Global · price of compute
What a million AI tokens costs: flagship prices are rising while capability gets cheaper
The leading labs' flagship models charge a median $12 per million output tokens, against $10 for their flagships a year earlier; the dearest, GPT-6 Astra, charges $50 (OpenRouter list prices, 2026-10-11). Open-weights models charge a median $0.65. Epoch AI estimates the price of a fixed level of capability falls about 13× a year.
Updated 11 October 2026 · data build 2026-10-10 · how we know · corrections
Flagship models charge a median $12 per million output tokens, 1.2× a year ago
List price per million tokens of each lab's top model, now against its top model on 2025-10-11. 7 labs with a flagship on OpenRouter at both dates. Measured
| Lab | Bloc | Flagship now | Released | In $/M | Out $/M | Flagship on 2025-10-11 | Out $/M then |
|---|---|---|---|---|---|---|---|
| OpenAI | US | GPT-6 Astra | 2026-09-04 | 10 | 50 | GPT-5 | 10 |
| Anthropic | US | Claude Fable 5.1 | 2026-09-01 | 10 | 50 | Claude Opus 4.1 | 75 |
| US | Gemini 3.1 Pro Preview | 2026-02-19 | 2 | 12 | Gemini 2.5 Pro | 10 | |
| xAI | US | Grok 4.7 | 2026-09-21 | 2 | 6 | Grok 4 | 15 |
| Mistral AI | Europe | Mistral Large 4 | 2026-10-06 | 0.68 | 2.09 | Mistral Medium 3 | 2 |
| DeepSeek | China | DeepSeek V4 Pro 0813 | 2026-08-12 | 0.66 | 1.98 | R1 | 2 |
| Moonshot AI | China | Kimi K3 | 2026-07-16 | 0.8 | 13.5 | Kimi K2 | 2.49 |
| Alibaba (Qwen) | China | – | – | – | – | Qwen3 Max | 6 |
A lab's flagship is its top-tier general model at the time (the tier the lab sells as most capable), as listed on OpenRouter. Batch, ':free', image, audio and 'pro'/'deep-think' add-on variants are excluded. Where a lab has two top tiers at once (Anthropic's Fable and Opus from June 2026), the dearer one is the flagship. Same rule for every lab.
Download CSV ↓Source: OpenRouter models and prices · Wayback Machine snapshots of the OpenRouter models API · as of 2026-10-11 · Compiled by StackPath
Flagship output prices a year apart: OpenAI $10 → $50, Anthropic $75 → $50, Google $10 → $12, xAI $15 → $6
Output price per million tokens of each lab's flagship when it launched, by quarter of release. Launch prices come from the first archived copy of OpenRouter's price list after release. Measured
* First price seen more than 45 days after release (no earlier archived copy).
Source: OpenRouter models and prices · Wayback Machine snapshots of the OpenRouter models API · as of 2026-10-11 · 32 snapshots since 2023-12-26 · Compiled by StackPath
Open-weights models cost a fraction: median $0.65 per million output tokens against $4.33 for closed models
Median output price of models released each quarter, open-weights (downloadable) against closed (API only), every paid text model OpenRouter has listed. Measured
| Quarter | Closed models | Closed median out $/M | Open models | Open median out $/M |
|---|---|---|---|---|
| 2024-Q1 | 26 | 1.38 | 2 | 4.65 |
| 2024-Q2 | 35 | 0.81 | 17 | 0.65 |
| 2024-Q3 | 24 | 0.8 | 24 | 0.4 |
| 2024-Q4 | 24 | 5.5 | 15 | 0.5 |
| 2025-Q1 | 30 | 1.2 | 36 | 0.25 |
| 2025-Q2 | 35 | 2.5 | 42 | 0.36 |
| 2025-Q3 | 26 | 2 | 49 | 0.75 |
| 2025-Q4 | 27 | 8 | 38 | 0.5 |
| 2026-Q1 | 30 | 5.25 | 19 | 0.5 |
| 2026-Q2 | 37 | 6.24 | 22 | 1 |
| 2026-Q3 | 69 | 6 | 32 | 0.49 |
| 2026-Q4 | 5 | 2.7 | 0 | – |
Source: OpenRouter models and prices · Wayback Machine snapshots of the OpenRouter models API · as of 2026-10-11 · developer bloc by headquarters, from /global/ai-model-rankings · Compiled by StackPath
The price of fixed capability falls about 13× a year, Epoch AI estimates
Paraphrased from Epoch AI (CC BY 4.0), with the price cuts we see on the same model in OpenRouter's archived price lists. Estimate
| Model | Seen on | Out $/M before | Out $/M after | Cut |
|---|---|---|---|---|
| Gemma 2 9B | 2025-08-14 | 0.2 | 0.01 | −95% |
| Qwen3 235B A22B Thinking 2507 | 2026-06-16 | 1.5 | 0.1 | −93% |
| Qwen2.5 VL 32B Instruct | 2025-08-14 | 0.9 | 0.08 | −91% |
| Mistral Small | 2024-09-18 | 6 | 0.6 | −90% |
| GPT-5.6 Luna Pro | 2026-08-14 | 6 | 0.6 | −90% |
| DeepSeek V4 Pro | 2026-10-11 | 3.2 | 0.42 | −87% |
| Qwen2.5 Coder 7B Instruct | 2025-05-13 | 0.2 | 0.03 | −85% |
| Qwen2.5 7B Instruct | 2025-01-17 | 0.27 | 0.05 | −82% |
| Mistral Small 3.1 24B | 2026-04-10 | 0.56 | 0.11 | −80% |
| Gemini Flash 1.5 (preview) | 2024-09-18 | 0.75 | 0.15 | −80% |
| o3 | 2025-06-16 | 40 | 8 | −80% |
| Magistral Small 2506 | 2025-07-14 | 1.5 | 0.3 | −80% |
Same model id, lower list price in a later snapshot; the date is the first snapshot showing the new price. For open-weights models a cut can mean a cheaper host became OpenRouter's default.
Source: Epoch AI, The plunging price of thought (CC BY 4.0) · OpenRouter models and prices · Wayback Machine snapshots of the OpenRouter models API · as of 2026-09-22 · Compiled by StackPath
A token routed through OpenRouter cost about $0.41 per million at list prices; Chinese models carried 66% of tokens and 45% of the value
Tokens routed in the 30 days to 2026-10-11 by model (OpenRouter rankings, CC BY 4.0) multiplied by each model's list price. 89% of tokens matched a priced model; 98% of them were input tokens. Calculated
| Developer bloc | Tokens (trillion) | Share of tokens | List-price value ($m) | Share of value | $ per million |
|---|---|---|---|---|---|
| US | 185.9 | 32.6% | 127 | 54.9% | 0.683 |
| China | 373.9 | 65.5% | 103.4 | 44.7% | 0.277 |
| Other | 7.3 | 1.3% | 0.7 | 0.3% | 0.091 |
| Europe | 1.6 | 0.3% | 0.2 | 0.1% | 0.117 |
| Unclassified | 1.8 | 0.3% | 0.2 | 0.1% | 0.092 |
| Model | Bloc | Open | Tokens (T) | In $/M | Out $/M | $ per M routed |
|---|---|---|---|---|---|---|
| DeepSeek: DeepSeek V4.1 Flash (batch) | China | yes | 100.5 | 0.112 | 0.336 | 0.116 |
| Z.ai: GLM 5.3 Flash (batch) | China | yes | 55.1 | 0.06 | 0.2 | 0.063 |
| Tencent: Hy4 preview | China | yes | 40.6 | 0.751 | 2.251 | 0.764 |
| OpenAI: GPT-5.6 Luna (batch) | US | no | 33.9 | 0.1 | 0.6 | 0.109 |
| DeepSeek: DeepSeek V4 Flash 0731 | China | yes | 32.2 | 0.018 | 0.64 | 0.045 |
| Xiaomi: MiMo-V2.6-Flash | China | yes | 25.1 | 0.14 | 0.28 | 0.142 |
| NVIDIA: Nemotron 3 Ultra (free) | US | yes | 23.1 | 0 | 0 | 0 |
| OpenAI: GPT-6 Luna (batch) | US | no | 15.4 | 0.05 | 0.25 | 0.056 |
| DeepSeek: DeepSeek V4 Flash 0423 | China | yes | 14.7 | 0.03 | 1.28 | 0.101 |
| Xiaomi: MiMo-V2.5 | China | yes | 14.6 | 0.14 | 0.28 | 0.141 |
| Tencent: Hy3 | China | yes | 13.2 | 0.083 | 0.33 | 0.087 |
| Z.ai: GLM 5.3 (batch) | China | yes | 12.7 | 0.45 | 2 | 0.475 |
| Google: Gemini 3.8 Flash (batch) | US | no | 9 | 0.375 | 1.875 | 0.421 |
| Meta: Muse Spark 1.3 Contributor | US | no | 7.9 | 0.1 | 0.2 | 0.102 |
| Anthropic: Claude Opus 5.5 (batch) | US | no | 7.9 | 2 | 10 | 2.096 |
Source: OpenRouter rankings (CC BY 4.0) · OpenRouter models and prices · as of 2026-10-11 · Compiled by StackPath
Mostly input: agents resend long contexts, so the cheap input price dominates and the average routed token costs far less than the output prices above. OpenRouter is one router, used mainly by developers and price-sensitive apps; it excludes ChatGPT, Claude.ai, Gemini and direct API traffic, so this is not market-wide spend.
Our reading. "AI is getting cheaper" is true for a fixed task and false for the frontier. The labs at the top are pricing their newest models up (GPT-6 Astra $50, Claude Fable 5.1 $50 per million output tokens), while open-weights models, mostly Chinese, sell capability that was frontier a year or two ago for under a dollar. Rising GPU rental prices (GPU prices) and the cost of the debt behind the chips (cost of GPU debt) may be one reason flagship prices are not falling; we have not tested that link.
List prices per million tokens on OpenRouter, which usually matches the lab's own API price for closed models; open-weights models are served by many hosts and OpenRouter shows the price of its default (often cheapest) route. Output tokens cost several times input tokens; reasoning models also spend hidden 'thinking' tokens billed as output, so a cheaper token can still mean a dearer answer. Launch prices come from the first Wayback snapshot after a model appeared (marked exact when within 45 days of release); before 2024 and between snapshots the record is incomplete. The usage-weighted figure multiplies the last 30 days of OpenRouter's routed tokens by today's list prices (cached input at the cache-read price where OpenRouter gives one; ':free' variants at zero). OpenRouter is one router and over-represents developers and price-sensitive traffic; it is not market-wide spend. Open-weights = the model has a Hugging Face id in OpenRouter's catalogue. Epoch's falling price of fixed capability is an estimate across five benchmarks, not a price list.
Questions people ask
- How much does a million tokens cost?
- It ranges from a few cents to $600 per million output tokens across the 343 paid text models on OpenRouter (2026-10-11). Open-weights models charge a median $0.2 in and $0.65 out; closed models $1.15 in and $4.33 out. A million tokens is roughly 750,000 English words.
- What is the most expensive AI model per token?
- Among lab flagships, GPT-6 Astra at $10 in / $50 out and Claude Fable 5.1 at $10 in / $50 out per million tokens (OpenRouter, 2026-10-11). 'Pro' variants that think longer cost more still.
- Are AI prices going up or down?
- Both. The flagships of OpenAI, Anthropic, Google, xAI, Mistral AI, DeepSeek, Moonshot AI charge a median $12 per million output tokens against $10 for their flagships a year earlier. But the price of a fixed level of capability falls fast: Epoch AI estimates about 47% a quarter (13× a year) since 2023, because cheaper models catch up with last year's best.
- Are Chinese AI models cheaper than American ones?
- On list price, yes. The US labs' flagships charge a median $31 per million output tokens, the Chinese labs' $7.74, Mistral's Mistral Large 4 $2.09. On OpenRouter in the last 30 days, Chinese models carried 66% of routed tokens but 45% of their list-price value.
- What is the difference between input and output tokens?
- Input tokens are what you send the model (the prompt, documents, chat history); output tokens are what it writes back. Output costs several times more because each output token takes a full pass through the model. Reasoning models also bill their hidden 'thinking' as output tokens.