Gemini 4 Argon looks strongest on professional, domain-heavy work such as legal, finance and long engineering tasks, and less dominant on general intelligence rankings, where Claude Opus 5.5 and Sonnet 5.5 still score higher. Until 30 November it costs the same per token as Sonnet 5.5, so if you’re curious, now is the cheapest time to test it on your own workload. From 1 December the price doubles to Opus level, and at that price I’d only keep it where it clearly beats the alternatives in your own tests. I’d also treat the “one million output tokens” headline with caution: Google’s own description of it changed between announcements.
Live API prices: Gemini 4 Argon’s price changes on 1 December. To see what it and more than 400 other models cost per million tokens today, check the AI models table on our dashboard, read twice a day from OpenRouter’s public model list. Gemini plan prices and outages are on our Google AI (Gemini) page.
I haven’t used Gemini myself; I work with Claude. Everything here comes from Google’s announcement, independent coverage and my own cost arithmetic.
Gemini 4 Argon at a Glance
| Announced | 30 September 2026, first to trusted cyber defenders through Google’s Fairwind Program |
|---|---|
| Broad availability | 13 October 2026: paid Gemini API, Google AI Studio, Antigravity, eligible Gemini Enterprise plans, and Google AI Pro and Ultra subscribers in the Gemini app |
| API price until 30 November | $2 input / $10 output per million tokens; cached input 95% off (about $0.10) |
| API price from 1 December | $4 input / $20 output per million tokens |
| Output limit | Described as up to 1 million tokens (see the caveat below) |
| Built for | Long software-engineering, knowledge-work and defensive cybersecurity tasks |
How It Scores: Strong on Professional Work, Not the Overall Leader
Independent figures below come from VentureBeat’s 13 October report, which I read in full; it cites Vals AI and Artificial Analysis.
| Benchmark | Gemini 4 Argon | Best rival listed | Who reported it |
|---|---|---|---|
| Vals Index (professional tasks) | 68.90%, first | Claude Sonnet 5.5, 67.04% | Vals AI (independent) |
| Finance Agent v2 | 65.40%, first | n/a | Vals AI (independent) |
| Artificial Analysis Intelligence Index | 53 | Claude Opus 5.5, 58 | Artificial Analysis (independent) |
| Harvey legal benchmark | 19.6% | GPT-6 Astra, 5.4% | |
| Zapier AutomationBench | 51.3% | GPT-6 Astra, 42.5% | |
| DeepSWE v1.1 | 77.9% | Claude Opus 5.5, 74.2% | |
| Terminal-Bench 4.0 | 57.4% | Claude Opus 5.5, 66.4% |
The pattern is consistent across Google’s and independent numbers: Argon leads on specialised professional work and trails on general intelligence and terminal-style agent tasks. Sam Witteveen, a consultant who tested it, summed it up for VentureBeat as a very good model but not the best, and said he doesn’t think Google’s Antigravity is close to Claude Code or Codex for autonomous software engineering. I’d add that the legal benchmark gap is big enough to matter if that’s your field; the coding gaps aren’t.
The 1M Output Limit: Read the Fine Print
On 30 September Google described raising Argon’s output limit from 64K to as much as one million tokens. The interesting use isn’t a million-word document; it’s letting an agent run a very long reasoning-and-action trajectory without being cut off.
But VentureBeat points out that the 13 October announcement describes it differently: a unified limit of up to 1 million input and output tokens, reached through a special continuation protocol that needs API configuration. Google hasn’t said how that relates to the largest single uninterrupted response. Witteveen also cautions it isn’t how the model normally behaves in the Gemini app. My advice: don’t design anything around the 1M figure until you’ve confirmed it in Google’s API documentation for your setup.
What It Will Cost: Our Calculation
Method: we costed three fixed workloads with Google’s announced Argon prices and Anthropic’s published Claude prices (both checked 15 October 2026) in a short script. This is arithmetic, not a test of the model. Google hasn’t published cache-storage fees, so our cached figures for Argon include only the 95%-off cached-input rate.
| Workload | Argon until 30 Nov | Argon from 1 Dec | Claude Sonnet 5.5 | Claude Opus 5.5 |
|---|---|---|---|---|
| 100,000 short requests (1,500 in / 150 out) | $450 | $900 | $450 | $900 |
| 1,000 agent sessions × 20 turns, cached context (36K cached + 4K new in, 800 out per turn) | $460.40 | $920.80 | $478.40 | $956.80 |
| One maximum-length run: 200K in + 1M out | $10.40 | $20.80 | n/a | n/a |
The same job doubles in price on 1 December: until then Argon matches Sonnet 5.5 per token, afterwards Opus 5.5. The long-output run is the part to watch. One run that uses the full allowance costs about $10 in output alone during the promotion and $20 afterwards, and an agent looping through several of those can reach hundreds of dollars fast. Set output limits and spending caps before you experiment with long trajectories. For other models’ prices, see my AI pricing comparison.
What Google Says It’s Using Argon For
Google says thousands of employees use Argon for specialised coding, research and writing. Its examples include agents that found memory optimisations freeing more than 300 TiB across Google infrastructure, C and C++ to Rust migrations including work on the Fuchsia Zircon kernel, a Chromium refactor of roughly 6 million lines, and about 100,000 employees a month using it through an internal version of Antigravity. These are company-reported results, not independent evaluations, but they show the intended workload: large engineering projects over many iterations, not ten-line demos.
Cybersecurity and the Phased Rollout
Google reports a 68% score on CWE-bench v1 for vulnerability remediation and says Argon can find, validate and patch vulnerabilities. That’s why the first access went to vetted defenders, and why Google is expanding the Fairwind Program to hundreds more security partners, including versions without some cybersecurity restrictions for authorised defensive work. Google also emphasises misuse prevention, prompt-injection resistance and sandboxed environments.
Who Can Use It Now
- Developers: the paid Gemini API, Google AI Studio, Android Studio and Firebase AI Logic.
- Individuals: Google AI Pro and Ultra subscribers in the Gemini app and AI Mode in Search; Antigravity requires the Ultra tier.
- Businesses: eligible Gemini Enterprise plans.
Should You Use It?
| If you… | My recommendation |
|---|---|
| Work on legal, finance or other domain-heavy tasks | Test it now, while it’s priced like Sonnet |
| Build terminal-style coding agents | Stay with Claude Code or Codex for now |
| Already pay for Google AI Pro | Try it in the Gemini app at no extra cost |
| Plan to rely on the 1M output limit | Wait until Google documents how it works |
| Are price-sensitive after 1 December | Keep it only where your tests show a clear win over Sonnet 5.5 |
Related: Gemini agent vs ChatGPT workspace agents · Claude vs ChatGPT for coding · Sonnet 5.5 pricing
More from EvalAITools: Gemini agent vs ChatGPT workspace agents · Gemini connected apps and privacy
