Gemini 4 Argon is broadly available, but its benchmark lead has limits

Gemini 4 Argon became globally available to developers and enterprises through the Gemini API on October 13, with access also offered through Google Antigravity, the Gemini app, and AI Mode in Search for eligible Google AI subscribers. Google lists an introductory API price of $2 per million input tokens and $10 per million output tokens through November 30, after which standard prices are $4 and $20. The model card describes Argon as a new model rather than a modification or fine-tune of an earlier Gemini model, and says its combined input-output context window reaches up to 1 million tokens. That makes the availability change concrete for teams that want to test long documents or multi-step work through an API, but the low introductory price is time-limited.

This is a second stage in Argon’s rollout. TechCrunch reported on September 30 that access was then limited to selected cybersecurity partners through Google’s Fairwind Program. Google’s October 13 announcement now describes global API availability, while retaining separate Fairwind access for vetted defenders who need the model without cyber guardrails. The distinction matters: general availability does not mean every user receives the unmitigated version. Google’s own announcement says eligible consumer access is restricted by subscription tier, and the API carries its own terms and setup requirements.

Google’s model card gives a more mixed picture than a single “frontier” label suggests. In Google-reported evaluations, Argon scores 99.7% on GraphWalks up to 128K tokens and 84.2% from 256K to 1M, ahead of the listed GPT-6 Astra and Claude comparison models. It also leads the listed Vals Index at 68.9%, and DeepSWE v1.1 at 77.9%. But Argon trails GPT-6 Astra on FrontierSWE v2 and Terminal-Bench Science, and Claude Opus 5.5 on Terminal-bench 4.0 and PostTrainBench. Those results make a better case for testing Argon on long-context research and knowledge work than for assuming it is the best choice for every coding agent or scientific workflow.

These are provider-published benchmark results, not an independent head-to-head replication. Google’s model card describes the evaluation set and links to its methodology, while TechCrunch’s September coverage documented the earlier restricted rollout and noted that Google’s comparative benchmark claims came from Google’s announcement. The launch also brings meaningful operational caveats: Google lists hallucinations, occasional slowness or timeouts, and a March 2026 knowledge cutoff among Argon’s limitations. For technical teams, the useful next step is to compare it on their own representative tasks, including latency, token use, and failure handling, and to treat the promotional price as a temporary evaluation window rather than a settled long-term cost.