Gemini 4 Argon deep dive: Google's new frontier model

The short version

Gemini 4 Argon (model ID gemini-4-argon) has been GA since October 13 on the paid Gemini API tier, in AI Studio, on Google Cloud's Gemini Enterprise Agent Platform, in the Gemini app and in AI Mode (Google). It was announced on September 30, but for the first two weeks only some partners in the Fairwind cyber-defense program and US government evaluators could use it (endue).

Google calls it its most capable model and says it beats rivals across domains. Independent numbers partly back that up: on Artificial Analysis's index it ties GPT-6 Astra, while Claude Opus 5.5 and Sonnet 5.5 stay ahead (Artificial Analysis).

Timeline and availability

September 30: announcement with gated access. October 13: global GA. Google Cloud's docs list an October 13 release date and a retirement date of October 13, 2027 or later. The model runs in the global, us and eu regions with standard and priority pay-as-you-go plus provisioned throughput (Google Cloud docs).

Who gets what: developers need the paid Gemini API tier; it isn't available on the free tier or through the Google AI Pro plan on the API (pricing). In the app and in Search, Pro and Ultra subscribers get it, but in Google Antigravity, the agent development environment, it's Ultra only. That confused people on Hacker News on launch day (HN).

Variants: one public model, gemini-4-argon, with low, medium and high thinking levels (medium by default). There is also a Gemini 4 Argon Cyber Permissive variant for security teams, behind an allowlist via the Fairwind program. There's no Pro, Flash or Lite sibling yet; Google only says the rest of the family is coming.

Specs: the headline is the 1M-token window

According to the developer docs, input, thinking and output share one 1,048,576-token window (Gemini API docs). Default synchronous output is 65,536 tokens, but with long decode a single response can run up to 1M tokens minus the input. On the Interactions API that requires background mode and a max_output_tokens value above 65,536 (What's new).

Inputs are text, image, video, audio and PDF; output is text only, with no image or audio generation. There's a discrepancy here: Artificial Analysis's model page lists only text and image input (AA model page). Google's docs are the authoritative source, but test it yourself before building on video or audio.

Per the docs, tools include function calling, structured output, code execution, Google Search and Maps grounding, URL context, file search, agentic video understanding and computer use (still in preview). There's no Live API. Batch, flex and priority inference are available.

Migration gotchas: custom temperature, top_p and top_k settings must be removed, and there's no explicit context caching, only implicit. If you're porting older Gemini calls, clean those parameters out.

Pricing: introductory rates until November 30

On the paid Gemini API tier, per million tokens: $2 input and $10 output (including thinking tokens) through November 30, then $4 and $20 from December 1. Cached input is 95% off: $0.10, later $0.20. Search and Maps grounding are free up to 5,000 requests a month, then $14 per 1,000 (pricing).

On Google Cloud, global pricing matches that; non-global regions cost about 10% more ($2.20/$11, later $4.40/$22). The Cyber Permissive variant is billed at the standard rate.

Why it matters: the introductory price matches GPT-6.1 Sol and the standard price matches Opus 5.5's list price. More important than the rate is that Argon writes a lot. In Artificial Analysis's runs it averaged 62,000 output tokens per task versus 27,000 for Astra. That makes it $1.99 per task now and $3.98 at standard pricing (Artificial Analysis). One Hacker News commenter noted that this is already more per task than Astra high or Opus 5.5 high (HN). Budget with the $4/$20 rate.

Benchmarks: what Google claims

Google says Argon ranks first on the Vals Index (68.9%) and leads 12 of the 18 benchmarks in its own table, tying one more. These are the vendor's own figures. The biggest leads are in knowledge work and long context: 19.6% on Harvey's legal agent benchmark (next best 6.7%), 84.2% on GraphWalks between 256K and 1M tokens (others 65–72%), and 65.4% on Vals Finance Agent v2 (endue summary).

Source: Google, vendor-reported (Google blog)

Where it loses: it trails Astra by 10.5 points on FrontierSWE v2 and Opus 5.5 by 9 points on Terminal-Bench 4.0. Coding is mixed: of four agentic coding benchmarks, it wins two and comes last on the other two.

Methodology matters too. According to endue's analysis, Google ran several Argon scores itself, took the best of three runs on OSWorld-2.0, gave Argon a 6x verifier timeout on Terminal-Bench Science, and ran LVBench with different frame counts per model. Claude Sonnet 5.5 and GPT-6.1 Sol are missing from the table. The direction is more reliable than the size of each gap.

Further vendor claims: 87% on the ERQA embodied-reasoning benchmark and, by Google's internal measure, 35% more agent-submitted code in Antigravity with nearly a third fewer rollbacks. None of these has been independently verified.

Independent evaluations

Artificial Analysis scores Argon at 53 on its Intelligence Index in high mode: level with Astra, one point above GPT-6.1 Sol and 23 points above Gemini 3.1 Pro Preview. Claude Opus 5.5 still leads at 58 (Artificial Analysis).

Agentic work is where it improved most: first on AutomationBench-AA at 78%, seven points ahead of Sonnet 5.5. On Terminal-Bench 4 it scored 57%, behind Sonnet 5.5 (64%), Opus 5.5 (60%) and Astra (59%).

The most-quoted number is hallucination. On AA-Omniscience, Argon's hallucination rate is 15%, versus 51% for Astra and 54% for GPT-6.1 Sol. What that means matters: AA measures the share of wrong answers among the questions a model didn't answer correctly. Argon's accuracy on the same test is 50%, 13 points below Astra. Put simply, it knows fewer facts but far more often says it doesn't know. The combined scores end up about even (42 vs. 43).

Community reaction

In the GA thread, a team running a coding benchmark said that in its 100-environment test Argon sits between Fable 5 and Fable 5.1, roughly two months behind the frontier. They said Google has fixed older Gemini models' trouble working efficiently in a harness, but Argon was the first near-frontier model they'd caught in infinite reasoning loops in a long time, and it sometimes failed to delimit reasoning blocks properly (HN).

At the September announcement, the top complaint was access, since almost nobody could try it for two weeks. There was benchmark fatigue too, and many criticised Google's Antigravity tooling rather than the model. Bloomberg's anonymous sources said internal testers found it weaker in real work than on benchmarks, which Google denied (endue).

Safety

Under its Frontier Safety Framework, Google says in the model card that it cannot rule out Argon reaching the critical capability levels (CCLs) for biological information risk and cybersecurity, or the tracked capability level (TCL) for stealth and situational awareness. In each case it prepared a safety case or residual-risk assessment that, by Google's assessment, brings the risk below the threshold (model card).

Google says the model solved all 12 simulated end-to-end cyberattack scenarios, up from five for the previous generation (safety blog). On the defensive side, by Google's own figures: external bio red-teaming unlocked under 1% of the bio sub-tasks tested, violations on an adversarial set roughly halved, shortcut attempts on coding benchmarks fell to 2%, and on Gray Swan's indirect prompt injection test attacks succeeded 0.7% of the time across 15 attempts.

Source: Google, vendor-reported, Gray Swan IPI test (Google blog)

There's also an automated misalignment monitor that reads the agent's reasoning and actions and blocks unintended workarounds in real time. Fairwind partners, however, get the model without cyber guardrails, under strict conditions.

Limitations and open questions

It's verbose, which raises real costs. Long outputs need care: a 1M-token response can take hours, and most proxies and serverless runtimes will drop the connection long before that. Feedback on front-end work is mixed, and the reasoning-loop and block bugs are community reports, not systematic measurements. There's no free API tier, and the lack of explicit caching won't reassure users who were already unhappy with Google's caching.

What to do now

If long context, document analysis or knowledge work is your main use case, Argon is now one of the strongest candidates, and the introductory price makes testing cheaper until November 30. For coding agents, compare it against Opus 5.5 and Astra in your own harness. Log token counts everywhere, because verbosity, not list price, decides the real cost. Where wrong answers are expensive, the low hallucination rate can be a real advantage, but only if your system can handle an "I don't know".