Google DeepMind launches Gemini 4 Argon, independently corroborated atop three measurement platforms
2026-09-30
Google DeepMind released Gemini 4 Argon on September 30, 2026, positioned for coding, enterprise knowledge work, and cyber defense. Initial access is limited to government users and 'trusted cyber defenders' in Google's Fairwind Program, with Google saying it will open broader developer, enterprise, and consumer access 'as soon as possible' after refining guardrails. Google's own published benchmarks claim first place on 13 of 19 tests against GPT-6 Astra and Claude Opus 5.5, including 77.9% on DeepSWE versus 74.2% (Opus 5.5) and 74.1% (Astra). Independent measurement firms report a strong result too: Artificial Analysis scored Argon 53 on its Intelligence Index (tying GPT-6 Astra, ahead of GPT-6.1 Sol's 52) and ranked it #1 on AutomationBench-AA (77.5%); Vals AI ranked it #1 on the Vals Index (68.9%) and LMArena placed it #1 in Text Arena (1525). The model supports up to 1M output tokens via a new 'Long Decode Continuation' API feature. Standard pricing is $4/$20 per 1M input/output tokens, with a 50%-off introductory rate ($2/$10) and no stated end date. Google reports Argon agents freed over 300 TiB of internal data-center memory and are migrating 800K+ lines of C/C++ kernel code to Rust.
Significance 5: Independently corroborated (Artificial Analysis, Vals AI, LMArena) frontier-class release topping three separate measurement platforms simultaneously, which the rubric treats as top-tier even though real-world access remains restricted to a government/cyber-defense preview.
Implications · machine-drafted, not owner judgment
Argon's restricted initial rollout (government + cyber-defense preview only) delays any real verdict impact — Opus 4.8 and GPT-5.2 remain the practical picks for coding-agent and hard-problems until Google widens access. If Google opens access with these scores holding up under independent re-test, this would be the first time in months Google plausibly leads multiple trusted benchmarks at once, which per the google-strategy-tax thesis's own falsifier criteria would meaningfully challenge the view that Google can't compete in the commercial race — notably because Argon is explicitly positioned for coding/enterprise/cyber-defense rather than consumer use, a genuine break from the consumer-first release pattern that thesis has tracked all month.
- Broader developer/enterprise access opening within the next month
- An independent SWE-bench Verified or Terminal-Bench run once public access exists
- Whether access stays cyber-defense/government-gated longer than Google's 'as soon as possible' framing suggests
Sources
- primary Gemini 4 Argon: our next era of frontier intelligence retrieved 2026-10-01
- press Google releases Gemini 4 Argon, called its most powerful model yet retrieved 2026-10-01
- curator [AINews] Gemini 4 Argon: GDM's answer to Astra/Fable, with 1M output retrieved 2026-10-01
Affected axes
Hard reasoning · Coding · Agentic / tool use · Long context · Human preference