Google announced Gemini 4 Argon on September 30, 2026, describing it as its most advanced frontier AI model to date. The model is built for long-horizon software engineering, enterprise knowledge work in fields like law and finance, and defensive cybersecurity.
The headline technical change is a single-pass output limit of roughly 1 million tokens, up from 64,000 on previous Gemini models. Google says this headroom lets the model sustain deep reasoning across multi-step tasks instead of stopping and restarting around an output ceiling.
Access remains restricted. As of October 1, 2026, no Gemini 4 Argon entry or requestable model ID appeared in the public Gemini API catalog or changelog, and the model was absent from OpenRouter, Vertex AI, Gemini CLI, Cursor, and GitHub Copilot listings.
A mixed benchmark picture
Google's published results show Argon leading in several areas. It scores 77.9% on DeepSWE v1.1, a benchmark for real-world, long-horizon software engineering — ahead of Claude Opus 5.5 at 74.2% and GPT-6 Astra at 74.1%. It ranks first on Zapier's AutomationBench at 51.3%, on the Vals Index at 68.9%, and on LVBench for long video understanding at 91.7%. It ties for first on CWE-bench v1, a vulnerability-remediation benchmark, at 68%.
The results are less favorable elsewhere. Argon scores 55.0% on FrontierSWE v2, trailing GPT-6 Astra at 65.5% and Claude Opus 5.5 at 62.3%. On Terminal-Bench 4.0 it scores 57.4%, nine points behind Opus 5.5 at 66.4%.
These are vendor-reported figures. Google ran Argon's scores through the Gemini API at the highest thinking setting, while competitor numbers largely come from the labs' own disclosures or public leaderboards. No third party had reproduced Google's full table at launch.
What independent evaluators found
Artificial Analysis, which had pre-release access, gives Argon a composite Intelligence Index score of 52.6 — essentially level with GPT-6 Astra at 52.7 and behind Claude Opus 5.5 at 57.6. Its independent Terminal-Bench 4.0 run also places Argon behind Opus 5.5 and Sonnet 5.5.
The same evaluation found a notable tradeoff. Argon posts a 15% hallucination rate on AA-Omniscience versus 51% for GPT-6 Astra, but its factual accuracy is lower — 50% against Astra's 63%. In effect, Argon says "I don't know" far more often rather than guessing.
Artificial Analysis also measured Argon's cost per average task at $1.99 during the promotional period, below Astra's $3.26 — but found Argon averaged 62,000 output tokens per task against Astra's 27,000, meaning its advantage comes from token pricing rather than efficiency.
Bloomberg reported on launch day that some Google staff questioned the model's real-world coding performance, particularly in areas such as front-end design, despite the benchmark results — a discrepancy that prompted public discussion of whether the model had been optimized for benchmark scores. Google disputed the characterization, according to Axios.
Reported internal use cases
Google says Argon is already used internally. Argon agents reportedly rewrote 32,000 lines of hand-written SIMD code in the open-source libgav1 video decoder as safe Rust, producing a version that runs 2.7 times faster than the prior Rust port with identical output.
Another team of Argon agents analyzed fleet telemetry to find memory optimizations across Google data centers, freeing over 300 TiB of memory so far, with projected total savings of 500 TiB to 1 PiB. Google also reports Argon helped quantum researchers beat a published baseline by 40% in minutes. All three examples are vendor-reported and not independently verified.
Security firm Wiz says it used Argon through its Scan for Good initiative to find a critical vulnerability in healthcare software used by hospitals worldwide, one that earlier frontier models had missed.
Pricing and access
During the introductory period, Argon is priced at $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input. Standard pricing rises to $4 and $20 — matching Claude Opus 5.5's listed rates. Google has not said how long the introductory price lasts or how reasoning tokens will be billed.
Initial access is limited to trusted cyber defenders through Google's Fairwind Program, with over 650 partners. Google says broader availability will start with paid API customers and Google AI Ultra subscribers, but has given no date.
What is confirmed, unknown, and awaited
Confirmed: the September 30 announcement, the 1M-token output limit, the published benchmark table, the promotional and standard pricing, and the Fairwind-first rollout.
Unknown: the public API release date, the official model ID, the input context window, the duration of the promotional price, reasoning-token billing, and how Argon performs under independent, workload-specific testing.
Awaited: the paid API launch, third-party reproduction of Google's benchmark figures, and evidence of how the model performs in everyday developer workflows once access widens.