A month ago the frontier tier cost 10 dollars per million input tokens and 50 per million output, at both of the labs that had one. On 1 October 2026 that is no longer where the market is, and the move happened at three labs inside four weeks.
We priced Moonshot's model against Anthropic's in what tokens actually cost on 1 September. Enough has changed since that the conclusion needs redoing rather than updating, so this is the whole board: every current frontier and near-frontier model, the same two workloads costed against each, and the three things the table does not tell you.
1. The board
All figures are per million tokens, taken from the providers rather than from aggregators, and current on 1 October 2026. Anthropic's are from their published pricing. Google's are from the Gemini 4 Argon announcement. OpenAI's and Moonshot's are as reported in our notes on GPT-6.1 Sol and Kimi K3.
| Model | Input | Output | Cache read | Cache read as % of input |
|---|---|---|---|---|
| Claude Fable 5 | $10.00 | $50.00 | $1.00 | 10% |
| GPT-6 Astra | $10.00 | $50.00 | not published | n/a |
| Claude Fable 5.1 | $10.00 | $50.00 | $0.25 | 2.5% |
| Claude Opus 5 | $5.00 | $25.00 | $0.50 | 10% |
| Claude Opus 5.5 | $4.00 | $20.00 | $0.20 | 5% |
| Gemini 4 Argon, standard | $4.00 | $20.00 | $0.20 | 5% |
| Kimi K3 | $3.00 | $15.00 | $0.30 | 10% |
| Claude Sonnet 5.5 | $2.00 | $10.00 | $0.20 | 10% |
| GPT-6.1 Sol | $2.00 | $10.00 | $0.10 | 5% |
| Gemini 4 Argon, introductory | $2.00 | $10.00 | $0.10 | 5% |
| Claude Haiku 4.5 | $1.00 | $5.00 | $0.10 | 10% |
Two rows in that table are not similar. They are identical.
Gemini 4 Argon at its standard price and Claude Opus 5.5 have the same price list. Four dollars in, twenty out, twenty cents to read from cache. Every rate, to the cent.
Gemini 4 Argon at its introductory price and GPT-6.1 Sol have the same price list. Two dollars in, ten out, ten cents cached. Again every rate.
Three labs, three independent pricing decisions, two exact collisions. That is not coordination, and we are not suggesting it is. It is what a commodity price looks like when it arrives: the number stops being a positioning statement and starts being what the thing costs to serve plus a margin everyone has worked out the same way.
2. What it costs you, on two real workloads
A price list is not a bill. Here are the same ten models against two workload shapes, which between them cover most of what anyone actually runs.
The assistant. Fifty million input tokens and five million output per month, with a fifth of the input served from cache. An internal knowledge assistant, a support deflection surface, a search front end.
The agent. Five hundred million input tokens and twenty-five million output, with ninety per cent of the input served from cache. A coding agent or an operations agent resends most of its context on every turn, so it reads from cache far more than it writes anything new. This is the shape that surprises people.
| Model | Assistant / month | Agent / month |
|---|---|---|
| Claude Haiku 4.5 | $66.00 | $220.00 |
| GPT-6.1 Sol | $131.00 | $395.00 |
| Gemini 4 Argon, introductory | $131.00 | $395.00 |
| Claude Sonnet 5.5 | $132.00 | $440.00 |
| Kimi K3 | $198.00 | $660.00 |
| Claude Opus 5.5 | $262.00 | $790.00 |
| Gemini 4 Argon, standard | $262.00 | $790.00 |
| Claude Opus 5 | $330.00 | $1,100.00 |
| Claude Fable 5.1 | $652.50 | $1,862.50 |
| Claude Fable 5 | $660.00 | $2,200.00 |
Three things fall out of that which the price list hides.
Claude Fable 5.1 and Claude Fable 5 are a rounding error apart on the assistant and 338 dollars apart on the agent. Same input price, same output price, identical tier. The entire difference is the cache read rate, 25 cents against a dollar, and it only shows up once the workload reads from cache at volume. If you looked at the two headline rates you would conclude there was no difference at all.
The successor is cheaper than its predecessor. Claude Opus 5.5 costs less than Claude Opus 5 on every rate: four against five, twenty against twenty-five, twenty cents against fifty. A newer model priced below the one it replaces, in the same tier, is unusual. It says the cost of serving moved, not the positioning.
Kimi K3 has become mid-priced without changing its price. When we audited it on 1 September it sat above Claude Sonnet 5 and the point of that note was that the cheap Chinese model story had stopped being true at the frontier. It still costs 3 dollars and 15 dollars. The board moved around it, and it is now the sixth cheapest row on a table of ten.
3. The rate nobody leads with
Look again at the right-hand column of the board. Until recently a cache read cost ten per cent of the input rate, everywhere, as a convention. It no longer does.
- Claude Fable 5.1 reads cache at 2.5% of its input price.
- Claude Opus 5.5, GPT-6.1 Sol and Gemini 4 Argon all read at 5%.
- Everything else is still at 10%.
Three labs deepened the same discount inside one quarter, on the rate that nobody puts in a launch headline.
That is not an accident of accounting. It is the rate that decides an agent's bill. An agent resends its working context on every turn, so a long task is overwhelmingly cache reads: in our agent workload above, 450 million of the 500 million input tokens. Dropping that rate from 10% to 5% of input is the single largest discount any of these labs has issued this year, and it is aimed squarely at the workload they all want.
We made the narrow version of this point reading GPT-6.1 Sol in the model that was cancelled, and again reading Argon in the output limit is the story. It generalises: if you run agents, the cache read rate and the output rate are your price list, and the input rate is close to decorative.
The practical consequence for anyone picking a model: get your prompt prefix stable and cacheable before you do anything else. A system that invalidates its cache on every request, because a timestamp or a request id sits near the front of the prompt, pays the input rate on everything and throws away the discount that the entire industry just spent a quarter competing on.
4. The thing that invalidates this entire table
Every figure above is a price per token. Two prices per token are only comparable if a token means the same amount of text on both sides of the comparison. That is an assumption, it is almost never stated, and it is false at least once inside this table.
Anthropic publish it plainly. Their documentation says that Claude 4.7 and later models use a newer tokenizer which "produces approximately 30% more tokens for the same text" than their earlier models. That is not a criticism of the tokenizer, which they attribute to better performance. It is a statement that you cannot compare the per-token price of Claude Sonnet 4.6 to Claude Sonnet 5.5 by reading the two numbers, because the same document is a different number of tokens in each.
Apply the published 1.3 factor to our agent workload and the ordering moves:
| Model | Nominal | Adjusted for the same text |
|---|---|---|
| Claude Sonnet 5.5 | $440.00 | $572.00 |
| Claude Opus 5.5 | $790.00 | $1,027.00 |
| Claude Fable 5.1 | $1,862.50 | $2,421.25 |
The Opus 5.5 and Argon rows, which are the same to the cent on the price list, are not the same at all once one side is counting about thirty per cent more tokens for the same work.
And here is the honest limit of that correction. The 1.3 figure is Anthropic's, and it compares their own models to their own earlier models. It is not a cross-lab ratio. No lab publishes how its tokenizer compares to another's, so nobody can size the error term between the Argon row and the Opus 5.5 row, including us. We have shown the adjusted figures because the direction is documented and the effect is large; treat the exact numbers as an illustration of the size of the problem rather than as a corrected answer.
Which is the real finding of this paper: a cross-lab price-per-token table carries an unquantified error term, and every one you will read, this one included, presents itself as more precise than it is.
5. Two discounts that are real and routinely ignored
Batch. Anthropic discount batch processing by 50% on both input and output. That puts Claude Sonnet 5.5 at 1 dollar and 5 dollars, and Claude Opus 5.5 at 2 and 10, for work that does not need an answer in the next few seconds. An enormous amount of what gets run synchronously has no human waiting on it: overnight enrichment, backfills, evaluation runs, report generation. Those are batch workloads being billed as interactive ones.
Introductory pricing is not pricing. Gemini 4 Argon's 2 dollars and 10 dollars is scheduled to become 4 and 20. A rate that is already announced to double is not a rate you can build a gross margin on, and modelling your unit economics against it is how a product launches with a cost base that doubles on a date somebody else chose. Use the standard rate, and treat the introductory period as a discount you happen to be receiving.
6. So which one
We are not going to give you a ranking, because a ranking of these ten would be a ranking of price lists and section 4 is an argument that price lists do not rank.
What we do instead, and what we would advise anyone to do:
Judge cost per completed task, not cost per request. A cheaper model that needs three turns where another needed one is not cheaper, it is more expensive and slower. This is the single most common error we see, and it survives because per-request cost is easy to look up and per-task cost has to be measured.
Measure on your own traffic. Take fifty real requests, run them on the two or three candidates, count the tokens each actually consumed and whether the job got done. That exercise takes an afternoon and beats every table on the internet, this one included.
Fix the cache before you pick a model. Moving from a 10% cache read rate to a 5% one is worth less than moving from a 0% cache hit rate to a 90% one, and the second is in your control today, on whichever model you are already using.
Then pick the cheapest model that passes your bar, and re-run the exercise when the board moves, which on the evidence of the last four weeks is roughly monthly.
This is the same argument we make to founders in AI SaaS development: cost per action and architecture are the same decision, and you want to have made it before the first line of code rather than after the first invoice. It is also why AI implementation is an architecture decision: where the model sits in your system moves your bill by more than which model you chose.
7. Sources
- Anthropic, Claude API pricing, read 1 October 2026. Every Claude figure above, including the per-model cache read multipliers and the tokenizer statement in section 4, is from this page.
- Koray Kavukcuoglu, Google DeepMind, Gemini 4 Argon, 30 September 2026, for Argon's introductory and standard rates and the 95% cache discount. Read at length in the output limit is the story.
- OpenAI's GPT-6.1 Sol rates, as reported by TechCrunch in OpenAI launches GPT-6.1 Sol, says it nearly matches GPT-6 Astra and costs less, 29 September 2026, and set out in the model that was cancelled.
- Moonshot AI's Kimi K3 rates, audited in what tokens actually cost, with the architecture behind them in the Kimi K3 technical report and our reading of it in where the compute went.
- Anthropic's prompt caching documentation, for the cache write and read mechanics behind the rates in section 3.
The arithmetic in sections 2 and 4 is ours and is reproducible from the rates in section 1: two workload shapes, multiplied out, nothing else in it. The workload shapes are illustrative and chosen to be recognisable, not measured from any particular customer. We have not run these ten models against a common task and the comparison in section 6 is advice, not a result. Prices at this tier have moved three times in the four weeks before publication, so check them rather than trusting a table with a date on it.
This note sits in our Research track, alongside the operations and foundation models threads. The full library is at Research.