its own label names, and Mistral Large 4's point still described a launch sale that ended three days ago. Both are corrected; no provider's list prices actually moved this week.

    • Corrected Gemini 4 Argon's frontier point from $4/$20 to $2/$10 input/output per million tokens: that was always the current intro price (the label already said so), the $4/$20 figure is the price that takes effect 1 Dec 2026, re-verified live 26 Oct 2026 (https://ai.google.dev/gemini-api/docs/pricing).
    • Relabeled Mistral Large 4's frontier point to drop the 'launch sale $0.68/$2.09' note, since that two-week 50% sale ended 20 Oct 2026 (edition 15); the point's price was already the correct list $1.36/$4.18. Open weights, promised for 'end of October,' had not appeared on Hugging Face as of 26 Oct 2026; contemporary coverage points to 27 or 31 Oct (https://docs.mistral.ai/inference/pricing).
    • Re-verified live against each provider's own pricing page with no other change: OpenAI (GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna), Anthropic (Claude Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5), Google (Gemini 3.1 Pro, 3.8 Flash, 3.5 Flash-Lite), xAI (Grok 4.7), DeepSeek (V4 Pro, V4.1 Flash, peak and off-peak), Meta (Muse Wonder) and Mistral (Medium 3.5, Small 4) frontier and tier-ladder prices, and every tracked cache-read ratio, are all unchanged as of 26 Oct 2026.

12. Edición 15 · 2026-10-23

    Kimi K3.1 Pro (23 Oct) scores 55 on Artificial Analysis's index for $2.11 per task, two points above GPT-6 Astra, with weights promised by 4 November; when they land, the best open-weights score jumps. MiniMax published the weights of M3.1 Flash (index 43, $0.22 per task), Beam 1.0's hosts cut prices, and Mistral Large 4's launch sale ended. Meta narrowed its 2026 capex range upward to $135-145B with Q3 results; frontier list prices at OpenAI, Anthropic, Google, xAI, DeepSeek and Meta are unchanged.

    • Kimi K3.1 Pro, released 23 Oct 2026 at $2/$10 per million tokens, scores 55 at $2.11 per task on Artificial Analysis; Moonshot says the full weights will be released by 4 Nov 2026 (https://artificialanalysis.ai/models/kimi-k3-1-pro; https://www.kimi.ai/blog/kimi-k3-1-pro).
    • MiniMax M3.1 Flash (19 Oct 2026, $0.15/$0.60 official) scores 43 at $0.22 per task; weights on Hugging Face 23 Oct; cheapest of 9 OpenRouter endpoints InferenceNet at $0.10/$0.55 (https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash; https://openrouter.ai/api/v1/models/minimax/minimax-m3.1-flash/endpoints).
    • Meta narrowed 2026 capex guidance to $135-145B from $130-145B on 21 Oct 2026 (midpoint 137.5 to 140); Q3 capex $34.91B (https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-Third-Quarter-2026-Results/default.aspx).
    • Mistral Large 4's two-week 50% launch sale (from 6 Oct 2026) has ended; the ladder's Mistral flagship is back at list $1.36/$4.18 (https://docs.mistral.ai/inference/pricing).
    • Beam 1.0's cost per task on Artificial Analysis fell from $0.57 to $0.23 (index unchanged at 30) as its hosts cut prices: Io Net is now cheapest at $0.24/$0.98, CoreWeave cut to $0.30/$1 (20 Oct) and Baseten to $0.40/$1.20 (21 Oct 2026) (https://artificialanalysis.ai/leaderboards/models; https://openrouter.ai/api/v1/models/reflection/beam-1.0-501b-a23b/endpoints).
    • Muse Wonder (max) cost per task moved from $4.52 to $4.54, index 52, as of 23 Oct 2026 (https://artificialanalysis.ai/models/muse-wonder).
    • DeepInfra raised Nemotron 3.5 Super VL to $0.15/$0.50 from $0.10/$0.40 (first seen 20 Oct 2026); Crusoe stays cheapest at $0.10/$0.40 (https://openrouter.ai/api/v1/models/nvidia/nemotron-3.5-super-vl/endpoints).
    • Gemini 4 Argon's cache-read ratio added: $0.10 against $2 input = 0.05, implicit caching only, unchanged when both double on 1 Dec 2026 (https://ai.google.dev/gemini-api/docs/pricing).
    • OpenAI added 12-hour prompt caching to the Responses API on 20 Oct 2026: writes cost 2.5x the input price, reads bill at the cached-input price (https://developers.openai.com/api/docs/changelog).

13. Edición 14 · 2026-10-21

    OpenRouter's cheapest DeepSeek V4.1 Flash host moved again, and this month's Ramp AI Index shows OpenAI closing on Anthropic in business adoption alongside a record level of providers-switching. Both are dated, primary-sourced updates rather than a shift in the broader picture.

    • DeepSeek V4.1 Flash's cheapest host on OpenRouter moved from OpenInference to Sail Research ($0.101/$0.15, output-sorted); OpenInference itself repriced to $0.049/$0.375 (from $0.0352/$0.0112 last edition), across 38 endpoints as of 21 Oct 2026 (https://openrouter.ai/api/v1/models/deepseek/deepseek-v4.1-flash/endpoints).
    • Ramp's October AI Index ('Cracks in the AI Thesis, Part 3', 14 Oct 2026) puts September paid-AI adoption at 55.9% of Ramp businesses (+0.62pp from August); Anthropic leads on provider share at 43.8% (+0.25pp) to OpenAI's 40.6% (+1.12pp), narrowing the gap (https://ramp.com/data/ai-index-oct-2026).
    • The same Ramp edition reports a record 8% of businesses switched AI model providers in September, the highest share yet, which Ramp's economist reads alongside slowing new-customer growth as a sign adoption may be nearing a plateau (https://ramp.com/data/ai-index-oct-2026).

14. Edición 13 · 2026-10-20

    DeepSeek quietly replaced V4.1 Flash with a new checkpoint on 17 October, and it closes most of the gap to the top of the open-weight ladder: Artificial Analysis's Intelligence Index for the model jumps from 39 to 45, tied with GLM-5.3 and Qwen3.8 Max and one point behind MiMo-V2.6-Pro, at essentially the same cost per task. The official API price is unchanged. Separately, today's drift check found the cheapest OpenRouter host for the prior build had moved from Baidu to OpenInference since last edition.

    • DeepSeek-V4.1-Flash-1017, a 17 Oct 2026 checkpoint update, now serves DeepSeek's deepseek-flash API model and the OpenRouter model 'DeepSeek V4.1 Flash' on Artificial Analysis's leaderboard; openVsClosed index rises 39 to 45, cost per task $0.27 to $0.26. DeepSeek's own release notes call it a minor update with no training-data or architecture change, citing gains on Terminal-Bench 4.0 (31.2 to 42.3) and HLE without tools (36.8 to 42.6) (https://api-docs.deepseek.com/news/news261017; https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017; https://artificialanalysis.ai/leaderboards/models).
    • DeepSeek's official deepseek-flash price is unchanged at $0.3/$1.2 peak, $0.15/$0.6 off-peak per million tokens; the 1017 checkpoint is also live on OpenRouter under its own model id with 20 endpoints, cheapest at Relace ($0.05/$0.024), added as a new hostedOpen row (https://api-docs.deepseek.com/quick_start/pricing; https://openrouter.ai/api/v1/models/deepseek/deepseek-v4.1-flash-1017/endpoints).
    • The prior DeepSeek V4.1 Flash build (OpenRouter's default deepseek/deepseek-v4.1-flash, canonical slug 20260910) is still served separately; its cheapest OpenRouter host moved from Baidu ($0.0357/$0.1428) to OpenInference fp4 ($0.0352/$0.0112) across 36 endpoints as of 20 Oct 2026, correcting last edition's row (https://openrouter.ai/api/v1/models/deepseek/deepseek-v4.1-flash/endpoints).
    • Checked and not added: K2 Horizon 375B A23B (MBZUAI Institute of Foundation Models, Apache 2.0, index 31, released on or before 28 Sep 2026, no commercial hosted price found) and Inkling (Thinking Machines, released 15 Jul 2026 at index 41 per contemporary coverage but shown at 25 on today's Artificial Analysis leaderboard, now hosted on OpenRouter at $1/$4.05) both predate this edition's window; Inkling's index discrepancy is unresolved and needs a primary-source check before it's added (https://artificialanalysis.ai/models/open-source; https://huggingface.co/IFM/K2-Horizon-375B-A23B; https://openrouter.ai/api/v1/models/thinkingmachines/inkling/endpoints).

15. Edición 12 · 2026-10-17

    The open-weight floor stopped being a Chinese-only story this week, at least in headcount. Meta published the weights of Muse Spark Open (816B parameters, 42B active) on 15 October, Reflection shipped Beam 1.0 (501B, 23B active) on the 16th, and NVIDIA released Nemotron 3.5 Super VL (120B, 12B active) the same day. On Artificial Analysis's index they score 39, 30 and 26, which puts the best US open model six points below GLM-5.3 and seven below MiMo-V2.6-Pro, and all of them below the Step 5 Preview weights that landed on the 15th at 44. The floor argument survives unchanged: Muse Spark Open sells at an identical $1.25/$4.25 from five hosts on day one, Beam from $0.25/$1 and Nemotron from $0.10/$0.40, so the new entrants priced themselves straight into the commodity band. Four rows added to the open-versus-closed chart and three to the hosted floor; nothing at the ceiling moved.

    • Muse Spark Open: Meta's first frontier-class open weights, meta-models/Muse-Spark-Open-816B-A42B on Hugging Face (15 Oct 2026, licence listed as 'other'); Artificial Analysis Intelligence Index 39 at $0.84 per task; hosted at $1.25/$4.25 by Together, Nebius, Fireworks, Baseten and Modal, the same price Meta charges for the closed Muse Spark 1.3 (https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B; https://artificialanalysis.ai/leaderboards/models; https://openrouter.ai/api/v1/models/meta/muse-spark-open/endpoints).
    • Beam 1.0 (Reflection): 501B total, 23B active, 1M context, on OpenRouter 16 Oct 2026 from $0.25/$1 (Nebius, FP8) to $1.30/$4.30 (Baseten); Artificial Analysis index 30 at $0.57 per task, second among US open models behind Muse Spark Open, using about 26k output tokens per index task against about 52k for K2 Horizon and about 27k for GPT-6 Astra; 3.0% on Terminal-Bench 4.0 and a 47% hallucination rate on AA-Omniscience. Reflection had promised Apache 2.0 weights; the licence was not confirmed on a primary page this edition (https://x.com/ArtificialAnlys/status/2111087606759030831; https://openrouter.ai/api/v1/models/reflection/beam-1.0-501b-a23b/endpoints).
    • Nemotron 3.5 Super VL (NVIDIA): 120B total, 12B active, nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B on Hugging Face (16 Oct 2026, licence listed as other); index 26 at $0.19 per task; $0.10/$0.40 at Crusoe and DeepInfra with a free endpoint on OpenRouter (https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16; https://openrouter.ai/api/v1/models/nvidia/nemotron-3.5-super-vl/endpoints).
    • Step 5 Preview now appears on Artificial Analysis's open-weights list at index 44, third among open models behind MiMo-V2.6-Pro (46) and GLM-5.3 (45), at $1.03 per task; added to openVsClosed and the hosted row updated with the four hosts at $1/$2.70 (https://artificialanalysis.ai/models/open-source).
    • Reading for the thesis: the top of the open ladder is still Chinese (MiMo 46, GLM-5.3 45, Step 5 44, Kimi K3 44, GLM-5.3 Flash 42) and the Western entrants sit at 39 and below, so the essay's claim that the open frontier is in Hangzhou and Beijing stands; what changed is that Meta, Reflection and NVIDIA now price open weights into the same $0.10 to $1.25 hosted band on launch day, which is the floor doing what the essay said it would. Ceiling unchanged: no provider moved a frontier or flagship list price this week.

16. Edición 11 · 2026-10-16

    A new series and chart, no change to the existing ones. 'What a cached token costs' tracks the cache-read price as a share of the input price for each provider's flagship-or-above model, from the first prompt-caching launches in August 2024 (Anthropic and DeepSeek at one tenth, OpenAI at one half two months later) to today's one twentieth at OpenAI, Anthropic and Meta and one fortieth at Fable 5.1. Beside it sit the two numbers that explain why list prices can stand still while the bill falls: the share of agentic tokens served from cache on OpenRouter (about 60% in January, just under 90% in September) and the effective price paid per million tokens on Ramp cards ($1.15 in March, $0.68 in August). This is the third speed in the two-speeds story and the subject of the companion essay on the harness tax.

    • cacheEconomics.readRatio added with fifteen dated points: Anthropic 0.10 (Claude 3.5 Sonnet, Aug 2024; Opus 4.5, Nov 2025) to 0.05 (Opus 5.5) and 0.025 (Fable 5.1) in Sep 2026; OpenAI 0.50 (GPT-4o, Oct 2024), 0.25 (GPT-4.1, Apr 2025), 0.10 (GPT-5, Aug 2025), 0.05 (GPT-6.1 Sol, Sep 2026); Google 0.25 (Gemini 2.5 implicit caching, May 2025) to 0.10 (Gemini 3.1 Pro, Feb 2026); DeepSeek 0.10 (V2 context caching, Aug 2024) to 0.02 (V4.1 Flash, Sep 2026); xAI 0.25 (Grok 4.7); Mistral 0.10 (Large 4); Meta 0.05 (Muse Wonder). Current ratios read on each provider's pricing page on 16 Oct 2026; the 2024 and 2025 launch ratios are the launch prices of each provider's caching feature (https://platform.claude.com/docs/en/about-claude/pricing; https://developers.openai.com/api/docs/pricing; https://ai.google.dev/gemini-api/docs/pricing; https://api-docs.deepseek.com/quick_start/pricing; https://docs.x.ai/developers/pricing; https://docs.mistral.ai/inference/pricing; https://dev.meta.ai/docs/pricing-rate-limits/).
    • cacheEconomics.agenticCacheShare: about 60% of agentic tokens were cache reads in January 2026 and just under 90% in September, per OpenRouter (https://openrouter.ai/data/newsletter/cache-is-crucial). cacheEconomics.effectivePrice: the effective price paid per million tokens on Ramp cards peaked at $1.15 in March 2026 and was $0.68 as of the September index, a 41% fall with no list-price cuts at the top (https://ramp.com/data/ai-index-sept-2026).
    • Chart code: one new section after the two-speeds chart, a step chart of the cache-read ratio by provider (fixed provider colours, log scale, direct labels on the latest point) with two stat tiles for the cache share and the effective price; the table view lists every point. No other chart changed.

17. Edición 10 · 2026-10-16

    A verification edition. Every quantitative claim in editions 8 and 9 was re-checked against its primary source on 16 Oct 2026; the current list prices, the METR rows added in edition 8, and every OpenRouter, Ramp and Epoch figure held. Three things are corrected. The open-versus-closed chart mixed two units (Artificial Analysis's cost per index task for closed models against a per-million-token blend for open models), so every row now uses cost per task; the open models remain a fraction of the closed models' cost at five to fifteen index points behind, but the gap is narrower than the old chart implied. Gemini 3.1 Pro's frontier date is corrected from September to February 2026. Six legacy METR rows are brought onto the current results file. Gemini 4 Argon reached the API price list on 13 Oct, the same day edition 8 said it was absent; the ladder now carries its dates.

    • openVsClosed now uses Artificial Analysis's leaderboard 'Cost per Task' (USD) for every row: GLM-5.3 Flash $0.28, DeepSeek V4.1 Flash $0.27, MiMo-V2.6-Pro $0.13, GLM-5.3 $2.01, Kimi K3 $2.00, Qwen3.8 27B $1.01, replacing the 7:2:1 cache/input/output per-million blend those open rows carried; closed rows were already on this basis and are unchanged. Index scores unchanged. The chart's column label and caption now say cost per task (https://artificialanalysis.ai/leaderboards/models; https://artificialanalysis.ai/models/open-source).
    • Gemini 3.1 Pro: frontier row dated 2026-02, not 2026-09; the preview launched 19 Feb 2026 at $2/$12 and is still labelled Preview on the price list (https://ai.google.dev/gemini-api/docs/changelog). Gemini 4 Argon is on the Gemini API price list since 13 Oct 2026 at $2/$10 (cache $0.10) through 30 Nov 2026 and $4/$20 from 1 Dec 2026; the tierLadder label carries those dates and the frontier row stays at the $4/$20 list price (https://ai.google.dev/gemini-api/docs/pricing).
    • METR: the six rows that predate edition 8 (GPT-4 0314 and 1106, Claude Sonnet 3.7, o3, Claude Opus 4, GPT-5) carried the January 2026 blog table's values; they now match METR's current results file to rounding (GPT-4 rows 4 min; Sonnet 3.7 60 min, CI 33-104; o3 120, CI 75-191; Opus 4 100, CI 60-164; GPT-5 203, CI 113-406), so the whole series rests on one file (https://metr.org/assets/benchmark_results_1_1.yaml).
    • Capex basis notes: Microsoft restated its CY2026 expectation to approximately $175B on the 29 Jul 2026 call after moving some finance leases to operating leases, with no change in planned investment; the ~$190B pre-reclassification figure is kept and the basis says so. Oracle's FY27 guide was first given with the Q4 FY26 release on 10 Jun 2026 and reaffirmed on 10 Sep 2026, not first given in September (https://www.microsoft.com/en-us/investor; https://www.oracle.com/news/announcement/q1fy27-earnings-release-2026-09-10/).
    • OpenAI run-rate marker now names the source of each point: ~$25B (The Information, early March; Semafor, April), $40B (Bloomberg, August), ~$50B booked at end-September (FT, 8 Oct), with the ~$70B late-September figure explained as gross cloud-partner sales (https://the-decoder.com/openai-revenue-keeps-surging-as-company-seeks-30-billion-in-fresh-capital/). Muse Wonder's $4/$20 is now confirmed on Meta's own API pricing page, so the frontier label drops the OpenRouter qualifier (https://dev.meta.ai/docs/pricing-rate-limits/).
    • Hosted open-model floor, 16 Oct 2026: gpt-oss-120b at $0.03/$0.15 is Venice and Near AI only (CoreWeave, DekaLLM, AkashML are $0.03/$0.17-0.19); DeepSeek V4.1 Flash is now cheapest at Baidu, $0.0357/$0.1428, across 34 endpoints; CoreWeave serves Llama 3.3 70B at $0.71 flat; StepFun released Step 5 Preview weights in BF16 and FP8 on 15 Oct and DeepInfra lists it at $0.95/$2.565 (https://openrouter.ai/api/v1/models/openai/gpt-oss-120b/endpoints; https://openrouter.ai/api/v1/models/deepseek/deepseek-v4.1-flash/endpoints; https://openrouter.ai/api/v1/models/stepfun/step-5-preview/endpoints).
    • Checked and unchanged: Anthropic, OpenAI, xAI, Mistral, DeepSeek, Z.ai and Moonshot list prices; Opus 5.5 Fast mode left research preview on 15 Oct at the same $8/$40; Ramp, OpenRouter and Epoch markers; D.C. Circuit opinion (petitions denied 25 Sep 2026, Henderson dissenting) with Anthropic petitioning for rehearing on 13 Oct per press; GSA OneGov Claude listing expires 31 Oct 2026. The 8,271-session corpus figures in the edition 9 markers appear only as dashboard alt text on the AgentX methodology page and are labelled derived; the Anthropic S-1 figures rest on Reuters' review of a confidential draft, with no public filing on EDGAR (https://platform.claude.com/docs/en/about-claude/pricing; https://media.cadc.uscourts.gov/opinions/docs/2026/09/26-1049-2194984.pdf; https://itvmo.gsa.gov/onegov).

18. Edición 9 · 2026-10-13

    Supplement to edition 8, same day: three markers from the only published request-level trace of an agentic coding workload, SemiAnalysis's AgentX corpus of Claude Code sessions. The median request carries 142,016 input tokens against 444 output, roughly 320 to 1, and the corpus runs at a 99% cache-hit rate, which is the mechanism behind the 90% cache-read share OpenRouter reports for agentic traffic and the reason the cost of serving an agent is set by KV-cache memory rather than by output price. Nothing else in the dataset changed.

    • SemiAnalysis AgentX v1.0 replay (built 21 Jun 2026, published with InferenceX v3 on 24 Aug 2026): 393 Claude Code sessions, 135,282 requests, median request 142,016 input tokens and 444 output tokens; 175 of the 393 sessions use subagents, in 1,697 subagent groups with a median duration of 2.27 minutes. Added as three enterprise markers (https://inferencex.semianalysis.com/agentx/methodology; https://newsletter.semianalysis.com/p/agentx-inferencexv3-does-cuda-moat).
    • The 99% cache-hit figure and the ~74M tokens-per-session figure come from SemiAnalysis's selection-stage corpus snapshot (8,271 sessions, 3.41M requests, 613.27B tokens), not from the released 393-session replay, and the per-session figure is derived by division; both are labelled as such in the marker. The 320:1 ratio is a ratio of medians, not a median of per-request ratios (https://inferencex.semianalysis.com/agentx/methodology).
    • Not added: SemiAnalysis's 5 Oct 2026 estimate that subscriptions are about 10% of Anthropic revenue but over 40% of its inference compute, and that a Claude Pro subscriber uses about 2.9B tokens a month, because the post is paywalled past the headline and the figures could only be read through The Register (https://newsletter.semianalysis.com/p/anthropic-subscriptions-offer-5x).

19. Edición 8 · 2026-10-13

    A day after edition 7, the movement was at the bottom of the ladder and in the measurement series, not at the ceiling. Anthropic's Claude Haiku 5.5 (7 Oct) lands at $0.10/$0.50, the same price as GPT-6 Luna and a 10x cut to Anthropic's own small tier in one step; Meta's first paid frontier model, Muse Wonder (12 Oct), arrives at $4/$20; and METR's own results file, read directly for the first time, replaces two secondary horizon figures with primary ones and adds three models, including a 17-hour point for an early Claude Mythos preview that METR says sits beyond its suite's reliable range. OpenRouter's own newsletter gives the first primary weekly total (128T), reports that agents now generate about seven times the tokens humans do and that nearly 90% of agentic tokens are cache reads, which is the mechanism behind Ramp's 41% fall in the effective price paid per token while list prices at the top stood still. Astra, Fable, Argon and Grok are unchanged; Q3 hyperscaler earnings (21-29 Oct) are the next likely movement in capex.

    • Claude Haiku 5.5 launched 7 Oct 2026 at $0.10 input / $0.50 output per million tokens for prompts up to 100K ($0.50/$2.50 above), 1M context, on all platforms; Anthropic says it is about 75% cheaper to run than Haiku 4.5, which stays listed at $1/$5. Sonnet 5.5's cache-read price was cut from $0.20 to $0.10 the same day; Fable 5.1, Opus 5.5 and Sonnet 5.5 list prices unchanged. tierLadder's Anthropic small row now carries Haiku 5.5 (https://www.anthropic.com/claude-haiku-5-5; https://platform.claude.com/docs/en/about-claude/pricing).
    • Meta released Muse Wonder and Muse Wonder Full on 12 Oct 2026; Wonder is served by Meta itself on OpenRouter at $4/$20 with a 1M context, Wonder Full is withheld from the API, and Meta's post gives no price or weights. Artificial Analysis scores Wonder (max) 52 at $4.52 and Wonder Full 56 with no price. Meta is added to frontier and tierLadder at $4/$20 and the unreproducible Muse Spark 1.3 row in openVsClosed is replaced by Muse Wonder (https://about.fb.com/news/2026/10/introducing-muse-wonder-models/; https://openrouter.ai/api/v1/models/meta/muse-wonder/endpoints; https://artificialanalysis.ai/leaderboards/models).
    • OpenAI standard prices unchanged (Astra $10/$50, GPT-6.1 Sol $2/$10, Luna $0.10/$0.50); the changelog adds an Ultrafast tier for GPT-6.1 Sol at $12/$60 on 8 Oct, Astra Ultrafast at $60/$300, and billing for gpt-rosalind-research ($5/$25) from 5 Oct. Google's Gemini 4 Argon is still absent from the API price list two weeks after announcement, gemini-3.7-flash was deprecated on 8 Oct, and Reuters reports Gemini 3.5 Pro will not ship. xAI unchanged at $2/$6 for Grok 4.7 (https://developers.openai.com/api/docs/changelog; https://ai.google.dev/gemini-api/docs/changelog; https://docs.x.ai/developers/pricing).
    • METR's primary results file (METR-Horizon v1.1) replaces the secondary figures on record: Claude Opus 4.6 is 719 min (95% CI 317-3,634), not 870; GPT-5.3-Codex 350 min (195-816), not 390; Claude Opus 4.5 is revised to 293 min (162-624). Added: Gemini 3.1 Pro 384 min (234-695, Feb 2026), GPT-5.4 342 min (187-769, Mar 2026) and an early Claude Mythos preview at 1,045 min (509-3,304, Apr 2026), which exceeds the 16-hour limit METR calls unreliable and is excluded from its doubling-time fit. The since-2023 doubling time is 128.7 days (CI 104-158); the since-2024 figure is not in the file and is retained from the Jan 2026 post. METR's time-horizons page says it was last updated 8 Sep 2026 and is no longer actively updated; no horizon exists for Opus 5.5, Fable 5.1 or GPT-6 Astra (https://metr.org/assets/benchmark_results_1_1.yaml; https://metr.org/time-horizons/).
    • OpenRouter's own newsletter reports 128.4T tokens in the week of 14 Sep 2026 (0.5T in Jan 2025), replacing the 134T third-party figure; agents generated about 7x the tokens humans did in late September with just under 90% of agentic tokens served from cache (60% in January); OpenAI took 78% of OpenAI+Anthropic tokens and just over half of their combined spend; open-weight models have held over 62% of tokens since July while closed models take 73% of spend; Chinese labs serve 69% of tokens. Week to 12 Oct: DeepSeek V4.1 Flash 40.5T, GLM 5.3 Flash 12.7T, Step 5 Preview 12.1T in its first week at $1/$2.70 with weights promised 15 Oct, Claude Opus 5.5 5.35T (+106%) (https://openrouter.ai/data/newsletter/the-token-state-of-play; https://openrouter.ai/data/newsletter/cache-is-crucial; https://openrouter.ai/rankings).
    • Ramp's September index (August card data, 9 Sep 2026): effective price paid fell 41% to $0.68 per million tokens from a $1.15 peak in March; frontier models' token share fell to 45% from a 53% peak; 6.4% of AI-paying businesses use open-source models; top-1% spend per employee fell 9.7%. OpenAI's run rate: the FT puts OpenAI-booked revenue near $50B at end-September with a $70B year-end target, and the late-September $70B figure counted gross cloud-partner sales, so the marker now carries both. The $100B Anthropic point for Sep 2026 is NYT's 'on track to top $100B by year-end', a projection, and is left as recorded with that caveat. Epoch: the median OpenAI researcher's daily coding-agent spend rose from under $1 in January to $601 in August at list prices, doubling every ~34 days; SemiAnalysis via Epoch puts Anthropic's direct API gross margin near 85% against ~45% through Bedrock, and Z.ai's share of GLM-5.3 Flash tokens on OpenRouter fell from 88% to 22% once the weights were public (https://ramp.com/data/ai-index-sept-2026; https://the-decoder.com/openai-revenue-keeps-surging-as-company-seeks-30-billion-in-fresh-capital/; https://epoch.ai/data-insights/openai-coding-agent-spending; https://epoch.ai/publications/how-do-chinese-ai-companies-make-money).
    • Hosted open-model floor, 13 Oct 2026: gpt-oss-120b from $0.03/$0.15 (Venice) across 24 OpenRouter endpoints with the $0.15/$0.60 mode unchanged; DeepSeek V4.1 Flash from $0.09/$0.18 (Decart) against a $0.30/$1.20 mode; Llama 3.3 70B from $0.10/$0.32 (DeepInfra). Mistral's page now calls the model Mistral Large 4 with weights promised by end-October and the half-price sale still running. Artificial Analysis adds Qwen3.8 27B (index 34, $0.5) as the sixth open model, marks GPT-6 Sol deprecated (row removed, GPT-6.1 Sol stands in its place) and now shows Claude Sonnet 5.5 (max) at $5.46 on the leaderboard (was $7.67). Caveat recorded for a future chart fix: AA's open-model price column is a 7:2:1 cache/input/output blend while the closed rows use the leaderboard cost column, so openVsClosed mixes two bases; no number was changed to fix this (https://openrouter.ai/api/v1/models/openai/gpt-oss-120b/endpoints; https://docs.mistral.ai/inference/pricing; https://artificialanalysis.ai/models/open-source).
    • Unchanged, with reasons: capex, because no hyperscaler has reported (Meta 21 Oct, Microsoft and Alphabet 28 Oct, Amazon 29 Oct); Epoch's decline estimate, whose 23 Sep update only added a citation; DeepSeek, Z.ai, Moonshot and Qwen official prices, though DeepSeek's 10 Sep note routes V4 Pro requests to V4.1 Flash at Flash rates while its pricing page still lists Pro separately; twoSpeeds, since neither the OpenAI frontier output price nor the cheapest GPT-4-level blended price moved. Policy: the D.C. Circuit stayed its 25 Sep judgment seven days for an en banc petition and the separate statutory designation remains enjoined in the Northern District of California; GSA extended Anthropic's $1 OneGov deal to 31 Oct 2026 (https://api-docs.deepseek.com/news/news260910; https://www.hklaw.com/en/insights/publications/2026/09/dc-circuit-upholds-exclusion-of-anthropic-from-dow-supply-chain; https://fedscoop.com/anthropic-extends-gsa-onegov-deal-for-claude-by-one-month).

20. Edición 7 · 2026-10-12

    A quiet six days between editions: every pricing page, the Artificial Analysis leaderboard and METR's blog were re-read and matched the recorded figures exactly, with no hyperscaler yet into Q3 earnings season and no new enterprise run-rate figure past September's. The one real event of the week was cosmetic - Mistral dropped the 'Preview' suffix from Large 4's name without moving its price - which is itself consistent with the thesis: nothing moved on the ceiling or the floor because nothing forced it to. This is a verification edition; Q3 hyperscaler earnings later in October are the likely next source of real movement in capex and the frontier ceiling.

    • Verified on the primary pricing pages, 12 Oct 2026: OpenAI gpt-6-astra $10/$50, gpt-6.1-sol $2/$10, gpt-6-luna $0.10/$0.50, all unchanged since edition 6; the page now also publishes Batch/Flex/Fast/Ultrafast service-tier prices for Sol and Luna, not only Astra, but the Standard-tier figures already on file are unchanged (https://developers.openai.com/api/docs/pricing).
    • Anthropic unchanged: Claude Fable 5.1 $10/$50, Claude Opus 5.5 $4/$20, Claude Sonnet 5.5 $2/$10, Claude Haiku 4.5 $1/$5, all matching the recorded frontier and tierLadder rows exactly (https://platform.claude.com/docs/en/about-claude/pricing).
    • Google unchanged: Gemini 3.1 Pro Preview $2/$12, Gemini 3.8 Flash intro $0.75/$3.75 through 31 Dec 2026, Gemini 3.5 Flash-Lite $0.30/$2.50. Gemini 4 Argon is still absent from Google's own pricing page twelve days after announcement, so its frontier/tierLadder rows still rest on the launch post via secondary coverage rather than a primary price list (https://ai.google.dev/gemini-api/docs/pricing).
    • xAI grok-4.7 unchanged at $2/$6 for prompts under 200k tokens; the page separately lists a $4/$12 rate at 200k tokens or more, a context-length tier not previously recorded here, but the standard rate used in this series is unchanged (https://docs.x.ai/developers/pricing).
    • DeepSeek unchanged: DeepSeek V4.1 Flash $0.30/$1.20 peak ($0.15/$0.60 off-peak), DeepSeek V4 Pro $1.32/$3.96 peak ($0.66/$1.98 off-peak) (https://api-docs.deepseek.com/quick_start/pricing).
    • Mistral unchanged on price: Large 4 sale $0.68/$2.09 (list $1.36/$4.18), Small 4 $0.15/$0.60. Mistral's own pricing page has quietly dropped the 'Preview' suffix from Large 4's name, while Artificial Analysis still lists it as 'Mistral Large 4 Preview'; recorded as a naming note only, no price changed (https://docs.mistral.ai/inference/pricing).
    • Artificial Analysis Intelligence Index scores unchanged across every tracked model (Claude Opus 5.5 58, Claude Sonnet 5.5 56, Claude Fable 5.1 53, GPT-6 Astra 53, GPT-6.1 Sol 52, Gemini 4 Argon 53, Mistral Large 4 Preview 38; open-weight leaders MiMo-V2.6-Pro 46, GLM-5.3 45, Kimi K3 44, GLM-5.3 Flash 42, DeepSeek V4.1 Flash 39). This session's read of the leaderboard surfaced a 'Cost per Task' column rather than the blended $/M figure the openVsClosed series uses, so no blended price was touched this edition to avoid mixing units (https://artificialanalysis.ai/leaderboards/models; https://artificialanalysis.ai/models/open-source).
    • OpenRouter's rankings page did not return weekly token-volume figures to this session's fetch, only its Intelligence Index ranking, so tokenVolumes is left unchanged rather than sourced from a partial read (https://openrouter.ai/rankings). METR's only post this week (6 Oct 2026, 'AI systems could cover up misbehavior') concerns observability, not a time-horizon measurement, and no new horizon has been published since the 22 Sep Opus 5.5 summary, which itself gave none; metrHorizons is unchanged (https://metr.org/blog/). No hyperscaler reported or re-guided capex this week; Q3 2026 earnings calls fall after this edition's window, so capex is unchanged. No enterprise run-rate figure newer than September's was found for OpenAI or Anthropic, so enterprise.markers is unchanged.

21. Edición 6 · 2026-10-06

    Verification edition, run the same day as edition 5 after this session's network access was widened to the primary sources. Every edition 5 row checked out against the provider pages themselves: OpenAI's pricing page lists gpt-6.1-sol at $2/$10 beside Astra's $10/$50, and Artificial Analysis now scores GPT-6.1 Sol one index point below Astra at about a fifth of the cost to run, which is the two-speeds thesis in a single table row. New today: Mistral launched Large 4 Preview, a 1T-parameter model it says will ship as open weights by the end of October, listed at $1.36/$4.18 with a launch sale at half that, so the open frontier gains a European entrant; and Anthropic's S-1, as reviewed by Reuters, puts 2025 revenue at $4.6B against $518B of committed compute with a quarter of revenue from two customers, the capital wall and the narrow buyer in one filing. METR, OpenRouter volume and capex are unchanged, with the reasons below.

    • Verified on the primary pricing pages, 6 Oct 2026: OpenAI gpt-6-astra $10/$50, gpt-6.1-sol $2/$10 (cached $0.10), gpt-6-luna $0.10/$0.50 (https://developers.openai.com/api/docs/pricing); Google Gemini 3.1 Pro Preview $2/$12, 3.8 Flash intro $0.75/$3.75 through 31 Dec 2026, 3.5 Flash-Lite $0.30/$2.50 (https://ai.google.dev/gemini-api/docs/pricing); xAI grok-4.7 $2/$6 (https://docs.x.ai/developers/pricing); DeepSeek V4.1 Flash $0.30/$1.20 and V4 Pro $1.32/$3.96 at peak, half off-peak (https://api-docs.deepseek.com/quick_start/pricing); Anthropic unchanged (https://platform.claude.com/docs/en/about-claude/pricing). Gemini 4 Argon is not yet on Google's pricing page, so its rows still rest on Google's launch statement as quoted by 9to5google.
    • Mistral Large 4 Preview launched 6 Oct 2026: 1T parameters, 49B active, trained on 3,800 Grace Blackwell GPUs in Mistral's own European datacenters, weights promised by the end of October; list $1.36/$4.18, launch sale $0.68/$2.09. Added to frontier at list price and as Mistral's flagship tier at the sale price; Small 4 ($0.15/$0.60) added as the small tier and Large 3 dropped from the ladder as superseded (https://mistral.ai/news/mistral-large-4/; https://mistral.ai/pricing/api).
    • Artificial Analysis Intelligence Index v4.3.2, read 6 Oct 2026: Claude Opus 5.5 (max) 58 at $5.98, Claude Sonnet 5.5 (max) 56 at $7.67, GPT-6 Astra (max) 53 at $3.26, Gemini 4 Argon (high) 53 at $1.99, Claude Fable 5.1 (max) 53 at $7.63, GPT-6.1 Sol (max) 52 at $0.72, Qwen3.8 Max 45 at $5.41, Mistral Large 4 Preview 38 at $1.13. Open-weight rows (MiMo-V2.6-Pro 46, GLM-5.3 45, Kimi K3 44, GLM 5.3 Flash 42, DeepSeek V4.1 Flash 39) are unchanged; the new closed rows are added to openVsClosed (https://artificialanalysis.ai/leaderboards/models; https://artificialanalysis.ai/models/open-source).
    • METR's 22 Sep 2026 predeployment summary of Claude Opus 5.5 gives no time-horizon figure. Its 26 Jun 2026 summary of GPT-5.6 Sol reports a 50% horizon near 11.3 hours (95% CI 5 to 40 hours) only after scoring detected cheating as failure, and states that none of its figures is a robust measurement, so metrHorizons is unchanged (https://metr.org/blog/2026-09-22-claude-opus-5-5/; https://metr.org/blog/2026-06-26-gpt-5-6-sol/).
    • OpenRouter, week to 5 Oct 2026: stealth Space Bunny Alpha 38.5T tokens, DeepSeek V4.1 Flash 28.4T, GLM 5.3 Flash 10T, MiMo-V2.6-Flash 9.72T, Hy4 preview 6.36T, GPT-6 Luna 6.21T. The top ten sum to about 117T but the page publishes no platform total, so tokenVolumes is unchanged; the GLM 5.3 Flash row in hostedOpen now carries its current rank (https://openrouter.ai/rankings).
    • Anthropic's S-1 as reviewed by Reuters, 28 Sep 2026: 2025 revenue about $4.6B (12x on 2024), operating loss $8.06B, net loss $42B including a roughly $34B non-cash financing charge, $7.33B spent on compute in 2025, $518B of cloud and compute obligations in coming years, and nearly a quarter of 2025 revenue from two customers. Added as two enterprise markers (https://www.cnbc.com/2026/09/28/anthropics-ipo-prospectus-shows-sweeping-ai-vision-surging-costs-reuters.html).
    • OpenAI DevDay, 29 Sep 2026, beyond GPT-6.1 Sol: a $500/month Pro plan with Pro 200 usage halved, and the October GPT-6.1 Astra launch cancelled after failed safety tests; OpenAI's pricing page also lists Astra at $100/M output in Fast and $300/M in Ultrafast. Added as two markers (https://thenextweb.com/news/openai-gpt-6-1-sol-price-astra-devday; https://developers.openai.com/api/docs/pricing).
    • Run rates reconfirmed in press readable this session: Anthropic pacing above $100B annualized, up 50% in two months (Bloomberg relaying the New York Times, 18 Sep 2026, https://www.bloomberg.com/news/articles/2026-09-18/anthropic-s-annualized-revenue-to-top-100-billion-in-2026-nyt); OpenAI nearing $70B, up more than 70% since the start of Q3 (Yahoo Finance relaying Axios, 29 Sep 2026, https://finance.yahoo.com/technology/article/openais-annualized-revenue-nearing-70-billion-report-164411134.html). Epoch's open-closed gap of four months and 8 ECI points reconfirmed (https://epoch.ai/data-insights/open-closed-eci-gap). Capex unchanged: no earnings in the window.

22. Edición 5 · 2026-10-06

    Two frontier labs moved in the same week, and in the direction the thesis predicts. OpenAI's GPT-6.1 Sol (29 Sep) puts near-Astra capability at one fifth of Astra's price while leaving Astra's $10/$50 ceiling untouched, and Google announced Gemini 4 Argon (30 Sep) above Gemini 3.1 Pro with a $4/$20 list price, temporarily $2/$10, its first new ceiling since Gemini 3 Pro. OpenAI's run rate was reported near $70B, up from a ~$25B plateau in the spring; with Anthropic reported near $100B, the two labs' combined pace has roughly tripled since mid-year. Primary sources other than Anthropic's remain blocked from this session, so these rows rest on reputable press plus the official announcement URLs that search surfaced; METR, Artificial Analysis, OpenRouter and capex series are unchanged.

    • OpenAI launched GPT-6.1 Sol at DevDay on 29 Sep 2026 at $2/$10 per million tokens (cached input $0.10), the same list price as GPT-6 Sol and one fifth of GPT-6 Astra's $10/$50, which is unchanged; tierLadder's OpenAI flagship row is renamed, no price change (https://openai.com/index/introducing-gpt-6-1-sol/; https://techcrunch.com/2026/09/29/openai-launches-gpt-6-1-sol-says-it-nearly-matches-gpt-6-astra-and-costs-less/). openai.com was unreachable from this session; the figures come from TechCrunch and The Next Web coverage.
    • Google announced Gemini 4 Argon on 30 Sep 2026: introductory $2/$10 per million tokens, $4/$20 once the unannounced intro period ends, with a 1M-token output limit; rollout starts with cyber defenders in DeepMind's Fairwind Program before paid API customers and AI Ultra subscribers. Added to frontier at its $4/$20 list price (the series records list price at launch; the intro is in the label) and dated 2026-10, when access begins; added to tierLadder as Google's premium tier at the $2/$10 intro price actually charged, following the Gemini 3.8 Flash convention (https://x.com/Google/status/2105388143902175529; https://9to5google.com/2026/09/30/gemini-4-argon-announcement/). ai.google.dev was unreachable.
    • OpenAI's annualized revenue run rate was reported nearing $70B in late September 2026, up more than 70% since July, after a $40B figure in August broke a five-month plateau near $25B (Axios scoop, 29 Sep 2026, with a source also speaking to Reuters); added to enterprise.markers (https://www.axios.com/2026/09/29/scoop-openais-annual-recurring-revenue-nears-70b).
    • Reconfirmed with no change: Grok 4.7 (21 Sep) kept Grok 4.6's $2/$6; DeepSeek V4.1 Flash has been $0.30/$1.20 peak since 10 Sep and V4 Pro $1.32/$3.96 since 16 Aug; Mistral Medium 3.5 $1.50/$7.50 and Large 3 $0.50/$1.50; Anthropic's table is identical to edition 4 (https://docs.x.ai/developers/release-notes; https://api-docs.deepseek.com/updates/; https://platform.claude.com/docs/en/about-claude/pricing). The xAI and DeepSeek pages were unreachable; their figures were checked against digitalapplied.com's Q3 price tracker and Reuters via investing.com.
    • METR has published no time horizon for Claude Fable 5.1, GPT-6 Astra or Claude Opus 5.5. The 153h, 38.8h and 88h figures circulating are a third party's ECI-fit extrapolations (https://github.com/joansam/metr-chart/pull/3), not measurements, and are not added.
    • OpenRouter's week to 4 Oct 2026: a stealth model, space-bunny-alpha, took 38.66T tokens, 23.3% of the platform's volume (implying roughly 166T for the week), ahead of DeepSeek V4.1 Flash at 25.61T, and was withdrawn on 5 Oct. The implied total is derived from a newsletter's percentage, so no tokenVolumes point is added until openrouter.ai can be read directly (https://chinaaidispatch.substack.com/p/the-monday-brief-22-everything-else).
    • Epoch AI's open-closed gap stands at four months and 8 ECI points since January 2026, consistent with the essay's 'about four months behind'; the dataset has no series for it, so no change (https://epoch.ai/data-insights/open-closed-eci-gap).
    • Network egress from this session still blocks every primary source except Anthropic's domains; hostedOpen, openVsClosed, metrHorizons, tokenVolumes and capex (no earnings in the window) are left as-is.

23. Edición 4 · 2026-10-05

    This session's network egress policy blocked almost every primary source this report relies on: OpenAI, Google, xAI, DeepSeek and Mistral's pricing pages, METR, Epoch AI, Artificial Analysis, OpenRouter and Google's blog all returned egress-blocked errors, as did most reputable press (Reuters, TechCrunch, VentureBeat, CNBC, Axios article pages) - only Anthropic's own domains were reachable. The one confirmed, dated change is on Anthropic's side: Claude Sonnet 5.5 (launched 2026-09-28, same $2/$10 price as Sonnet 5) is now the current mid-tier model, and Anthropic's revenue run rate is reported pacing toward roughly $100B annualized by mid-September, up from $65B in July. Every other series could only be checked against conflicting third-party aggregator snippets that could not be verified on a primary page, so it is left unchanged per the no-invented-figures rule.

    • Claude Sonnet 5.5 launched 2026-09-28 at the same $2/$10 per-million-token price as Sonnet 5 and is now Anthropic's current mid-tier model; tierLadder's Anthropic 'mid' row is renamed accordingly, no price change (https://platform.claude.com/docs/en/about-claude/pricing; https://www.anthropic.com/news).
    • Claude Mythos 5.1, a new invite-only 'Project Glasswing' model, is priced identically to Fable 5.1 ($10/$50) per Anthropic's pricing page; not added as its own row since it moves no price point and is not generally available (https://platform.claude.com/docs/en/about-claude/pricing).
    • Anthropic's annualized revenue run rate was reported pacing above $100B by mid-September 2026, up from the $65B figure recorded in July, per Axios, 18 Sep 2026; added as a new point in enterprise.anthropicRunRate (https://www.axios.com/2026/09/18/anthropic-100-billion-revenue).
    • developers.openai.com, ai.google.dev, docs.x.ai, api-docs.deepseek.com, mistral.ai, metr.org, epoch.ai, artificialanalysis.ai, openrouter.ai, blog.google, and most press domains (reuters.com, techcrunch.com, venturebeat.com, cnbc.com, axios.com) returned EGRESS_BLOCKED from this session; frontier/tierLadder entries for OpenAI, Google, xAI, DeepSeek and Mistral, priceToCapability, epochDecline, metrHorizons, tokenVolumes, openVsClosed, hostedOpen and selfHost are left unchanged this edition rather than sourced from unverifiable aggregator sites.
    • Search-engine snippets (not independently fetchable this edition) suggested Alphabet's and Meta's 2026 capex guidance remains at the $195-205B and $130-145B ranges already recorded in this dataset, consistent with no change, so capex is left as-is.
    • A search-engine summary referenced an OpenAI 'GPT-6.1 Sol' model from an alleged October 1, 2026 DevDay, said to match GPT-6 Sol's $2/$10 price, but this came only from SEO aggregator sites rather than openai.com (unreachable) and was not added, per the no-invented-figures rule.

24. Edición 3 · 2026-09-28

    Capex was the one series with real, confirmed news this edition: Amazon raised its 2026 guide to ~$220B (from ~$200B) on rising memory costs back on 30 Jul 2026, and Oracle gave its first FY2027 capex guidance (~$70B net cash outlay) on its Q4 FY26 call in early September 2026 - both predate this edition's window but were missing from the dataset, so they are filled in now. Every other series (frontier, tierLadder, METR) was reconfirmed against Anthropic's own pricing page and reputable coverage of OpenAI's official GPT-6 Sol/Luna launch post, with no material change since edition 2. OpenRouter and Artificial Analysis were unreachable this edition, so hostedOpen, openVsClosed and tokenVolumes are left as-is rather than sourced from inconsistent third-party aggregators.

    • Amazon's 2026 capex guidance was raised to ~$220B from ~$200B, cited to rising memory/DRAM costs, on its Q2 2026 earnings call, 30 Jul 2026 (https://www.cnbc.com/2026/07/30/amazon-amzn-q2-earnings-report-2026.html; https://www.fool.com/investing/2026/08/24/jeff-bezos-amazon-just-raised-its-ai-spending-to-2/).
    • Oracle's first FY2027 capex guidance, ~$70B net cash outlay of its own money plus an estimated $20-25B repaid by customers (~$90-95B gross total), was given on its Q4 FY2026 earnings call around 10 Sep 2026; recorded under this series' existing 'Oracle FY shown one year back' convention, so it lands in the 2026 column (https://seekingalpha.com/news/4602439-oracle-expects-34-percent-fy27-revenue-growth-as-it-plans-70b-net-cash-capex-outlay; https://thebuildout.ai/intel/3e5ab1c5-857d-4585-982f-6d8082874d28).
    • Anthropic's official pricing page confirms Claude Fable 5.1 ($10/$50), Claude Opus 5.5 ($4/$20) and Claude Sonnet 5's introductory $2/$10 now made the permanent, ongoing rate - all already matched the recorded frontier/tierLadder rows, no change (https://platform.claude.com/docs/en/about-claude/pricing).
    • OpenAI's GPT-6 Sol ($2/$10) and GPT-6 Luna ($0.10/$0.50), launched 22 Sep 2026 as permanent (not promotional) price cuts, already matched the recorded tierLadder (https://openai.com/index/introducing-gpt-6-sol-and-luna/; https://venturebeat.com/technology/openai-releases-gpt-6-sol-and-luna-models-slashing-api-costs-50-or-more).
    • METR's Time Horizon 1.1 post remains the latest published methodology; this week's independent commentary on it describes the same doubling cadence already reflected in metrNotes, no new dated model horizon to add (https://metr.org/blog/2026-1-29-time-horizon-1-1/).
    • developers.openai.com, openai.com, openrouter.ai and artificialanalysis.ai could not be reached from this session; third-party aggregator figures for OpenRouter weekly volume and open-weight leaderboard scores found this week were inconsistent with each other (mismatched units, conflicting top-model rankings) and were not used, per the no-invented-figures rule.

25. Edición 2 · 2026-09-27

    Verification pass run the same day as edition 1's launch; no material change to any series. Checked provider pricing pages, Artificial Analysis, OpenRouter, and METR for anything published since edition 1 and found only confirmations of figures already recorded, so every series is left as-is this edition.

    • OpenAI GPT-6 Astra ($10/$50, launched 2026-09-03), GPT-6 Sol ($2/$10) and GPT-6 Luna ($0.10/$0.50), both launched 2026-09-22, already match the recorded tierLadder (https://developers.openai.com/api/docs/pricing).
    • Anthropic Claude Fable 5.1 ($10/$50, launched 2026-09-01) and Claude Opus 5.5 ($4/$20 flagship tier, launched 2026-09-22) already match the recorded frontier and tierLadder rows (https://platform.claude.com/docs/en/about-claude/pricing).
    • Google Gemini 3.8 Flash intro pricing ($0.75/$3.75 input/output through 2026-12-31, reverting to $1.50/$7.50) already matches the recorded tierLadder note (https://ai.google.dev/gemini-api/docs/pricing).
    • 2026 hyperscaler capex guidance reconfirmed in the same range already on file: Amazon ~$200B, Alphabet $195-205B (one report cites 'as much as $205B'), Meta $130-145B (https://www.globaldatacenterhub.com/p/microsoft-q3-fy2026-the-190b-capex, https://valueaddvc.com/blog/meta-145b-ai-capex-2026-why-zuckerberg-raised-guidance-twice).
    • METR's Time Horizon 1.1 methodology (the source already cited for metrHorizons/metrNotes) remains the latest published update; no new dated model horizon point has been released since edition 1 (https://metr.org/blog/2026-1-29-time-horizon-1-1/).
    • No new open-weight leaderboard score, OpenRouter volume figure, or enterprise usage marker was found beyond what edition 1 already recorded.

26. Edición 1 · 2026-09-27

    First edition. Six research strands compiled into the dataset: frontier and mid-tier price history, price-to-capability decline, hyperscaler capex, METR horizons, token volumes, open-weight and self-hosted prices, and enterprise usage markers.

    • GPT-6 Astra ($10/$50) and Opus 5.5 ($4/$20) added as the newest frontier prices; Fable 5.1 unchanged at $10/$50.
    • DeepSeek V4 Pro peak pricing ($1.32/$3.96) recorded after the August increase.
    • OpenRouter weekly volume ~134T tokens (17-22 Sep 2026), top ten all at $0.84 input or less.
    • Capex 2026 guidance: Microsoft ~$190B, Alphabet $195-205B, Amazon ~$200B, Meta $130-145B.

Método: los precios son precios de lista por millón de tokens tomados de las páginas de precios y anuncios de lanzamiento de cada proveedor; el capex sale de los reportes de resultados y filings; los horizontes son mediciones de METR time-horizon 1.1; los volúmenes los reportan las propias empresas. Las cifras marcadas como secundarias vienen de prensa o de rastreadores de terceros cuando no se pudo consultar la página primaria. Cada gráfica tiene vista de tabla y cada cifra su fuente.