Gemini 4 Argon, GPT-6.1 Sol, Claude Sonnet 5.5: the same $2/$10 list price, nearly seven times apart per task

Suppose you are choosing an AI model for your company's internal assistant in mid-October 2026. You open the pricing pages of the three big labs and notice something unusual: their three newest mid-range models have identical list prices — $2 per million input tokens and $10 per million output tokens.

(A token is the unit a model uses to read and write text — roughly one short English word. AI models are priced by the number of tokens going in and coming out.)

Same price, so pick the strongest? The question sounds reasonable, but the independent numbers published this week show it is aimed at the wrong thing. The three models cost the same per token, but the cost of finishing a task differs by nearly seven times. The two most-cited independent leaderboards disagree about who is on top. And one of the three prices is only a promotion.

Three models, one price

• Claude Sonnet 5.5 (Anthropic), released 28 September. $2/$10, unchanged from Sonnet 5. Context window up to 1 million tokens.
• GPT-6.1 Sol (OpenAI), released 29 September, a week after GPT-6 Sol. $2/$10, about a fifth of GPT-6 Astra ($10/$50).
• Gemini 4 Argon (Google), announced 30 September and opened to paying users from 13 October. Introductory price of $2/$10 until the end of 30 November, then $4/$20 — the same as Claude Opus 5.5. Free accounts cannot use it yet.

To be clear: Argon is Google's most capable model, while Sonnet and Sol are the mid tier of Anthropic and OpenAI — both of which also sell Opus 5.5 ($4/$20) and Astra ($10/$50) above them. The three only meet at the same price for as long as Google's promotion lasts.

Two leaderboards, two winners

Two independent leaderboards are cited more than any others right now, and they give different results.

The Artificial Analysis Intelligence Index (version 4.3) combines around ten tests of reasoning, knowledge, coding and autonomous tasks. On this index:

• Claude Opus 5.5: 58
• Claude Sonnet 5.5: 56
• Gemini 4 Argon (high reasoning): 53, level with GPT-6 Astra
• GPT-6.1 Sol: 52

The Vals Index from Vals AI focuses on professional work such as finance, law and coding, and records the actual cost of running the whole benchmark. The 16 October table:

• Gemini 4 Argon: 68.9% — first
• Claude Sonnet 5.5: 67.04%
• Claude Opus 5.5: 66.97%
• GPT-6.1 Sol: 61.15% — ninth

Argon tops this table but sits five points behind Opus 5.5 on the other. Sonnet 5.5 is second on both. Neither leaderboard is wrong — they measure different work. That is the first reason not to choose a model from a single ranking: the answer depends on the work you intend to give it.

Same price per token, nearly seven times apart per task

The most striking number is not in the score column but in the cost column. Vals records the actual money spent running each test in the benchmark:

• GPT-6.1 Sol: $3.24 per test
• Gemini 4 Argon: $15.68 per test (at the standard $4/$20 price; at the introductory price this falls to roughly half)
• Claude Sonnet 5.5: $21.34 per test
• Claude Opus 5.5: $32.14 per test

Sonnet 5.5 and GPT-6.1 Sol share a list price, yet on the same tests Sonnet costs about 6.6 times as much — in return for nearly 6 percentage points more accuracy. Whether that is worth it depends on the task: for contract analysis where a mistake means paying a lawyer to review it, 6 points may be well worth the money. For sorting emails, almost certainly not.

Why such a gap at the same price? Because the list price is only half the calculation. The other half is how many tokens each model uses to finish the job, and that varies a lot:

• Today's models all "think" before answering — writing an internal chain of reasoning the user usually does not see. Those thinking tokens are billed as output tokens.
• On multi-step work (calling tools, reading documents, fixing code), one model may need fewer turns than another. Every extra turn rereads the whole context — more input tokens.
• The reasoning effort you set when calling the API can change output tokens several times over on the same model.

A small calculation to show the mechanism: a task needs to read 30,000 tokens of documents. Model A writes 2,000 tokens (thinking included), model B writes 8,000. Both at $2/$10:

• A: 30,000 × $2 / 1 million + 2,000 × $10 / 1 million = $0.06 + $0.02 = $0.08
• B: $0.06 + 8,000 × $10 / 1 million = $0.06 + $0.08 = $0.14

Same list price, and B costs 75% more per task. Multiply by ten thousand tasks a month and that is the difference between $800 and $1,400.

The promotional price has an end date

Google states plainly that Argon's $2/$10 is an introductory price lasting until the end of 30 November. From December it rises to $4/$20 — double. Artificial Analysis measured the cost per task on its index: $1.99 at the introductory price and $3.98 at the standard price. At the promotional price, Argon matches Astra's score at about 60% of the cost per task; at the standard price, that cost advantage is gone.

Gemini 3.8 Flash, Google's smaller model, is also on an introductory price of $0.75/$3.75, rising to $1.50/$7.50 from 1 January 2027.

There is nothing wrong with introductory pricing — it is a sensible way to get people to try a model. But if you are budgeting for 2027 or signing customer contracts based on your AI costs, budget at the standard price. A system chosen because it was cheapest in November may not be the cheapest in December.

A dimension rarely discussed: when the model doesn't know

Artificial Analysis also measures something composite scores do not show: when a model meets a question it does not know the answer to, does it say "I don't know" or invent a plausible answer? On this test, Argon answered wrongly instead of declining in about 15% of those cases, compared with 51% for GPT-6 Astra — even though the two have the same overall score.

For an internal assistant answering questions about company policy, or a tool extracting figures from financial reports, this measure may matter more than a few points on the overall table. An answer of "no information found" tells the user to ask again. An invented figure looks exactly like a real one.

This is one independent party's measurement on a specific question set, not a guarantee for every situation. But it suggests something worth doing in your own tests: include a few questions whose correct answer is "not in the documents", and see which model is willing to say so.

Four steps to choose with your own numbers

One — pick 30 real tasks. Take about 30 requests from the daily work you plan to hand to AI, including easy ones, hard ones, and a few where the correct answer is "I don't know". Write down the pass criterion for each in advance. Thirty tasks are not enough to separate two models a few percentage points apart, but they are enough to reveal large cost gaps and obvious failures.

Two — run them and record the tokens. Every API response includes the actual tokens used. With Anthropic and OpenAI that is the usage field (with input_tokens and output_tokens). With Gemini it is usageMetadata, where thinking tokens are recorded separately in thoughtsTokenCount — remember to add them to the output. Store these numbers alongside whether each task passed.

Three — cost it at standard prices. Cost per task = input tokens × input price + output tokens × output price. Use the post-promotion price. If you cache repeated documents, count the tokens read from cache separately at the cache price.

Four — divide by the tasks that passed. Total cost divided by the number of passing tasks gives the cost of one usable result — the only number that is truly comparable across models. A cheap model that fails a third of the tasks can end up dearer than an expensive one that gets nearly all of them right, once you count the human time spent checking and redoing.

Finally, build the system so the model name is one configuration line, not scattered through the code. In the three weeks from 22 September to 13 October, the three labs released at least six new models between them. The best model for you today is almost certainly not the best one in three months.

Conclusion

Three models at $2/$10 is a temporary coincidence, not a fair contest: one side is a flagship on promotion, the other two are mid-tier models at standard prices. And that coincidence exposes something the price list hides: price per token is not price per task.

To sum up with the numbers available: Sonnet 5.5 sits consistently near the top of both tables but uses the most tokens of the three; GPT-6.1 Sol is far cheaper per task with a lower score; Argon leads on professional work and invents answers least often, but its cost advantage expires on 30 November. There is no overall winner — only a model that fits a kind of work at a cost you can accept.

The next article in this series looks at small models, where the cost gaps are even larger, and why much of a business's work may not need a mid-tier model at all. If you want to test models on your own data and processes before committing, contact us or see KonexForge's AI & ML capability.