{"title":"Token Economics report","humanPage":"https://www.geoffsdesk.com/human/token-economics","essay":"https://www.geoffsdesk.com/machine/blog/token-economics.md","cadence":"weekly","units":"USD per million tokens unless stated; capex in USD billions; horizons in minutes; volumes in trillions of tokens","asOf":"2026-09-27","edition":2,"notes":{"prices":"USD per million tokens, list price at launch unless noted. 'Frontier' means each provider's newest top non-Pro model at that date.","blended":"Blended $/M uses a 3:1 input:output mix where a single figure was needed."},"frontier":[{"date":"2023-03","provider":"OpenAI","model":"GPT-4","input":30,"output":60},{"date":"2023-11","provider":"OpenAI","model":"GPT-4 Turbo","input":10,"output":30},{"date":"2024-05","provider":"OpenAI","model":"GPT-4o","input":5,"output":15},{"date":"2024-08","provider":"OpenAI","model":"GPT-4o","input":2.5,"output":10},{"date":"2025-04","provider":"OpenAI","model":"GPT-4.1","input":2,"output":8},{"date":"2025-08","provider":"OpenAI","model":"GPT-5","input":1.25,"output":10},{"date":"2025-12","provider":"OpenAI","model":"GPT-5.2","input":1.75,"output":14},{"date":"2026-03","provider":"OpenAI","model":"GPT-5.4","input":2.5,"output":15},{"date":"2026-04","provider":"OpenAI","model":"GPT-5.5","input":5,"output":30},{"date":"2026-06","provider":"OpenAI","model":"GPT-5.6 Sol","input":5,"output":30},{"date":"2026-09","provider":"OpenAI","model":"GPT-6 Astra","input":10,"output":50},{"date":"2023-07","provider":"Anthropic","model":"Claude 2","input":8,"output":24},{"date":"2024-03","provider":"Anthropic","model":"Claude 3 Opus","input":15,"output":75},{"date":"2025-05","provider":"Anthropic","model":"Claude Opus 4","input":15,"output":75},{"date":"2025-11","provider":"Anthropic","model":"Claude Opus 4.5","input":5,"output":25},{"date":"2026-06","provider":"Anthropic","model":"Claude Fable 5","input":10,"output":50},{"date":"2026-09","provider":"Anthropic","model":"Claude Fable 5.1","input":10,"output":50},{"date":"2024-05","provider":"Google","model":"Gemini 1.5 Pro","input":3.5,"output":10.5},{"date":"2024-10","provider":"Google","model":"Gemini 1.5 Pro","input":1.25,"output":5},{"date":"2025-04","provider":"Google","model":"Gemini 2.5 Pro","input":1.25,"output":10},{"date":"2025-11","provider":"Google","model":"Gemini 3 Pro","input":2,"output":12},{"date":"2026-09","provider":"Google","model":"Gemini 3.1 Pro","input":2,"output":12},{"date":"2024-11","provider":"xAI","model":"grok-beta","input":5,"output":15},{"date":"2025-04","provider":"xAI","model":"Grok 3","input":3,"output":15},{"date":"2026-02","provider":"xAI","model":"Grok 4.20","input":1.25,"output":2.5},{"date":"2026-07","provider":"xAI","model":"Grok 4.5","input":2,"output":6},{"date":"2026-09","provider":"xAI","model":"Grok 4.7","input":2,"output":6},{"date":"2024-12","provider":"DeepSeek","model":"DeepSeek V3","input":0.27,"output":1.1},{"date":"2025-04","provider":"DeepSeek","model":"DeepSeek R1","input":0.55,"output":2.19},{"date":"2026-04","provider":"DeepSeek","model":"DeepSeek V4 Pro","input":1.74,"output":3.48},{"date":"2026-08","provider":"DeepSeek","model":"DeepSeek V4 Pro (peak)","input":1.32,"output":3.96},{"date":"2024-02","provider":"Mistral","model":"Mistral Large","input":8,"output":24},{"date":"2024-08","provider":"Mistral","model":"Mistral Large 2","input":2,"output":6},{"date":"2025-12","provider":"Mistral","model":"Mistral Large 3","input":0.5,"output":1.5},{"date":"2026-04","provider":"Mistral","model":"Mistral Medium 3.5","input":1.5,"output":7.5}],"twoSpeeds":[{"date":"2023-03","frontier":60,"fixed":37.5,"frontierModel":"GPT-4","fixedModel":"GPT-4 (Epoch AI, GPT-4-level MMLU)"},{"date":"2023-11","frontier":30,"fixed":15,"frontierModel":"GPT-4 Turbo","fixedModel":"GPT-4 Turbo, blended"},{"date":"2024-05","frontier":15,"fixed":7.5,"frontierModel":"GPT-4o","fixedModel":"GPT-4o, blended"},{"date":"2024-08","frontier":10,"fixed":0.26,"frontierModel":"GPT-4o","fixedModel":"GPT-4o mini, blended"},{"date":"2025-02","frontier":10,"fixed":0.18,"frontierModel":"GPT-4o","fixedModel":"Epoch AI: GPT-4-level MMLU at $0.18"},{"date":"2025-08","frontier":10,"fixed":0.14,"frontierModel":"GPT-5","fixedModel":"GPT-5 nano, blended"},{"date":"2025-12","frontier":14,"fixed":0.14,"frontierModel":"GPT-5.2","fixedModel":"GPT-5 nano, blended"},{"date":"2026-04","frontier":30,"fixed":0.14,"frontierModel":"GPT-5.5","fixedModel":"GPT-5 nano, blended"},{"date":"2026-09","frontier":50,"fixed":0.1,"frontierModel":"GPT-6 Astra","fixedModel":"GLM-5.3 Flash, Artificial Analysis blended"}],"tierLadder":[{"provider":"OpenAI","tier":"premium","model":"GPT-6 Astra","input":10,"output":50},{"provider":"OpenAI","tier":"flagship","model":"GPT-6 Sol","input":2,"output":10},{"provider":"OpenAI","tier":"small","model":"GPT-6 Luna","input":0.1,"output":0.5},{"provider":"Anthropic","tier":"premium","model":"Claude Fable 5.1","input":10,"output":50},{"provider":"Anthropic","tier":"flagship","model":"Claude Opus 5.5","input":4,"output":20},{"provider":"Anthropic","tier":"mid","model":"Claude Sonnet 5","input":2,"output":10},{"provider":"Anthropic","tier":"small","model":"Claude Haiku 4.5","input":1,"output":5},{"provider":"Google","tier":"flagship","model":"Gemini 3.1 Pro","input":2,"output":12},{"provider":"Google","tier":"mid","model":"Gemini 3.8 Flash (intro; $1.50/$7.50 list)","input":0.75,"output":3.75},{"provider":"Google","tier":"small","model":"Gemini 3.5 Flash-Lite","input":0.3,"output":2.5},{"provider":"xAI","tier":"flagship","model":"Grok 4.7","input":2,"output":6},{"provider":"DeepSeek","tier":"flagship","model":"DeepSeek V4 Pro (peak)","input":1.32,"output":3.96},{"provider":"DeepSeek","tier":"small","model":"DeepSeek V4.1 Flash (peak)","input":0.3,"output":1.2},{"provider":"Mistral","tier":"flagship","model":"Mistral Medium 3.5","input":1.5,"output":7.5},{"provider":"Mistral","tier":"mid","model":"Mistral Large 3","input":0.5,"output":1.5}],"priceToCapability":[{"label":"GPT-3 quality (MMLU ~42)","start":"Nov 2021, $60/M","end":"Late 2024, $0.06/M (Llama 3.2 3B)","factor":1000,"months":36,"source":"https://a16z.com/llmflation-llm-inference-cost/"},{"label":"GPT-3.5 level (MMLU 64.8)","start":"Nov 2022, $20/M","end":"Oct 2024, $0.07/M","factor":280,"months":23,"source":"https://hai.stanford.edu/news/ai-index-2025-state-of-ai-in-10-charts"},{"label":"GPT-4 level (MMLU)","start":"Mar 2023, $37.50/M","end":"Feb 2025, $0.18/M","factor":200,"months":23,"source":"https://epoch.ai/data-insights/llm-inference-price-trends"},{"label":"75% on GPQA Diamond","start":"o3, ~$0.30/question","end":"GPT-5.6 Luna, $0.0004/question","factor":725,"months":18,"source":"https://epoch.ai/publications/the-plunging-price-of-thought"},{"label":"87.5% on ARC-AGI-1","start":"Dec 2024, o3 $4,560/task","end":"Aug 2026, Grok 4.6 $0.30/task","factor":15200,"months":20,"source":"https://r40.io/blog/arc-agi-verified-cost-per-task-price-step/"},{"label":"~27% on FrontierMath","start":"Apr 2025, o4-mini ~43M tokens","end":"Dec 2025, GPT-5.2 ~5M tokens","factor":8.6,"months":8,"source":"https://epoch.ai/gradient-updates/how-persistent-is-the-inference-cost-burden"}],"epochDecline":{"perQuarter":0.47,"perYear":13,"atDebutPerQuarter":0.66,"source":"https://epoch.ai/publications/the-plunging-price-of-thought"},"capex":{"years":[2023,2024,2025,2026],"companies":[{"name":"Microsoft","values":[28.1,44.5,64.6,190],"basis":"FY to June for 2023-2025 (cash additions to P&E); 2026 is the ~$190B calendar-2026 guide"},{"name":"Alphabet","values":[32.3,52.5,91.4,200],"basis":"Calendar; 2026 midpoint of $195-205B guidance"},{"name":"Amazon","values":[52.7,83,131.8,200],"basis":"Calendar, net of proceeds; 2026 'about $200B'"},{"name":"Meta","values":[28.1,39.2,72.2,137.5],"basis":"Calendar incl. finance-lease principal; 2026 midpoint of $130-145B"},{"name":"Oracle","values":[6.9,21.2,55.7,null],"basis":"FY to May, shifted one year (FY24 shown under 2023); FY27 guidance not captured"}],"source":"https://insight.factset.com/hyperscalers-tap-external-financing-as-ai-capex-outruns-cash-flow"},"metrHorizons":[{"model":"GPT-4 (0314)","date":"2023-03","minutes":3.5,"secondary":false},{"model":"GPT-4 (1106)","date":"2023-11","minutes":3.6,"ciLow":1.6,"ciHigh":7.5,"secondary":false},{"model":"Claude Sonnet 3.7","date":"2025-02","minutes":60,"ciLow":32,"ciHigh":106,"secondary":false},{"model":"o3","date":"2025-04","minutes":121,"ciLow":74,"ciHigh":201,"secondary":false},{"model":"Claude Opus 4","date":"2025-05","minutes":101,"ciLow":58,"ciHigh":170,"secondary":false},{"model":"GPT-5","date":"2025-08","minutes":214,"ciLow":117,"ciHigh":480,"secondary":false},{"model":"Claude Opus 4.5","date":"2025-11","minutes":320,"ciLow":170,"ciHigh":729,"secondary":false},{"model":"GPT-5.3-Codex","date":"2026-02","minutes":390,"secondary":true},{"model":"Claude Opus 4.6","date":"2026-02","minutes":870,"ciLow":360,"ciHigh":5880,"secondary":true}],"metrNotes":{"doublingDaysSince2024":88.6,"doublingDaysSince2023":130.8,"unreliableAboveHours":16,"source":"https://metr.org/blog/2026-1-29-time-horizon-1-1/"},"tokenVolumes":{"googleMonthlyTrillions":[{"date":"2024-05","value":9.7},{"date":"2025-05","value":480},{"date":"2025-07","value":980},{"date":"2026-05","value":3200}],"openRouterWeeklyTrillions":[{"date":"2025-11","value":5},{"date":"2026-05","value":25},{"date":"2026-09","value":134}],"sources":["https://blog.google/innovation-and-ai/sundar-pichai-io-2026/","https://www.investing.com/news/stock-market-news/openrouter-raises-113m-as-token-volume-surges-to-25t-weekly-93CH-4710030","https://www.teamday.ai/blog/top-ai-models-openrouter-2026"]},"openVsClosed":[{"model":"GLM-5.3 Flash","open":true,"blended":0.1,"index":42},{"model":"DeepSeek V4.1 Flash","open":true,"blended":0.2,"index":39},{"model":"MiMo-V2.6-Pro","open":true,"blended":0.2,"index":46},{"model":"GLM-5.3","open":true,"blended":0.9,"index":45},{"model":"Kimi K3","open":true,"blended":2.3,"index":44},{"model":"GPT-6 Sol","open":false,"blended":1.06,"index":48},{"model":"Muse Spark 1.3","open":false,"blended":1.6,"index":48},{"model":"GPT-6 Astra","open":false,"blended":3.26,"index":53},{"model":"Claude Opus 5.5 (max)","open":false,"blended":5.98,"index":58}],"openVsClosedSource":"https://artificialanalysis.ai/models/open-source","hostedOpen":[{"setup":"gpt-oss-120b, cheapest hosts (AkashML, CoreWeave, FlexAI)","input":0.03,"output":0.17,"note":"24 providers listed"},{"setup":"gpt-oss-120b, modal price (Together, Groq, Nebius, Azure, Bedrock, OpenRouter)","input":0.15,"output":0.6,"note":""},{"setup":"gpt-oss-120b, Cerebras","input":0.35,"output":0.75,"note":"~1,700-3,000 tok/s"},{"setup":"Llama 3.1 8B (Groq)","input":0.05,"output":0.08,"note":"Apr 2026"},{"setup":"Llama 3.3 70B (Groq)","input":0.59,"output":0.79,"note":"Apr 2026"},{"setup":"DeepSeek V4.1 Flash, official (peak / off-peak)","input":0.3,"output":1.2,"note":"off-peak $0.15 / $0.60"},{"setup":"DeepSeek V4 Pro, official (peak / off-peak)","input":1.32,"output":3.96,"note":"off-peak $0.66 / $1.98"},{"setup":"GLM-5.3 Flash (OpenRouter)","input":0.15,"output":0.5,"note":"#1 by weekly tokens, Sep 2026"},{"setup":"Datacenter TCO, 397B MoE on TPU v7 (SemiAnalysis)","input":null,"output":0.181,"note":"total $/M at 100 tok/s per user; B200 $0.222, B300 $0.276"},{"setup":"Owned RTX 4090, Qwen 3.8 27B, batched, 100% utilisation","input":null,"output":0.28,"note":"$1.03 at 20% utilisation"},{"setup":"Rented RTX 5090 ($0.76/h), Llama 3.1 8B, batch 32","input":null,"output":0.06,"note":""}],"selfHost":{"gpu":"RTX 4090","model":"Qwen 3.8 27B, 16 concurrent requests","hardwareUsd":2600,"years":3,"watts":300,"kwhUsd":0.17,"tokensPerSec":147,"referenceApiOutputUsd":1.7,"source":"https://paraloncloud.com/resources/own-gpu-vs-pay-per-token-break-even"},"enterprise":{"anthropicRunRate":[{"date":"2024-12","value":1},{"date":"2025-12","value":9},{"date":"2026-04","value":30},{"date":"2026-07","value":65}],"markers":[{"label":"Opus share of Claude Code sessions vs chat","value":"54% vs 10%","date":"Jun 2026","source":"https://www.anthropic.com/research/economic-index-june-2026-report"},{"label":"Fable 5 share of Anthropic tokens / spend on Ramp, one month after launch","value":"6% / 11.4%","date":"Jul 2026","source":"https://ramp.com/data/ai-index-august-2026"},{"label":"Top 1% of customers' share of revenue (OpenAI, Anthropic)","value":"80%","date":"2026","source":"https://www.pymnts.com/news/artificial-intelligence/2026/openai-and-anthropic-get-80percent-of-revenue-from-1percent-of-customers/"},{"label":"Open-source share of enterprise LLM usage","value":"19% (2024) → 11% (2025)","date":"2025","source":"https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/"},{"label":"US labs' share of OpenRouter tokens","value":"~70% (Jun 2025) → ~30% (Jun 2026)","date":"Jun 2026","source":"https://pro.stockalarm.io/blog/openrouter-llm-rankings-investor-analysis"},{"label":"Cursor Router cost per commit: Balance / Opus 4.8 / Fable 5","value":"$4.63 / $7.34 / $12.69","date":"Jul 2026","source":"https://cursor.com/blog/router"},{"label":"OpenRouter price elasticity","value":"+0.5-0.7% usage per 10% price cut","date":"2025","source":"https://openrouter.ai/state-of-ai"},{"label":"Enterprise deployments that are true agents","value":"16%","date":"2025","source":"https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/"}]},"changelog":[{"date":"2026-09-27","edition":2,"summary":"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.","changes":["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."]},{"date":"2026-09-27","edition":1,"summary":"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.","changes":["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."]}]}