Gemini 4 Argon is Google DeepMind’s September 30, 2026 frontier model — the start of the Gemini 4 generation and Google’s answer to GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5. Google positions it for complex, long-horizon work in software engineering, enterprise knowledge work (legal, finance) and cybersecurity defense.
One important caveat: at launch Argon is not broadly available. It is rolling out to trusted cyber defenders through the new Fairwind Program while Google takes part in the US government’s voluntary pre-release access process; developers, enterprises and consumers follow “as soon as possible”, beginning with paid API customers and Google AI Ultra subscribers.
Benchmarks
Google’s published comparison (all Google-reported; methodology on DeepMind’s evals page). Each lab runs rivals under its own harness, so numbers disagree across vendors — Anthropic’s and OpenAI’s tables list different scores for the same models.
| Benchmark | What it measures | Gemini 4 Argon | GPT-6 Astra | Claude Fable 5.1 | Claude Opus 5.5 |
|---|---|---|---|---|---|
| DeepSWE v1.1 | Long-horizon software engineering | 77.9% | 74.1% | 67.4% | 74.2% |
| Vibe Code Bench | Agentic coding | 91.9% | 89.6% | 90.3% | 90.3% |
| FrontierSWE v2 | Agentic coding | 55.0% | 65.5% | 56.3% | 62.3% |
| Terminal-Bench 4.0 | Agentic coding | 57.4% | 58.2% | 57.9% | 66.4% |
| Vals Index | Knowledge work (finance, coding, legal, tax) | 68.9% | 63.1% | 65.8% | 67.0% |
| AutomationBench | Cross-app business workflows | 51.3% | 41.4% | 31.4% | 42.5% |
| Vals Finance Agent v2 | Multi-step financial research | 65.4% | 53.5% | 58.9% | 58.6% |
| Harvey Legal Agent Benchmark | Legal research and drafting | 19.6% | 5.4% | 6.7% | 3.8% |
| Terminal-Bench Science 0.1 | Science and math | 57.6% | 68.1% | 52.6% | 63.3% |
| LABBench 2 | Science and math | 88.8% | 85.4% | 68.6% | 73.1% |
| GraphWalks 256K–1M | Long context (BFS F1) | 84.2% | 71.8% | 65.0% | 66.8% |
| OSWorld-2.0 (offline, partial) | Computer use | 69.2% | 72.6% | — | — |
| Chartography | Multimodal understanding | 71.6% | 71.0% | 46.2% | 66.3% |
| LVBench | Long-video understanding | 91.7% | 87.5% | 79.7% | 83.7% |
| CWE-bench v1 | Vulnerability remediation | 68.0% | 68.0% | 58.0% | 67.0% |
What Google highlights
- 1M-token output. The output limit rises from 64K to 1M tokens so the model can think and generate hundreds of thousands of tokens in a single trajectory.
- Internal use at Google. Argon agents found memory optimizations that free over 300 TiB of memory across data centers (an estimated 500 TiB to 1 PiB once fully rolled out), are migrating C/C++ to Rust at scale (from re2 and libgav1 up to 800K+ lines for the Fuchsia Zircon kernel), and beat a published quantum-circuit baseline by 40% in minutes. A libgav1 rewrite replacing 32K lines of SIMD code ran 2.7x faster than the earlier Rust port with identical output.
- Cyber defense. Argon can autonomously find, validate and patch vulnerabilities. Trusted defenders get it without cyber guardrails. Wiz reports it found a critical vulnerability exposing personal data in healthcare software that earlier frontier models had missed.
- Safeguards. Activation monitoring for misuse, leading prompt-injection robustness on Gray Swan’s indirect-injection benchmark, chain-of-thought and action monitors for misalignment, and sealed sandboxes for high-risk evaluations.
Pricing and availability
| Item | Detail |
|---|---|
| Introductory price | $2 input / $10 output per 1M tokens |
| Cached input | 95% off the input price |
| Output limit | Up to 1M tokens |
| Access now | Trusted cyber defenders via the Fairwind Program |
| Access next | Paid API customers and Google AI Ultra subscribers, then developers, enterprises and consumers |
Google has not published a context window, model ID or full spec sheet yet. Its sibling, Gemini 3.8 Flash, is the Gemini model you can use today.
How Argon compares
- vs GPT-6 Astra: one-fifth of Astra’s list price at launch and ahead on most knowledge-work and multimodal rows, but behind on FrontierSWE v2, Terminal-Bench Science and OSWorld-2.0.
- vs Claude Opus 5.5: Opus 5.5 ($4 / $20) is generally available now and leads Terminal-Bench 4.0 by about nine points; Argon leads DeepSWE v1.1 by about four points.
- vs Claude Fable 5.1: Argon leads on every row of Google’s table except FrontierSWE v2 and Terminal-Bench 4.0.
- vs Gemini 3.8 Flash: Flash is the cheap, available workhorse; Argon is the capability ceiling and still gated.
Limitations
- Not generally available. You cannot plan production work around Argon until Google confirms broad API availability and final pricing.
- Self-reported and cross-vendor numbers. Google, OpenAI and Anthropic each publish tables where the same rival model scores differently.
- Open spec questions. Context window, API model ID, rate limits and rollout dates are not yet public.
For the wider picture, see the 2026 LLM guide.
Best for
- Long-horizon software engineering and codebase-scale migrations (C/C++ to Rust at Google)
- Cybersecurity defense — finding, validating and patching vulnerabilities
- Enterprise finance, legal and tax agents (Vals Index, Harvey's Legal Agent Benchmark)
- Long-video and chart understanding that mixes visual and document inputs
Pros & cons
Strengths
- Leads Google's published table on DeepSWE v1.1 (77.9%), AutomationBench (51.3%), Vals Index (68.9%), LABBench 2 (88.8%) and LVBench long-video understanding (91.7%)
- Output limit raised to 1M tokens (from 64K), letting it reason and generate hundreds of thousands of tokens in one trajectory
- Priced at $2 / $10 per 1M at launch with cached input at 95% off — a fifth of GPT-6 Astra's and Claude Fable 5.1's list price
- Google's own engineers report concrete wins — a libgav1 decoder 2.7x faster than the prior Rust port and 300+ TiB of data-center memory freed by Argon agents
Limitations
- Not generally available — rolling out first to trusted cyber defenders via the Fairwind Program, with developers, enterprises and consumers "as soon as possible"
- Trails on several rows of Google's own table — Terminal-Bench 4.0 57.4% (Opus 5.5 66.4%), FrontierSWE v2 55.0% (Astra 65.5%), Terminal-Bench Science 0.1 57.6% (Astra 68.1%), OSWorld-2.0 69.2% (Astra 72.6%)
- Pricing is labelled introductory; final price after the launch window is not stated
- Context window and full spec sheet have not been published yet, and all scores are Google-reported
Frequently asked questions
When was Gemini 4 Argon released?
Google DeepMind announced it on September 30, 2026. It began rolling out to a set of trusted cyber defenders through the Fairwind Program; broad availability — starting with paid API customers and Google AI Ultra subscribers — was described as "soon" and tied to the US government's voluntary pre-release access process.
How much will Gemini 4 Argon cost?
Google announced an introductory price of $2 per 1M input tokens and $10 per 1M output tokens, with cached input tokens at 95% off the input price (about $0.10 per 1M). That is the same list price as GPT-6.1 Sol and a fifth of GPT-6 Astra's $10 / $50.
Is Gemini 4 Argon better than GPT-6 Astra and Claude Opus 5.5?
On Google's own table it leads on DeepSWE v1.1 (77.9% vs Astra 74.1% and Opus 5.5 74.2%), AutomationBench (51.3%), Vals Index (68.9%) and LVBench (91.7%), but trails Opus 5.5 on Terminal-Bench 4.0 (57.4% vs 66.4%) and Astra on Terminal-Bench Science 0.1 (57.6% vs 68.1%). Labs report each other's numbers under their own harnesses, so wait for independent tests.
What is the Fairwind Program?
Google's trusted-access program for cybersecurity. It gives vetted government authorities, critical-infrastructure operators and software maintainers prioritized access to Argon and Gemini 3.8 Flash Cyber with fewer cyber guardrails, so defenders can use the models' full vulnerability-finding and patching capability.
What is Gemini 4 Argon's output limit?
Up to 1M output tokens, up from 64K on previous Gemini models, so a single run can sustain hundreds of thousands of reasoning and output tokens.