Why a Cheaper AI Model Can Cost More per Finished Task

A cheaper price per token does not guarantee a cheaper finished task. The bill depends on tokens burned, input reused from cache, results rejected and redone, and reviewer time.

The October 2026 data shows the gap but cannot settle it. No source reports cost per accepted task, or an acceptance rate, for any of these models.

The short answer

List prices leave out the real terms: tokens used per task, the cache-hit share and the acceptance rate decide a real bill.
Rankings can flip: in the FrontierSWE leaderboard, Opus 5.5, whose token price is double Argon’s introductory price, cost less per trial than Argon.
Token use varies a lot: Artificial Analysis reports Muse Wonder used about 7.69 million input tokens per task, two to seven times Argon, Sol and Astra.
The honest formula: cost per accepted task is cost per attempt divided by acceptance rate.

Why do list prices mislead?

Per-token prices are the only numbers vendors publish in a comparable form, so they dominate comparisons. They leave out three terms that decide a real bill: tokens used per task, the cache-hit share, and the acceptance rate.

Simon Willison made the same point on 13 October. What matters, he noted, is how many tokens a model needs to finish a task inside an agent harness. That is his opinion, not a measurement.

What are the list prices?

Prices are per million tokens.

Model	Input	Cache read	Output	Notes
GPT-6.1 Sol	$2	$0.10	$10	Prompts over 272K input tokens cost 2x on input and 1.5x on output for the whole request Batch and Flex are 50% below standard
Claude Opus 5.5	$4	$0.20	$20	5-minute cache write $5; Batch 50% off; Fast mode $8 / $40
Gemini 4 Argon, through 30 Nov 2026	$2	$0.10	$10	Output includes thinking tokens; implicit caching only
Gemini 4 Argon, from 1 Dec 2026	$4	$0.20	$20	Per Google’s Gemini API pricing page

From 1 December, Argon’s listed price doubles and matches Opus 5.5 line for line. Any comparison made at the introductory price will change then.

What did four sources actually measure?

Source	Unit of cost	Reported figures
Artificial Analysis, 12 Oct 2026	Cost per Intelligence Index task at max reasoning effort	Opus 5.5 $5.98 (labeled “max w/ fallback”) Muse Wonder $4.52 Gemini 4 Argon (high) $1.99 GPT-6.1 Sol $0.72
Leonardo Gonzalez, Trilogy AI, 14 Oct 2026	One Harbor migration attempt	Muse about $3 to $8 Argon on Google’s native API about $12 Opus $42 to $57 at metered-equivalent rates (it ran on a subscription) review and repair work excluded
FrontierSWE v2 leaderboard (Proximal Labs, as of 30 Sep 2026), as relayed by a third-party blog	Mean score over five attempts on 34 tasks (up to 20 hours each), and cost per trial	GPT-6 Astra 65.5% at $1,029.65 per trial Claude Opus 5.5 62.3% at $98.87 Gemini 4 Argon 55.0% at $129.36
Simon Willison, 13 Oct 2026	Entire benchmark run	$2,407 for Argon (High), quoting Artificial Analysis

These are four different units, so do not put them in one chart. The FrontierSWE leaderboard footnotes that Argon’s cache hit rate was “much lower due to a non-production setting”, and cost per trial is not cost per resolved task. Still, the table shows how rankings can diverge from list prices.

How does Artificial Analysis compute cost?

Its methodology page says it uses the token counts each provider’s API reports. It falls back to a canonical tokenizer when they are missing, and combines those counts with “live measurements of the model’s typical cache hit rate”. So the cache assumption in its numbers is a general provider measurement, not something observed in the benchmark run itself.

A version 4.1 changelog entry says token and cost metrics were upgraded to reflect cache hit rates and cache token pricing. The page defines no “cost per task” formula and does not say which token types are in the headline figure. Treat the numbers as a comparison aid, not a bill.

What moved the bill in these studies?

Tokens per task. Artificial Analysis ties Muse Wonder’s higher cost to both its per-token prices and its heavy token usage.
Cache reuse and route. In the Trilogy study, cached input was about 96% for Muse through the Claude Agent SDK on OpenRouter, but about 9% for Argon on that same route. For Argon on Google’s native API it was 92% to 98%. These are single-task figures.
Failed runs. Argon’s OpenRouter attempts hit upstream timeouts as context neared roughly 500,000 tokens and were abandoned, so tokens were spent without a result.

How does a cheaper token become a pricier task?

This is our own illustrative example, not a measurement. Formula: cost = uncached input x input price + cached input x cache-read price + output x output price. Our assumptions: 1 million input tokens, 50,000 output tokens, cache-write fees ignored.

Case	Opus 5.5	GPT-6.1 Sol
No caching	$5.00	$2.50
90% cache hits	$1.58	$0.79

Cost per accepted task = cost per attempt divided by acceptance rate. If Opus were accepted half the time, that is $1.58 / 0.5 = $3.16. If Sol were accepted a quarter of the time, that is $0.79 / 0.25 = $3.16, the same, even though Sol’s token price is half.

The 50% and 25% are invented to show the mechanism. No source publishes acceptance rates for these models.

How do you measure your own cost per finished task?

We suggest this routine. It is our method, not a vendor rule.

Pick 10 to 20 real tasks with a written acceptance rule (see when passing tests hide an unfinished task).
Run the same tasks on two models with the same settings and log actual tokens and cache share.
Add the cost of rejected attempts and the reviewer’s time.
Divide by accepted tasks only.

For the side-by-side list prices, see Gemini 4 Argon vs GPT-6.1 Sol vs Claude Opus 5.5.

Questions and answers

Is the model with the lowest token price always cheapest?

No. Tokens used, cache reuse, retries and review time all change the total.

Are Artificial Analysis cost figures a real bill?

No. They use provider-reported token counts and a general measure of cache hit rates, and the page gives no cost-per-task formula.

Why did Argon’s cache share differ so much in the Trilogy study?

The author reports about 9% on one OpenRouter route and 92% to 98% on Google’s native API, for one migration task. The route changed the cache share, so check yours before you compare prices.

When does Argon’s price change?

Google’s API pricing page lists $2 input and $10 output through 30 November 2026, and $4 and $20 from 1 December 2026.

Terms in plain language

Token: a small piece of text that AI services count and charge for.
Cache hit: input the service has already seen and can reuse at a lower price.
Acceptance rate: the share of attempts whose result you actually accept.
Harness: the software wrapper that lets a model use tools and run tasks.

Sources

Facts last checked against the linked pages on 15 October 2026. Prices can change.

Part of the guide: AI Model Pricing and API Cost Guides