What AI models really cost
Compare AI models by price per capability point, not by list price per million tokens. As of 24 October 2026, our AI Rating API shows Claude Sonnet 5.5 scoring 80 at a blended $4.00 per million tokens, while no model in the most expensive tier reaches 80, and three models scoring between 66.1 and 67.1 range from $0.54 to $8.00. Set a capability bar per task, route to the cheapest model that clears it, and check again every month.
AI pricing pages answer the wrong question. They tell you what a million tokens costs. What you need to know is what a given level of capability costs, and which model is the cheapest one that is good enough for the job in front of you.
We run a public product that answers exactly that: the AI Rating API, which reconciles OpenRouter’s catalog, the Artificial Analysis Intelligence Index and LM Arena every morning. Every figure below comes from its snapshot published at 06:00 UTC on 24 October 2026. The score measures capability only, on a fixed 0–100 scale, so a 66 today is comparable to a 66 next month. Prices are OpenRouter list prices in US dollars per million tokens. “Blended” is the API’s own tiering formula, (3 × input + output) ÷ 4, rounded to the cent. LM Arena was unavailable that morning; it only breaks ties, so no score here depends on it.
List price is the wrong unit
Here are eight text models from that snapshot, sorted by score:
| Model | Tier | Score | In / out ($/M) | Blended ($/M) |
|---|---|---|---|---|
| Claude Opus 5.5 | high | 82.3 | 4 / 20 | 8.00 |
| Claude Sonnet 5.5 | high | 80 | 2 / 10 | 4.00 |
| Claude Fable 5.1 | ultra | 76.3 | 10 / 50 | 20.00 |
| Claude Opus 5 | high | 72.6 | 5 / 25 | 10.00 |
| GPT-5.6 Sol | high | 67.1 | 4 / 20 | 8.00 |
| Grok 4.7 | high | 66.3 | 2 / 6 | 3.00 |
| MiMo-V2.6-Pro | low | 66.1 | 0.435 / 0.87 | 0.54 |
| Claude Haiku 5.5 | low | 62 | 0.10 / 0.50 | 0.20 |
Three things stand out.
The most expensive tier is not the most capable. No model in the API’s ultra tier (blended above $12) reaches 80 today. Its leader, Claude Fable 5.1, scores 76.3 at a blended $20.00. Claude Sonnet 5.5 scores 80 at $4.00, a fifth of the price.
The same capability can cost almost 15 times more. MiMo-V2.6-Pro, Grok 4.7 and GPT-5.6 Sol all score between 66.1 and 67.1. Their blended prices are $0.54, $3.00 and $8.00.
Newer is often cheaper and better. Claude Opus 5, released 24 July 2026, scores 72.6 at $5 / $25. Claude Opus 5.5, released 22 September 2026, scores 82.3 at $4 / $20.
And the list price is not even one number. Anthropic charges $0.10 / $0.50 for Claude Haiku 5.5 on prompts up to 100,000 tokens and $0.50 / $2.50 above that, and its Batch API halves both rates. Your real price depends on prompt length, caching and batching, which is one more reason to stop comparing sticker prices.
Price per point, within a tier
The API keeps price out of the score on purpose: tiers handle cost, the score measures capability, so a cheap, weak model can never outrank a strong one. The useful comparison is therefore the gap inside a tier. In the high tier, Claude Opus 5.5 costs twice as much as Claude Sonnet 5.5 for 2.3 more points. In the low tier, MiMo-V2.6-Pro buys 4.1 points over Claude Haiku 5.5 for about 2.7 times the blended price. Sometimes those points are worth paying for. Often they are not, and you only find out by asking the question task by task.
The cheapest adequate model per task
Pick the bar first, then the model. Set a minimum score for each task class from your own evaluations, then take the cheapest model that clears it. Here is that rule applied to today’s full board of 172 scored text models, with illustrative bars:
- Score of 60 or more: Claude Haiku 5.5, at $0.20 blended.
- 65 or more: MiMo-V2.6-Pro, at $0.54.
- 75 or more: a three-way tie at $4.00 between Claude Sonnet 5.5, Kimi K3.1 Pro and Gemini 4 Argon.
- 80 or more: Claude Sonnet 5.5, at $4.00. The only other model that qualifies is Claude Opus 5.5, at $8.00.
These bars are examples, not a claim about what any task needs. The index is a weighted blend of agentic, general, coding and scientific-reasoning evaluations. It is a screening tool, and a model can beat its score on a narrow task. The method is what matters: a bar per task, then the cheapest model that clears it.
The ranking moves every month
Whatever you pick will not stay the right answer for long. Eight of the ten models in the high tier were released on or after 21 September 2026. Kimi K3.1 Pro, released 23 October, ranked third the next morning. GPT-6 Sol (22 September) scores 68; GPT-6.1 Sol, released a week later at the same $2 / $10, scores 74. DeepSeek V4.1 Flash 1017 (17 October) scores 8 points above DeepSeek V4.1 Flash (10 September) at nearly the same blended price.
The far end of the board shows the cost of not looking. OpenAI’s o1-pro still lists at $150 / $600 and scores 17.7. Mercury 2.5 scores 17.6 at $0.04 / $0.15: for essentially the same score, o1-pro costs 4,000 times as much per output token. A model name hard-coded last year can keep billing at last year’s rates long after the market has moved on.
What this means for your team
Three controls turn this into policy:
- An approved-model list keyed by task, not by model name. Each task class gets a minimum score and a price ceiling. The model that meets both is resolved from data on a schedule, not from a wiki page someone wrote in the spring.
- Routing by task. Each request class goes to the cheapest model that clears its bar. The flagship is an exception with a reason attached, not the default.
- Spend caps. Budgets per team or per key, enforced in the request path, so a runaway loop or a misrouted batch job fails fast instead of arriving as an invoice.
A list that lives in a document is a suggestion. It becomes a control only when the gateway refuses a model that is not on it, and a cap becomes a control only when the call over budget fails. That enforcement is what AI guardrails are.
Our AI Guardrails engagement builds these controls as code in your accounts: the approved-model list, task routing and spend caps, enforced at your gateway and easy to update when next month’s data moves. The data behind this note is public, and how the AI Rating API works is documented on our site.