Gemini 4 Argon Cost per Task: The Hidden Token Bill Benchmarks Don't Show You

Artificial Analysis measures $1.99 per Intelligence Index task for Gemini 4 Argon (high), against $3.26 for GPT-6 Astra (max), at the same index score of 53. That price rests on a 50% launch discount: at standard pricing the same work would cost $3.98, about 1.2 times Astra. The gap between those two numbers is a token story, and per-token price lists hide it.

All figures below are as of 12 October 2026 unless stated.

How much does Gemini 4 Argon cost per task?

Independent measurement puts it at $1.99 per task at promo pricing and a projected $3.98 at standard pricing. Artificial Analysis (AA) describes the second number as what cost per task "will increase to" on standard pricing, so it is a projection, not an observed bill.

Vals AI reports a different unit: $15.68 per test on the Vals Index, where Argon scores 68.90%. The two vendors use different task sets, so $1.99 and $15.68 cannot be compared with each other. Each is useful only against other models on the same index.

What does Google charge per token?

Google's announcement of 30 September 2026 sets an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input at 95% off the input price. A footnote says $4 per million input and $20 per million output apply after the introductory period.

Google gives no end date. AA says the discount runs for at least one month and that Google has not confirmed when it stops. Google's Gemini API pricing page did not list Argon when checked, so the announcement is the only first-party price source.

Vals lists Argon at $4.00 and $20.00, the post-introductory rate.

Why does the token bill differ from the price list?

Because cost per task is tokens times price, and Argon uses a lot of tokens. AA reports Argon averaged 62k output tokens per task, compared with 27k for GPT-6 Astra (max), about 2.3 times as many. Across the full index run AA's model page records 110M output tokens against a median of 82M for comparable models.

A rough check using only the published prices (our arithmetic, not an AA figure): 62k output tokens at $10 per million is $0.62 per task at promo pricing, and $1.24 at $20 per million. Output alone does not explain $1.99 or $3.98. AA's method also counts input, cache hit, cache write and reasoning tokens, then divides by task count and weights by Intelligence Index weight.

AA also states that Argon's cost position is driven by lower token prices, rather than reduced token use. Once the promo ends, that advantage is gone against Astra: $3.98 versus $3.26.

How does Argon compare on cost with other models?

On AA's index, Argon at promo pricing is 60% of Astra's cost and 2.7 times the cost of GPT-6.1 Sol (max), which scores 52 on the same index.

Index version v4.3.2 comes from AA's model page. On Vals, Argon's $15.68 per test sits below Claude Sonnet 5.5 ($21.34), Claude Opus 5.5 ($32.14) and Claude Fable 5.1 ($28.71), while scoring 68.90% against 67.04%, 66.97% and 65.83%. Vals ran Argon at high reasoning effort and temperature 1.

Where do the costs climb?

On long agentic work. The same Vals page shows Argon's cost per test rising to $193.78 on CUA-bench, $57.82 on Code Migration, $44.90 on Terminal-Bench Science and $30.45 on SRE Bench, far above the $15.68 index average.

This is where an index-average price misleads. A team running computer-use agents will not see $15.68 per test. Cost per task depends on which tasks, and the workloads that most need a 1M-token output budget are the ones that burn it.

That budget is a company claim. Google says it expanded the output token limit to 1M tokens, up from 64K, so the model can generate hundreds of thousands of tokens in one trajectory. AA says it tested Argon with a Long Decode Continuation setup that lets reasoning run up to 1M output tokens without request timeouts. Vals capped output at 262,144 tokens in its own settings. Different caps can change both scores and cost, so the three sets of numbers are not run under identical conditions.

What do Google's own benchmarks say, and what do they leave out?

Google's benchmark table is lab-reported and has no cost column. It lists Argon at 68.9% on the Vals Index, 51.3% on AutomationBench and 77.9% on DeepSWE v1.1. It also shows Argon behind on some rows: Terminal-bench 4.0 at 57.4% against 66.4% for Claude Opus 5.5, and FrontierSWE v2 at 55.0% against 65.5% for GPT-6 Astra.

None of those rows says how many tokens each score took. A score gap of two points with a token gap of 2.3 times is a different purchasing decision from a two-point gap at equal usage. Google's page links a methodology document, but no model card for Argon is listed in Google's model card index.

AA adds a quality caveat on its office-work test: Argon's AA-Briefcase result is driven by a 65% rubric pass rate, with lower Analytical Quality (1576 Elo) and Presentation Quality (1308 Elo). A cheaper task that needs rework is not cheaper.

How should you estimate your own cost per task?

Run your own tasks and meter tokens. Do not multiply a benchmark's token count by a list price and call it a forecast. A workable method:

1. Log input, cached input, output and reasoning tokens for 50 to 100 representative tasks.
2. Price them at both $2/$10 and $4/$20, since the introductory period has no published end.
3. Add the cost of retries and failed tasks, which index averages hide.
4. Report cost per successful task, not per attempt.

Argon cannot be run this way by most teams yet. Google says it is rolling out to trusted cyber defenders through its Fairwind Program, with paid API customers and Google AI Ultra subscribers next and no date given.

What's still unclear?

• How long the discount lasts. Google's footnote gives no date. AA says at least one month.
• Cached input after the introductory period. Google states 95% off the input price but does not state a post-introductory dollar figure.
• Whether availability is limited. AA's article says Argon is not publicly available, while its model page FAQ says it is available through one API provider. The page does not name the provider.
• Cross-vendor comparability. AA prices Argon at the promo rate, Vals at $4/$20. The cost numbers come from different task sets, effort settings and output caps.
• Weakest point of this analysis. The headline "hidden token bill" rests on AA's per-task token counts for two models: 62k versus 27k. It is one comparison, not a survey of all models.
• Version drift. Vals's own text gives Argon a rank of #1 of 41 in the launch update and #1 of 45 on the model page, so its leaderboard is moving.

Disclosure: Model Report Card is affiliated with Humanbit, a data labelling and evaluation company.