Tracker
The AI token price index
Nine AI labs' prices per million tokens for a budget model, a flagship and, where they sell one, a premium tier, updated monthly. The gaps between tiers matter more to a budget than the gaps between labs.
Flagship models, median
$3
blended, per million tokens, 9 labs
Budget models, median
$0.25
8 labs sell a budget tier
Cheapest flagship
$0.99
DeepSeek; 2 labs also sell a premium tier
BudgetFlagshipPremium
View the data
| Lab | Tier | Model | In / out | Blended |
|---|---|---|---|---|
| OpenAI | Budget | OpenAI: GPT-6 Luna | $0.10 / $0.50 | $0.20 |
| Flagship | OpenAI: GPT-6.1 Sol | $2 / $10 | $4 | |
| Premium | OpenAI: GPT-6 Astra | $10 / $50 | $20 | |
| Anthropic | Budget | Anthropic: Claude Haiku 5.5 | $0.10 / $0.50 | $0.20 |
| Flagship | Anthropic: Claude Sonnet 5.5 | $2 / $10 | $4 | |
| Premium | Anthropic: Claude Fable 5.1 | $10 / $50 | $20 | |
| Budget | Google: Gemini 3.5 Flash Lite | $0.30 / $2.50 | $0.85 | |
| Flagship | Google: Gemini 4 Argon | $2 / $10 | $4 | |
| Moonshot (Kimi) | Flagship | MoonshotAI: Kimi K3.1 Pro | $2 / $10 | $4 |
| xAI | Budget | SpaceXAI: Grok 4.3 (previous generation) | $1.25 / $2.50 | $1.56 |
| Flagship | SpaceXAI: Grok 4.7 | $2 / $6 | $3 | |
| Alibaba (Qwen) | Budget | Qwen: Qwen3.8 Flash | $0.15 / $0.47 | $0.23 |
| Flagship | Qwen: Qwen3.8 Max (0902) | $2 / $6 | $3 | |
| Mistral | Budget | Mistral: Mistral Small 4 | $0.15 / $0.60 | $0.26 |
| Flagship | Mistral: Mistral Medium 3.5 | $1.50 / $7.50 | $3 | |
| Zhipu (GLM) | Budget | Z.ai: GLM 5.3 Flash | $0.15 / $0.50 | $0.24 |
| Flagship | Z.ai: GLM 5.3 | $1.40 / $4.40 | $2.15 | |
| DeepSeek | Budget | DeepSeek: DeepSeek V4.1 Flash 1017 | $0.15 / $0.60 | $0.26 |
| Flagship | DeepSeek: DeepSeek V4 Pro 0813 | $0.66 / $1.98 | $0.99 |
What it means for a budget
- An AI budget line set in March at March prices now sits about 40% too high on price alone. Treat that underspend as a reserve for volume yet to come, not a saving.
- Use is shifting to cheaper models, which cuts cost per head before any price falls. One annual total hides which effect is winning.
- A multi-year committed-spend contract agreed at this quarter's rate gives away the deflation the buyer was about to collect.
The argument in full: Budgeting for deflation.
Method and sources
Lab prices come from OpenRouter, fetched 25 October 2026. OpenRouter can route a model through third-party hosts, and large buyers agree discounts that no list records. Blended prices weight input three to one against output. I pick each lab's models by hand and review them monthly. The flagship is the lab's main general model, the budget model its cheapest current one, and the premium tier a stronger model sold well above the flagship. A faster serving of the same model does not count as premium. Every price was checked against the lab's own pricing page on 25 October 2026, and where OpenRouter differed (Zhipu's GLM-5.3) the index uses the lab's price. Several labs charge more for very long prompts, so the index shows the standard rate. DeepSeek charges double during weekday peak hours, and the index shows the off-peak rate. Qwen prices use Alibaba's international (Singapore) rates. What companies pay: Ramp spending data, reported by Fortune, 9 September 2026. Long-run series: Guido Appenzeller, “LLMflation”, Andreessen Horowitz. Updated monthly.