Gemini 4 Argon
Gemini 4 Argon is Google's new frontier model, announced on Sep 30, 2026 and, for now, available only to cyber defenders invited through the Fairwind Program. The introductory price is US$ 2 per million input tokens and US$ 10 per million output; later it becomes US$ 4 and US$ 20, on an unpublished date. In Google's own table it leads 13 of 19 tests, ties one and trails in five, and on Terminal-Bench 4.0 it comes last of the four. No independent source has measured Argon yet.
Gemini 4 Argon in official numbers
Data published by the maker in its documentation and announcement. The source of each item is at the end of the page.
- Maker
- Google DeepMind
- Announcement
- Sep 30, 2026
- API identifier
- not published
- Access today
- Fairwind Program invitees (cyber defense)
- Next in line
- paid API customers and Google AI Ultra subscribers, no date
- Input
- US$ 2 per million tokens (introductory)
- Output
- US$ 10 per million tokens (introductory)
- After the introductory period
- US$ 4 input and US$ 20 output
- End of introductory price
- not published
- Cached input
- 95% off the input price
- Max output
- 1 million tokens (previously 64 thousand)
- Context window (input)
- not published
- Knowledge cutoff
- not published
- Reasoning
- levels not documented
- Model card
- not published as of Oct 1, 2026
- Zero data retention
- yes, on Gemini Enterprise
Where Gemini 4 Argon does well, where it falls behind and who it is for
Where it does well
- Legal and finance: 19.6% on Harvey's Legal Agent Benchmark, against 6.7% for the runner-up, and 65.4% on Vals Finance Agent v2, against 58.9%, in Google's figures.
- Automation: 51.3% on AutomationBench, against 42.5% for Claude Opus 5.5 and 41.4% for GPT-6 Astra.
- Long-horizon coding: 77.9% on DeepSWE v1.1, against 74.2% for Opus 5.5 and 74.1% for GPT-6 Astra.
- Long context and long video: 84.2% on GraphWalks from 256 thousand to 1 million tokens, against 71.8% for the runner-up, and 91.7% on LVBench.
- Prompt injection: in the Gray Swan test, indirect attacks had a 0.7% success rate over 15 attempts, against 1.0% for Opus 5.5 and 8.5% for GPT-6 Astra. Lower is better.
Where it falls behind
- Nobody outside uses it: invitation-only access, with no API identifier and no opening date.
- Terminal work: 57.4% on Terminal-Bench 4.0, last of the four in Google's own table. Claude Opus 5.5 scores 66.4%.
- FrontierSWE v2: 55.0%, also last of the four, against 65.5% for GPT-6 Astra. On PostTrainBench it loses to Opus 5.5 (45.3% against 49.3%).
- Science in the terminal and computer use: behind GPT-6 Astra on Terminal-Bench Science 0.1 (57.6% against 68.1%) and OSWorld 2.0 (69.2% against 72.6%).
- Only the maker has spoken: no model card and no Frontier Safety Framework report, and some of Argon's scores were computed by Google itself.
Who it makes sense for
- Security teams at governments, critical infrastructure (health, telecom, energy, finance) and large technology platforms, which get priority in Fairwind.
- Academic labs doing defensive benchmarking, which can also apply.
- Paid Gemini API customers and Google AI Ultra subscribers, first in line: it is worth building now the test that will decide the switch.
- Large code migrations and other tasks that need a very long answer, where 1 million output tokens changes what fits in one request.
When to pick another model
- Projects that need to ship now: from Google, you can use Gemini 3.8 Flash or Gemini 3.1 Pro today, the latter in preview.
- Agents that spend hours in the terminal: in Google's own table, Opus 5.5 leads Terminal-Bench 4.0.
- Science in the terminal and computer use, where GPT-6 Astra scored higher in the same table.
- Long-term budgets at the introductory price: it doubles on a date Google has not disclosed.
What Google announced on September 30
Argon was presented by Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect at Google, as the company's frontier model for coding, enterprise knowledge work such as legal and finance, and cyber defense. According to Google, thousands of employees already use it internally.
The release is phased. First come trusted cyber defenders through the Fairwind Program, while Google takes part in the U.S. government's voluntary pre-release model access process. Fairwind, created on September 2, 2026 with Gemini 3.8 Flash Cyber, has more than 650 partners, with priority for governments, critical infrastructure and technology platforms, and requires phishing-resistant MFA, per-employee usage logs and background checks. Only a subset of partners has Argon, and for them it comes without the cybersecurity guardrails. Later, the promise is to open it to developers, enterprises and consumers, with no date.
The number that spread the most, 1 million tokens, is the output limit, up from 64 thousand. The summary at the top of the announcement itself mentions a 1 million token limit without saying of what, and several sites repeated the number without saying it refers to output. The input window has not been published; the only hint is the GraphWalks test in the 256 thousand to 1 million token range.
Google's table: 13 wins, one tie and five tests behind
Google compared Argon with GPT-6 Astra, from OpenAI, and with Claude Fable 5.1 and Claude Opus 5.5, from Anthropic. That is 13 wins in 19 rows, a tie on CWE-bench v1 and five tests behind: FrontierSWE v2 and Terminal-Bench 4.0, where it comes last; PostTrainBench, to Opus 5.5; Terminal-Bench Science 0.1 and OSWorld 2.0, to GPT-6 Astra.
Google's methodology carries caveats. Competitor scores are the ones each maker reported, at the highest reasoning level. Several Argon scores were computed by Google itself, and on Terminal-Bench Science it ran with a verifier timeout six times longer. Self-reported scores also move: GPT-6 Astra has 64.6% on Terminal-Bench Science in its own launch and 68.1%, the figure Google used, in a later OpenAI announcement.
| Argon | GPT-6 Astra OpenAI | Fable 5.1 Anthropic | Opus 5.5 Anthropic | |
|---|---|---|---|---|
| Vals Index knowledge work |
68.9% | 63.1% | 65.8% | 67.0% |
| AutomationBench process automation |
51.3% | 41.4% | 31.4% | 42.5% |
| Vals Finance Agent v2 financial research |
65.4% | 53.5% | 58.9% | 58.6% |
| Harvey's Legal Agent Benchmark legal work |
19.6% | 5.4% | 6.7% | 3.8% |
| DeepSWE v1.1 long-horizon coding |
77.9% | 74.1% | 67.4% | 74.2% |
| FrontierSWE v2 coding |
55.0% | 65.5% | 56.3% | 62.3% |
| Vibe Code Bench coding |
91.9% | 89.6% | 90.3% | 90.3% |
| Terminal-Bench 4.0 long terminal work |
57.4% | 58.2% | 57.9% | 66.4% |
| PostTrainBench machine learning engineering |
45.3% | 44.3% | 40.2% | 49.3% |
| Terminal-Bench Science 0.1 science in the terminal |
57.6% | 68.1% | 52.6% | 63.3% |
| LABBench 2 science |
88.8% | 85.4% | 68.6% | 73.1% |
| RiemannBench math |
76.0% | 72.0% | 65.6% | 69.6% |
| GraphWalks, up to 128k long context |
99.7% | 98.7% | 91.4% | 90.6% |
| GraphWalks, 256k to 1M long context |
84.2% | 71.8% | 65.0% | 66.8% |
| Agent's Last Exam computer use |
39.5% | 34.2% | no data | 38.2% |
| OSWorld 2.0 computer use, offline subset |
69.2% | 72.6% | no data | no data |
| Chartography chart reading |
71.6% | 71.0% | 46.2% | 66.3% |
| LVBench long video |
91.7% | 87.5% | 79.7% | 83.7% |
| CWE-bench v1 code security, tie |
68.0% | 68.0% | 58.0% | 67.0% |
Figures from the Gemini 4 Argon announcement, identical on the Google DeepMind page and in the methodology PDF. The dot marks the best in each row; on CWE-bench v1, Argon and GPT-6 Astra tie. "No data" marks tests Google did not publish for that model.
Pricing: the introductory rate matches the rivals' middle tier
At the introductory price, Argon costs the same as GPT-6 Sol and GPT-6.1 Sol, from OpenAI, and Claude Sonnet 5.5, from Anthropic, the middle-tier models of both. At full price, it matches Claude Opus 5.5. GPT-6 Astra, its direct rival in the table, costs US$ 10 and US$ 50.
The math cannot be closed yet: Google has not published when the introductory price ends, batch pricing, a long-context tier, or whether the 95% cache discount continues afterwards. By our math, that discount comes to US$ 0.10 per million cached tokens at the introductory price.
| Argon introductory | Argon after | GPT-6 Sol and 6.1 Sol OpenAI | Sonnet 5.5 Anthropic | Opus 5.5 Anthropic | Gemini 3.1 Pro | |
|---|---|---|---|---|---|---|
| Input | US$ 2 | US$ 4 | US$ 2 | US$ 2 | US$ 4 | US$ 2 |
| Output | US$ 10 | US$ 20 | US$ 10 | US$ 10 | US$ 20 | US$ 12 |
| Max output | 1 million | 1 million | 128 thousand | 128 thousand | 128 thousand | 65,536 |
Standard pricing. Gemini 3.1 Pro for prompts up to 200 thousand tokens. Google, OpenAI and Anthropic pricing pages checked on Oct 1, 2026.
Examples of using Gemini 4 Argon
Google says Argon handled migrations from tens of thousands of lines (re2, libgav1) up to more than 800 thousand lines of the Fuchsia Zircon kernel, the latter still under audit. In libgav1, the Rust decoder became 2.7 times faster than the previous Rust port.
On a Fairwind defense team, the model finds, validates and fixes flaws on its own. According to Google, Wiz used it to find a critical flaw in healthcare software used by hospitals, which earlier models had missed.
An agent reads contracts and financial statements, cross-checks the sources and delivers the analysis. That is what Vals Finance Agent v2 and Harvey's Legal Agent Benchmark measure, two of the tests where Argon pulls furthest ahead in Google's table.
Where Gemini 4 Argon stands out and where it falls behind, by the numbers
Gemini 4 Argon has published results on 20 tests compiled by Epoch AI. The yardstick here is its generation: its position among the models released between Jan 3, 2026 and Dec 29, 2026 that took the same test. Against the whole archive, which includes 2023 models, a recent release lands near the top of almost everything, and that says nothing.
Where it stands out
Tests where it sits in the top third of its own generation.
Where it falls behind
Tests where it sits in the bottom half of its own generation.
It is not in the bottom half of its generation on any test with 6 or more models compared.
See all 20 tests, one by one
| Test | Score | In its generation | Overall | Test record |
|---|---|---|---|---|
| GPQA Diamond | 96.4% | #1 / 73 | #1 / 173 | 96.4% Gemini 4 Argon |
| AIME 2024 and 2025 (OTIS) | 100.0% | #1 / 75 (tied with 11) | #1 / 168 | 100.0% GPT-5.5 |
| CritPt | 27.1% | #14 / 76 (tied with 1) | #14 / 112 | 32.3% GPT-5.6 Sol |
| SciCode | 61.8% | #3 / 74 | #3 / 104 | 66.9% Claude Opus 5.5 |
| Frontiermath Tiers 1 3 V2 | 86.0% | #10 / 64 (tied with 1) | #10 / 86 | 93.7% GPT-6 Astra |
| WebDev Arena | 1,680 | #8 / 67 | #8 / 82 | 1,820 Claude Opus 5.5 |
| SimpleQA Verified | 76.0% | #1 / 59 | #1 / 81 | 76.0% Gemini 4 Argon |
| Chess puzzles | 59.0% | #8 / 58 (tied with 2) | #9 / 80 | 76.0% GPT-6 Astra |
| Game puzzles | 64.0% | #5 / 58 | #5 / 77 | 84.0% GPT-6 Astra |
| ProofBench | 99.0% | #4 / 64 (tied with 3) | #4 / 74 | 100.0% Claude Fable 5.1 |
| Frontiermath Tier 4 V2 | 78.0% | #9 / 54 (tied with 1) | #9 / 67 | 100.0% GPT-6.1 Sol |
| Vending Bench 2 | 13,718 | #3 / 43 | #3 / 57 | 15,515 GPT-6 Astra |
| Apex Agents | 82.4% | #1 / 48 | #1 / 50 | 82.4% Gemini 4 Argon |
| FrontierCode | 47.6% | #12 / 43 | #12 / 43 | 54.6% Claude Opus 5.5 |
| Gdp Pdf | 28.4% | #9 / 36 | #9 / 37 | 35.0% GPT-6.1 Sol |
| Furniture Assembly | 72.5% | #5 / 32 | #5 / 34 | 83.3% Claude Opus 5.5 |
| Blueprint Bench 2 | 54.4% | #1 / 29 | #1 / 31 | 54.4% Gemini 4 Argon |
| Ebr Bench | 64.8% | #3 / 22 | #3 / 27 | 76.2% GPT-6 Astra |
| FrontierSWE | 55.0% | #5 / 20 | #5 / 20 | 65.5% GPT-6 Astra |
| CursorBench | 52.9% | #3 / 16 | #3 / 16 | 57.8% Claude Opus 5.5 |
For each test, the model's best effort level counts, as in the ranking. Source of the scores: Epoch AI, under CC BY 4.0.
What to compare Gemini 4 Argon with
The models from other labs closest in capability and released in the same period, at most two per lab. Click any of them to see its page.
| Model | Lab | ECI | ELO | LiveBench | Release |
|---|---|---|---|---|---|
| GPT-6 Astra | OpenAI | 166.41 | n/a | 82.2 | Sep 3, 2026 |
| GPT-6.1 Sol | OpenAI | 166.32 | n/a | 81.6 | Sep 29, 2026 |
| Claude Fable 5.1 | Anthropic | 164.65 | 1,586 | 83.4 | Sep 1, 2026 |
| Gemini 4 Argon | Google DeepMind | 164.56 | 1,608 | n/a | Sep 30, 2026 |
| Claude Sonnet 5.5 | Anthropic | 164.45 | 1,537 | 77.8 | Sep 28, 2026 |
| Kimi K3 | Moonshot | 157.42 | 1,528 | 79.2 | Jul 16, 2026 |
| Qwen 3.8 Max | Alibaba | 156.88 | 1,536 | 78.5 | Aug 2, 2026 |
Gemini 4 Argon or Claude Fable 5.1?
Claude Fable 5.1 is the model from another lab closest to Gemini 4 Argon in capability. Across the 19 tests both took, this is the score:
| Test | Gemini 4 Argon | Claude Fable 5.1 |
|---|---|---|
| Vending Bench 2 | 13,718 | 5,422 |
| Blueprint Bench 2 | 54.4% | 41.9% |
| Chess puzzles | 59.0% | 48.0% |
| Apex Agents | 82.4% | 68.6% |
| CritPt | 27.1% | 31.1% |
| Ebr Bench | 64.8% | 57.1% |
| Frontiermath Tier 4 V2 | 78.0% | 87.8% |
| Game puzzles | 64.0% | 58.0% |
| SimpleQA Verified | 76.0% | 70.8% |
| FrontierCode | 47.6% | 50.9% |
Unweighted score: each test counts one point, whether math or coding. The dot marks who did better on each row.
What Gemini 4 Argon costs to deliver each task
Price per token is what is on the price sheet; cost per task is what lands on the invoice, because it depends on how many tokens the model spends. Each point is an effort level, published by whoever ran the test and compiled by Epoch AI. The further up and to the left, the more score per dollar.
CursorBench
The score scale does not start at zero.
Other models from Google DeepMind
The models from the same lab closest in time, newest first.
| Model | Release | ECI | ELO | LiveBench | Access |
|---|---|---|---|---|---|
| Gemini 4 Argon | Sep 30, 2026 | 164.56 | 1,608 | n/a | Limited access |
| Gemini 3.8 Flash | Sep 2, 2026 | 157.20 | 1,551 | 75.8 | API access |
| Gemini 3.7 Flash | Aug 13, 2026 | 157.22 | 1,551 | 78.8 | API access |
| Gemini 3.6 Flash | Jul 21, 2026 | 154.24 | 1,530 | 73.6 | API access |
| Gemini 3.5 Flash-Lite | Jul 21, 2026 | 145.15 | 1,467 | 63.9 | API access |
| Gemini 3.5 Flash | May 19, 2026 | 154.38 | 1,526 | 74.6 | API access |
| Gemma 4 31B IT | Apr 2, 2026 | 142.86 | n/a | n/a | Open weights (restricted use) |
| Gemma 4 26B A4B | Apr 2, 2026 | 141.81 | 1,495 | n/a | Open weights (unrestricted) |
| Gemini 3.1 Pro | Feb 19, 2026 | 154.67 | 1,528 | 77.0 | API access |
Have you used Gemini 4 Argon? Rate it
Users' opinion, not measurement: the rating is kept apart from the ranking and does not move anyone's position. One vote per browser, no sign-up, and you can change your rating later.
What people ask about Gemini 4 Argon
Can I use Gemini 4 Argon yet?
Only by invitation, for cyber defenders in the Fairwind Program. Google promises to open it later to developers, enterprises and consumers, starting with paid API customers and Google AI Ultra subscribers, with no date. There is no model identifier on the API yet.
How much will Gemini 4 Argon cost?
At the introductory price, US$ 2 per million input tokens and US$ 10 per million output, with 95% off cached input. Afterwards, US$ 4 and US$ 20. Google has not published when the introductory period ends.
Does Argon have a 1 million token context window?
What Google published is a 1 million token output limit, up from 64 thousand: the model can generate up to 1 million tokens in one answer. The input size has not been published.
Is Argon better than GPT-6 Astra and Claude Opus 5.5?
In Google's table, it is ahead in 13 of 19 tests, ties one and loses five: two to Opus 5.5 (Terminal-Bench 4.0 and PostTrainBench) and three to GPT-6 Astra. Competitor scores are self-reported, and no independent source has measured Argon.
Why does Argon have no position in the ranking?
Because none of the ranking's three sources (Epoch AI, Arena and LiveBench) has measured the model yet. When the first one publishes, the page shows the score on its own.
Where does Gemini 4 Argon rank among AI models?
In the ROO3 AI Benchmark, ordered by the Epoch AI ECI, Gemini 4 Argon ranks #5 out of 267 models, with an ECI of 164.56. The ECI is a composite index that combines several hard benchmarks into a single scale. The position changes when a new model comes out, and the page is refreshed six times a day.
Is Gemini 4 Argon better than Claude Fable 5.1?
Across the 19 tests both models took and Epoch AI compiled, Gemini 4 Argon comes out ahead on 9, behind on 9 and tied on 1. It is a test-by-test count with no weighting: a win in math counts the same as a win in coding. To choose between the two, look at the tests that resemble your own use.
What is Gemini 4 Argon best at?
Against models of the same generation (released up to 9 months before and up to 3 months after it), the best relative results of Gemini 4 Argon are: GPQA Diamond (#1 of 73); SimpleQA Verified (#1 of 59); Apex Agents (#1 of 48). The yardstick is its position among the models of that period that took the same test.
Where do the numbers on this page come from?
From the same three sources as the ranking, all openly licensed: Epoch AI (ECI and individual tests, CC BY 4.0), Arena (human-preference ELO, CC BY 4.0) and LiveBench (overall average, CC BY-SA 4.0). ROO3 does not measure any model. When the page shows a figure published by the maker itself, it appears in the written analysis, labeled as the maker's, and never enters the ranking.
Where each number comes from
- Google, Gemini 4 Argon announcement
- Google DeepMind, Gemini models benchmark table
- Google DeepMind, Gemini 4 Argon evaluation methodology (PDF)
- Google DeepMind, Fairwind Program
- Google, Fairwind Program announcement
- Google DeepMind, model cards list (no Argon as of Oct 1, 2026)
- Gemini API, pricing (Gemini 3.1 Pro)
- Gemini API, Gemini 3.1 Pro page
- OpenAI, API pricing
- OpenAI, GPT-6 Astra announcement
- OpenAI, GPT-6.1 Sol announcement
- Anthropic, pricing (documentation)
- Anthropic, model comparison in the Fable 5.1 documentation
Epoch AI, "AI Benchmarking Hub". Published at epoch.ai. Data under the CC BY 4.0 licence, reordered by roo3.co.
The Arena (LMArena) leaderboard, from the public lmarena-ai/leaderboard-dataset on Hugging Face, under the CC BY 4.0 licence. We are not affiliated with the Arena.
LiveBench, by the LiveBench team, published at livebench.ai under the CC BY-SA 4.0 license. The overall average is recomputed by roo3.co with the same formula as the website (average of the categories), and those numbers stay under CC BY-SA 4.0. We are not affiliated with LiveBench.
← See the full ranking