Gemini 4 Argon is Google’s first Gemini 4 model, launched on September 30, 2026, and it leads most of the benchmarks Google published, including a 77.9% score on the DeepSWE v1.1 software engineering test. After a two week period limited to vetted cyber defenders, Google opened it to developers through the Gemini API and to Google AI Pro and Ultra subscribers in mid October. API pricing is $2 per million input tokens and $10 per million output tokens through November 30, then $4 and $20 from December 1.
One version of Argon is still not public: a build with its cyber guardrails switched off, reserved for Google’s own teams and members of the Fairwind Program. This guide covers the benchmark numbers, the price change, how to get access today, and what the gated cyber build means.
Key takeaways
- Launched September 30, 2026; broader availability announced October 13
- Price: $2 input and $10 output per million tokens until November 30, then $4 and $20
- Output limit of 1 million tokens, up from 64,000 on earlier Gemini models, with a 1,048,576 token context window
- Leads or ties 13 of 18 benchmarks Google disclosed, but trails on FrontierSWE v2 and Terminal-bench 4.0
- Guardrail free build goes only to vetted defenders through the Fairwind Program
Gemini 4 Argon at a glance
| Spec | Gemini 4 Argon | What it means |
|---|---|---|
| Release | September 30, 2026 (Fairwind); October 13 broader rollout | Public developers got it about two weeks after defenders |
| Context window | 1,048,576 tokens | Large codebases or contract sets fit in one prompt |
| Max output | 1 million tokens, up from 64K | Whole module rewrites or long reports in a single pass |
| Input price | $2 per 1M until Nov 30; $4 from Dec 1 | Doubles in December |
| Output price | $10 per 1M until Nov 30; $20 from Dec 1 | Thinking tokens are billed as output |
| Cached input | $0.10 per 1M until Nov 30; $0.20 from Dec 1 | Implicit caching only, per Google’s price page |
| Batch | $1 / $5 until Nov 30; $2 / $10 from Dec 1 | Half the standard rate |
Sources: Google’s Gemini API pricing page (checked November 3, 2026), Google’s launch post and Let’s Data Science.
Benchmarks: where Argon leads and where it does not
VentureBeat counted 18 disclosed benchmarks in Google’s comparison. Argon leads outright on 12 and ties for first on one, CWE-bench v1, for 13 of 18. GPT-6 Astra leads on three and Claude Opus 5.5 on two. The scores below come from Google’s own table as reported by VentureBeat.
| Benchmark | Gemini 4 Argon | Claude Opus 5.5 | GPT-6 Astra |
|---|---|---|---|
| DeepSWE v1.1 (software engineering) | 77.9% | 74.2% | 74.1% |
| GraphWalks (large context reasoning) | 84.2% | 66.8% | 71.8% |
| AutomationBench (Zapier) | 51.3% | 42.5% | 41.4% |
| Vals Finance Agent v2 | 65.4% | 58.6% | 53.5% |
| Harvey Legal Agent Benchmark | 19.6% | 3.8% | 5.4% |
| LVBench (long video) | 91.7% | 83.7% | 87.5% |
| CWE-bench v1 (vulnerability fixes) | 68% (tie) | 67% | 68% (tie) |
| Gray Swan prompt injection success, lower is better | 0.7% | 1.0% | 8.5% |
| FrontierSWE v2 | 55.0% | n/a | 65.5% |
| Terminal-bench 4.0 | 57.4% | 66.4% | n/a |
Vendor reported numbers via VentureBeat. n/a means the score was not given in the coverage we could verify.
The pattern is clear. Argon is strongest on long, knowledge heavy work: repository scale coding, large context reasoning, finance and legal agent tasks, and long video. It trails on the tests closest to driving a computer or terminal on its own, FrontierSWE v2 and Terminal-bench 4.0, plus Terminal-Bench Science and PostTrainBench. If your use case is an autonomous agent clicking through tools, those losses matter more than the DeepSWE lead.
There is also a caution on real world performance. Bloomberg, cited by The Next Web, reported that some Google employees felt the model did better on benchmarks than in daily work, including some coding tasks. Google said it would be inaccurate to say Gemini 4 underperforms in areas such as coding. Independent testing is the way to settle it.
The Fairwind Program and the unrestricted cyber build
Google built two versions of Argon. The one developers and subscribers use has cyber guardrails on. A second build, with those guardrails removed, goes only to Google’s internal teams and to Fairwind members. Fairwind launched in early September alongside Gemini 3.8 Flash Cyber, and Google’s program page now lists Gemini 4 Argon as its exclusive model.
| Fairwind detail | What Google says |
|---|---|
| Who can apply | Governments and national cyber authorities, critical infrastructure operators (healthcare, telecom, energy, finance), core technology platforms, and academic labs focused on defensive benchmarking |
| Vetting | Background checks covering security history and record of ethical operations |
| Allowed uses | Authorized threat simulation, reverse engineering and malware analysis for defensive and academic research |
| Not allowed | Malicious tasks such as creating malware; sharing, reselling or redistributing access |
| Account rules | Access limited to internal security, incident response or penetration testing teams, with phishing resistant MFA and tracked use |
| How to apply | Through a Google Cloud account team; no fixed response time |
Source: Google DeepMind Fairwind Program page.
The early proof point came from Wiz, which uses Argon in its Scan for Good initiative to find and fix exposures in public infrastructure for free. Google says Argon found a critical vulnerability exposing sensitive personal information in healthcare software used by hospitals worldwide, one earlier frontier models missed, and that it beat Gemini 3.8 Flash Cyber on Wiz’s black box penetration testing benchmark. Argon also went through the US government’s voluntary pre release model access process.
The case for and against gating by identity
Supporters argue that defenders need the strongest tools before attackers get similar ones, and that a vetted list is better than no limits at all. The obvious weakness is that a model without guardrails is only as safe as the least secure account that can reach it, which is why Google’s rules require phishing resistant MFA and tracked use. The same debate is playing out in Washington, where proposals such as the Bernie Sanders AI safety bill push for much tighter limits on frontier systems.
Latest update, November 2026: who can use Argon now
On October 13, Google expanded access beyond Fairwind. According to Let’s Data Science, Argon is now in the Gemini API, Google AI Studio and Google’s platform for enterprises, in Antigravity for Google AI Ultra subscribers, and in the Gemini app and AI Mode in Search for Google AI Pro and Ultra subscribers. The introductory API price runs through November 30. Google’s pricing page lists the higher rate starting December 1, 2026.
| Who you are | How to get Gemini 4 Argon |
|---|---|
| Developer | Gemini API or Google AI Studio with a paid key; budget at $4 and $20 per million tokens from December 1 |
| Google AI Pro subscriber | Gemini app and AI Mode in Search |
| Google AI Ultra subscriber | Gemini app, AI Mode in Search and Antigravity |
| Security team at an eligible organization | Apply for Fairwind through your Google Cloud account team for the guardrail free build |
| Free Gemini user | No Argon access announced; see our guide on how to use Google Gemini for what the free tier includes |
What to keep in mind
- The cheap rate ends November 30. Anything you prototype now will cost twice as much per token in December
- Benchmarks are Google’s. Replication by outside groups is still thin
- Agent gaps are real. Rivals still lead on terminal and computer control tests
- Dual use risk. The ability that finds and patches a bug can also find one to exploit
What it means for you
If you build software, Argon is worth testing on long context work, where its lead looks largest, and on code review across a whole repository. Compare it against GPT-6 Astra on any task that involves driving a terminal. Because Claude Opus 5.5 is priced at the same $4 and $20 that Argon moves to in December, according to VentureBeat, the decision should come down to quality on your own tasks, not price.
If you use Gemini as a subscriber, Argon in the app is mainly a quality upgrade for harder questions. For consumers comparing assistants, our guide to accessing GPT-5 for free shows what OpenAI offers without a subscription. And if your concern is security on your own devices rather than in the enterprise, start with Android’s Advanced Protection mode. Agents acting outside their owners’ intent remain a live problem, as shown when an AI agent wrote a hit piece on a developer. Follow our AI and software coverage for updates when the December pricing takes effect.
FAQ
What is Gemini 4 Argon?
Gemini 4 Argon is Google’s first Gemini 4 model, released September 30, 2026. Google positions it for software engineering, legal and finance work, long context analysis and cyber defense, with up to 1 million tokens of output.
How much does Gemini 4 Argon cost?
On the Gemini API it costs $2 per million input tokens and $10 per million output tokens through November 30, 2026. From December 1, 2026, the price is $4 input and $20 output.
Can I use Gemini 4 Argon now?
Yes. Since mid October it has been available in the Gemini API and Google AI Studio, in the Gemini app and AI Mode for Google AI Pro and Ultra subscribers, and in Antigravity for Ultra subscribers.
What is the Fairwind Program?
Fairwind is Google’s program that gives vetted cyber defenders, such as governments and critical infrastructure operators, access to a version of Gemini 4 Argon without cyber guardrails for defensive work. Applicants go through background checks and apply through a Google Cloud account team.
Is Gemini 4 Argon better than GPT-6 Astra and Claude Opus 5.5?
On Google’s published table, Argon leads or ties 13 of 18 benchmarks, including DeepSWE v1.1 at 77.9%. GPT-6 Astra leads on FrontierSWE v2 and Claude Opus 5.5 leads on Terminal-bench 4.0, so the best choice depends on the task.
What is the context window of Gemini 4 Argon?
Argon accepts 1,048,576 input tokens and can produce up to 1 million output tokens, up from 64,000 on earlier Gemini models.
Sources
- Google: Gemini 4 Argon launch post
- Google AI for Developers: Gemini API pricing
- Google DeepMind: Fairwind Program
- VentureBeat: Google retakes the benchmark lead, but in limited release
- The Next Web: Gemini 4 Argon reaches cyber defenders first
- Let’s Data Science: Google expands Gemini 4 Argon access
About this article: GeekBlog covers U.S. technology news, AI, phones, smartwatches and gaming. Every story is written and checked under our Editorial Policy. Spotted a mistake or have a story tip? Contact our editors.

