Gemini 4 Argon is broadly available. Its wins are real. So are its coding gaps.

Google announced Gemini 4 Argon on September 30. Access was limited: approved defenders in its Fairwind cybersecurity program only. The model shipped with a specific framing — built for complex reasoning, enterprise-grade security work, long-context tasks. Then, on October 13, it opened to any developer on a paid Gemini API or AI Studio plan.

That's the practical news. The complication is the benchmark picture. And the real complication is December 1.

What the benchmark table actually shows

Google published a benchmark table at launch. The headline number: 77.9% on DeepSWE v1.1, ahead of GPT-6 Astra at 74.1% and Claude Opus 5.5 at 74.2%. Meaningful gap — about 3.7 points.

The same table has other rows.

FrontierSWE v2: Argon at 55.0%. Astra at 65.5%. Opus 5.5 at 62.3%. That's not a small gap — that's 10 points off the leader. Terminal-bench 4.0: Argon at 57.4%, Opus 5.5 at 66.4%. Argon wins on knowledge-work evaluations and trails on real-world coding evaluations. That split is specific enough that it tells you something useful about where the model is and isn't the right choice.

Important caveat before relying on any of this: NeuralTrust reported that no independent third party had reproduced Google's benchmark table at launch. All figures above come from secondary blogs citing Google's own model card. The DeepSWE v1.1 win is Argon's; the FrontierSWE gap is also real. But "real" here means "appears in Google's documentation," not "independently verified." Treat the numbers as directionally useful, not authoritative.

The other number worth keeping is on LVBench: 91.7% on long video understanding. That's a different evaluation entirely — not coding, not agent task completion. If you're building anything that involves long video analysis, that figure is relevant even if the coding benchmarks are mixed.

Source spread

• Google Blog — Gemini 4 Argon launch — [hype]. The official announcement framing. Leads with cybersecurity use cases and the Fairwind rollout.
• VentureBeat — Broad availability, pricing — [builder]. Clean reporting on October 13 GA date, pricing structure, and December 1 transition.
• NeuralTrust — Benchmark analysis — [skeptic]. Notes the absence of independent benchmark verification at launch; good on FrontierSWE v2 specifics.
• DEV Community — Max Quimby's benchmark breakdown — [builder]. Breaks down which of 18 benchmarks Google won or lost; found coding specifically as the gap.

Pros and cons

What this model has going for it:

• The DeepSWE gap is real. 3.7 points over Opus 5.5 and Astra on a reasonably comprehensive software engineering evaluation is not nothing, especially if your use case looks more like knowledge-work tasks than shell-driven coding sessions.
• The 1M output token limit matters for code generation. If you're generating large files, translations, or long documents, having 1M output tokens instead of 64K changes what's possible in a single call.
• The introductory pricing is genuinely interesting. $2 per million input tokens puts Argon in a competitive tier for the next seven weeks. If you're evaluating it against Opus 5.5 or Astra for a specific use case, now is the time to do it — you won't be doing it at these economics after December 1.

What to watch:

• The FrontierSWE gap is also real. If your application involves autonomous agent loops doing real file manipulation, shell execution, or multi-step code workflows — Terminal-bench and FrontierSWE are more relevant than DeepSWE. Argon trails by 7–10 points there.
• Benchmark self-reporting at a frontier model launch is the norm. That doesn't mean the numbers are wrong. It means they haven't been tested on your actual workload, by someone who wasn't paid to report them.
• The December 1 cliff. $2/$10 becomes $4/$20 in forty-four days. If your evaluation goes well and you're building something that needs Argon specifically, you need to know what the model costs at standard pricing before you ship.

Samwise's take

What builders need to know

• Access: Paid Gemini API or AI Studio plan. Antigravity CLI is limited to Ultra subscribers. Free tier not yet included.
• Economics window: $2 input/$10 output per million tokens through November 30. Standard pricing ($4/$20) takes effect December 1. Build your evaluation into the next 44 days.
• 1M output token limit (up from 64K on previous models) — relevant for code generation, large document tasks. Confirm input context limits against the official API docs before building against them.
• Coding workloads: Run Terminal-bench and FrontierSWE-style tasks on your specific codebase before committing. The benchmark gap vs. Opus 5.5 is real and specific enough to matter in production.
• Benchmark verification: No independent third-party verification at launch. Use the numbers as a starting hypothesis, not a decision.

Further reading

• Google Blog — Gemini 4 Argon announcement
• VentureBeat — GA and pricing details
• NeuralTrust — Benchmark analysis and caveats
• DEV Community — Which benchmarks Argon won and lost
• DataCamp — Pricing, context, and access summary