Gemini 4 Argon launches on GPT-6.1 Sol's rate card and is scheduled to move to Claude Opus 5.5's. On Artificial Analysis, Opus 5.5 at high effort already scores higher than Argon for less than Argon's introductory price.
Google announced its first Gemini 4 model yesterday with two prices and no date for when one becomes the other. Neither price is new. Both already belong to someone else, which makes this the easiest flagship launch of the year to compare, as long as you compare tasks and not tokens.
- Intro price $2 / $0.10 / $10, later $4 / $0.20 / $20. No date for the switch.
- Artificial Analysis: 53 for $1.99 a task. Opus 5.5 on high gets 54 for $1.82.
- Vals: first of 41 at $15.68 a test, about half what Opus 5.5 costs there.
- Only Fairwind cyber defenders and early testers can call it today, and there is no model ID yet.
Sol's prices now, Opus's prices later
Argon starts at $2.00 in and $10.00 out, with cached input 95% off at $0.10. That is GPT-6.1 Sol's card on all three lines. "After the introductory period expires", in Google's words, it goes to $4.00 and $20.00, and the same 95% cache discount makes that $0.20. That is Claude Opus 5.5's card on all three lines.
| Per 1M tokens | Argon intro | GPT-6.1 Sol | Argon later | Opus 5.5 | Gemini 3.1 Pro |
|---|---|---|---|---|---|
| Input | $2.00 | $2.00 | $4.00 | $4.00 | $2.00 |
| Cached input | $0.10 | $0.10 | $0.20 | $0.20 | $0.20 |
| Output | $10.00 | $10.00 | $20.00 | $20.00 | $12.00 |
| Long prompts | not published | 2x / 1.5x over 272K | not published | flat to 1M | $4 / $18 over 200K |
| Batch | not published | 50% off | not published | 50% off | 50% off |
Argon rates from Google's launch post; the $0.20 is the stated 95% discount applied to $4.00. Other rows from the Gemini API, OpenAI and Anthropic pricing pages, read October 1, 2026. Gemini 3.1 Pro is still a preview.
The intro card undercuts Google's own previous flagship on output, $10 against Gemini 3.1 Pro's $12. The later card doubles every line. Google hasn't said when the switch happens. When it put Gemini 3.8 Flash on an introductory price a month ago, it printed December 31. Argon gets no date. It also gets no batch rate, no cache storage rate and no word on whether 3.1 Pro's 200K price step carries over.
Same card as Sol, 2.76x the bill
Matching prices make one thing easy. When two models charge the same per token, any gap in cost per task comes from how many tokens each one spends. Artificial Analysis has only run Argon at high so far, and its launch write-up puts Argon at 53 on the Intelligence Index for $1.99 a task at the intro price. At the later price AA prints $3.98.
- 58Opus 5.5 max$5.98
- 56Sonnet 5.5 max$7.62
- 54Opus 5.5 high$1.82
- 53Argon high, list price$3.98
- 53GPT-6 Astra max$3.26
- 53Argon high, intro price$1.99
- 52GPT-6.1 Sol max$0.72
- 30Gemini 3.1 Pro Preview$0.67
Score on the Artificial Analysis Intelligence Index v4.3.2 and cost per index task, read October 1, 2026.
GPT-6.1 Sol on max scores 52 for $0.72. Same three prices, one point lower, and Argon costs 2.76x as much per task. Argon wrote about 110M output tokens across the index, roughly 62,000 a task. AA counts about 27,000 a task for GPT-6 Astra on max.
Against Astra the intro price does its job. Both score 53, and Argon is 39% cheaper per task at $1.99 against $3.26. At the later price that flips: $3.98 is 22% more than Astra, even though Astra's input and output prices are 2.5x Argon's.
The Opus row nobody is quoting
Google's benchmark table puts Argon against Opus 5.5 on max. On AA's list, Opus 5.5 on high scores 54 for $1.82 a task. That is one point above Argon for 9% less than Argon's intro price. When Argon moves to Opus's own card, the same comparison is $3.98 against $1.82, or 2.19x, and that is a pure token gap because every price on the two cards is then the same.
So on AA, the question isn't whether Argon is cheaper than Opus 5.5. It is whether Argon at a lower effort lands near 53 for a lot less money, the way GPT-6.1 Sol on medium did. AA hasn't run a lower setting yet, and Google hasn't published which thinking levels Argon accepts.
One number on AA favours Argon without a cost caveat. Its hallucination rate on AA-Omniscience is 15%, the lowest of any model scoring 45 or more on the index. Astra sits at 51% and GPT-6.1 Sol at 54%.
Vals reads it the other way
The Vals Index has Argon first of 41 models, and cheaper per test than every Claude and Astra row near it.
| Model | Vals Index | Cost per test |
|---|---|---|
| Gemini 4 Argon (high) | 68.90% | $15.68 |
| Claude Sonnet 5.5 | 67.04% | $21.34 |
| Claude Opus 5.5 | 66.97% | $32.14 |
| Claude Fable 5.1 | 65.83% | $28.71 |
| GPT-6 Astra | 63.13% | $18.46 |
| GPT-6.1 Sol | 61.15% | $3.24 |
| Gemini 3.8 Flash | 54.83% | $5.73 |
Vals AI index page, read October 1, 2026.
That is 51% under Opus 5.5 and 27% under Sonnet 5.5 for a higher score. Vals lists Argon at $4 and $20 and doesn't say which rate its $15.68 used. If it is the later price, the intro price would put the same run near $7.84. Even then GPT-6.1 Sol does the index for $3.24, seven and three-quarter points lower.
Two indexes, two answers. Vals includes finance, legal and coding agent tasks, and those are where Google's own table has Argon well ahead: 19.6 on Harvey Legal Agent against 3.8 for Opus 5.5, and 51.3 on AutomationBench against 42.5. On Terminal-Bench 4.0 it trails Opus 5.5, 57.4 to 66.4, and on FrontierSWE v2 it trails Astra, 55.0 to 65.5.
You can't call it yet
Today Argon is open to Google's Fairwind Program, a vetted group of cyber defenders, and to early testers. Paid Gemini API customers and Google AI Ultra subscribers are next, then wider developer and enterprise access. Google gave no dates for any of it.
Nothing on Google's developer side has caught up with the blog post. The Gemini API pricing page, the models page and Vertex AI pricing had no Argon row when we checked this morning, and there is no published model ID. Artificial Analysis and Vals both list a 1M-token context window; Google's post doesn't state one.
Google does state 1M output tokens, up from 64K. AA says that comes from a new API feature, Long Decode Continuation, which pauses a long answer and resumes it across follow-up calls. Vals lists 262K as the max output. We'd read 1M as a total across several calls, and plan for every one of those tokens to bill at the output rate.
Budget at $4, test at $2
Budget at $4 and $20. The intro price has no end date, so any estimate built on $2 and $10 can double without warning. At $4 and $20, AA puts Argon at about 2.2x Opus 5.5 on high per task, and Vals puts it at about half Opus 5.5 per test.
If your work looks like Vals's finance, legal or long-context tasks, Argon is the one to test first when access opens. If it looks like terminal work or pure reasoning, Opus 5.5 on high or GPT-6.1 Sol already costs less per task on the numbers we have.
We've added Argon to the catalogue as gated, at the intro rate, with the later rate in its notes. The calculator prices any token mix against Argon, Opus 5.5 and Sol.
What we read
- Google: Gemini 4 Argon - introductory and later prices, cache discount, 1M output, access plan
- Google DeepMind: Gemini 4 Argon evaluation methodology - benchmark table against Opus 5.5, Fable 5.1 and GPT-6 Astra
- Gemini API pricing - Gemini 3.1 Pro and 3.8 Flash rows; no Argon row as of October 1, 2026
- Artificial Analysis: Gemini 4 Argon and leaderboard - Intelligence Index scores, cost per task, output tokens, Omniscience hallucination rate
- Vals AI: Vals Index - scores and cost per test
- OpenAI: API pricing and Anthropic: Pricing - GPT-6.1 Sol and Opus 5.5 rows
- 9to5Google: Gemini 4 Argon announcement - access and pricing coverage