From 22 September to 7 October 2026 Anthropic, OpenAI, Google and Mistral released models with lower list prices or launch prices. If you hold a quote for an automation or a chatbot, redo the cost per task before you sign: two of the new prices are launch prices, and the real cost moves with caching, long prompts and volumes.
What changed in the price lists
- On 22 September Anthropic released Claude Opus 5.5 at $4 for input and $20 for output, against $5 and $25 for Opus 5: 20% less at list price. Anthropic says the cost of running it falls by about 40% on typical workloads, because the model costs less per token and uses fewer tokens for each task.
- The same day OpenAI released GPT-6 Sol at $2 and $10 and GPT-6 Luna at $0.10 and $0.50. GPT-5.6 Sol costs $4 and $20 at a promotional price, so GPT-6 Sol costs half.
- On 28 September Anthropic released Claude Sonnet 5.5 ($2 and $10) and on the 29th OpenAI released GPT-6.1 Sol, at the same price as GPT-6 Sol but with cache reads at $0.10 instead of $0.20.
- On 30 September Google announced Gemini 4 Argon at $2 and $10 as an introductory price, rising to $4 and $20 when the period ends, with no end date given. At the announcement it was rolling out to "trusted cyber defenders" in the Fairwind Program, and on 7 October the Gemini API pricing page did not list it.
- On 6 October Mistral opened Large 4 in public preview at $1.36 and $4.18, with a 50% launch discount ($0.68 and $2.09) for two weeks. On 7 October Anthropic released Claude Haiku 5.5 ($0.10 and $0.50 up to 100,000 prompt tokens) and halved the cache-read price of Sonnet 5.5, from $0.20 to $0.10.
Launch price deadlines
According to the Gemini API pricing page, Gemini 3.8 Flash goes from $0.75 to $1.50 for input on 1 January 2027. For Large 4 Mistral's changelog says two weeks without dates: counting from 6 October, the discount would end around 20 October (our calculation). An announced increase does not always arrive: on 10 August Anthropic made the introductory price of Claude Sonnet 5 the standard price and cancelled the increase planned for 1 September.
What a thousand tasks cost
| Model | Dollars per million tokens (input / output) | Dollars for 1,000 tasks | Condition |
|---|---|---|---|
| GPT-6 Luna | 0.10 / 0.50 | 0.45 | List price |
| Mistral Large 4 | 1.36 / 4.18 | 4.81 | Preview; with the discount 2.41 |
| GPT-6 Sol, GPT-6.1 Sol, Claude Sonnet 5.5 | 2 / 10 | 9.00 | List price |
| Gemini 4 Argon | 2 / 10, then 4 / 20 | 9.00, then 18.00 | Introductory price |
| Claude Opus 5.5 | 4 / 20 | 18.00 | List price |
| Claude Fable 5.1 | 10 / 50 | 45.00 | List price |
The cost of 1,000 tasks is our own calculation: it assumes 2,000 input tokens and 500 output tokens a task, at list price, with no caching, batch or discounts [illustrative estimate, to be confirmed against your own project]. With 20,000 tasks a month (40 million input tokens and 10 million output tokens), the same assumption gives about $9 with GPT-6 Luna, $360 with Opus 5.5 ($450 with Opus 5, so $90 less) and $900 with Fable 5.1 [illustrative estimate, to be confirmed against your own project]. The models are not equivalent: quality on your task matters.
How to redo the sums before you sign
- Count the tokens on a sample of real requests, with every model you are considering: the same text produces different token counts, and Anthropic says about 30% more for models from version 4.7 onwards than for earlier ones.
- Work out the cost at the launch price and at the standing price; if the vendor does not say how long the first lasts, decide on the second.
- Add the list-price surcharges. At Anthropic a 5-minute cache write costs 1.25 times the input price (Opus 5.5: $5 against $4). At OpenAI, prompts above 272,000 tokens cost twice the input price and 1.5 times the output price (GPT-6 Sol: $4 and $15), and regional processing costs 10% more on models released from 5 March 2026.
- Split the tasks by difficulty and urgency: a cheap model for high volumes, a more capable one for the hard cases, and batch (50% less at Anthropic) for whatever can wait.
- Add up the costs that are not tokens: human review, monitoring, integration and maintenance. If one person reviews 1% of the answers of a service handling 20,000 requests a month, that is 200 reviews of 3 minutes, or 10 hours a month [illustrative estimate, to be confirmed against your own project].
- Write into the quote the token volume it is based on and how reductions and increases in list prices are shared.
- Try a second model on the same examples and plan the migration: on 30 September Anthropic announced the retirement of Claude Sonnet 4.5 on 30 November 2026, and for moving from Sonnet 5 to Sonnet 5.5 it lists five changes that can break code.
Where we start
Cost and answer-quality control is part of our AI automation and chatbot service. If you ask us to look at a quote that depends on a model, we start from the cost per task described above, with a sample of your real cases, two price scenarios and a second model already tried.
A quote also includes the work of building the solution. Estimating that work is the job of Metis, our software effort estimation engine: from a requirement and a little structural data it returns a median estimate, a calibrated 90% confidence interval and the most similar historical projects. It is in private pilot.
If the requests to your assistant repeat, Kmemo can reduce the calls to the model. It is our open source Kotlin library (Apache-2.0) for semantic caching of LLM calls: it recognises requests already answered even when rephrased, and rejects cases where similarity is high but the right answer differs, such as "convert 100 USD" and "convert 250 USD". It saves money only on requests that repeat, and we check that on your sample before it goes into the quote.
We cannot predict list prices, which is why the quote has to say how reductions and increases are shared.
What we do not know yet
- When Gemini 4 Argon will open to paying API customers: Google speaks of an imminent opening, with no date.
- What happens to the price of GPT-5.6 Sol after 21 November: OpenAI's page says only that the promotion stays available until at least that date.
- How far Anthropic's 40% for Opus 5.5 holds on your task: it refers to typical workloads and default settings.
If you want us to look at the quote for an automation or a chatbot, you will find the service under AI automation and chatbots and you can write to us.
Sources
- OpenAI: API pricing
- OpenAI: API changelog
- Anthropic: Pricing
- Anthropic: Release notes
- Anthropic: Introducing Claude Opus 5.5 (22 September 2026)
- Google: Gemini 4 Argon: our next era of frontier intelligence (30 September 2026)
- Google AI for Developers: Gemini API pricing
- Mistral AI Docs: Mistral Large 4
- Mistral AI Docs: Changelog
