Gemini 4 Argon is now generally available: price, who gets it and what Google claims

On October 13, 2026, Google made Gemini 4 Argon, its newest Gemini frontier model, generally available two weeks after announcing it. Developers can call it through the paid Gemini API as gemini-4-argon, and Google AI Pro and Ultra subscribers get it in the Gemini app and AI Mode in Search. The Gemini 4 Argon price is $2 per million input tokens and $10 per million output tokens through November 30, 2026, then $4 and $20. Google Cloud's post says Argon is "available for everyone on Google Cloud," but there is still no free tier.

This story is based on Google's October 13 posts, the Gemini API docs, the Gemini app release notes, Google's plan pages and the Argon model card, checked October 14, 2026.

The short version

• What changed: On September 30, Argon went only to vetted cyber defenders in Google's Fairwind Program (our original story). Since October 13 it's a stable, paid product.
• Who gets it: paid Gemini API users, AI Studio, Gemini for business, AI Pro and Ultra subscribers in the Gemini app and AI Mode, and Ultra subscribers in Antigravity.
• Price: $2 / $10 per million tokens through November 30, 2026; $4 / $20 from December 1. Cached input is 95% off.
• Limits: 1,048,576 tokens shared by input and output. Default output is 65,536 tokens, and up to about 1M with "long decode."
• Claims: mostly vendor-reported; Vals AI's independent Vals Index ranks Argon first.

What changed since September 30?

Google's September 30 post promised a wider release "starting with paid API customers and Google AI Ultra subscribers," with no date. The October 13 rollout went further than that. AI Pro subscribers ($19.99 a month in the US) also got Argon in the Gemini app and AI Mode, per the Gemini app release notes: "Starting today, Gemini 4 Argon is rolling out globally for Google AI Pro & Ultra subscribers in the Gemini app."

Three open questions from our first story now have answers:

• The introductory price ends November 30, 2026. The pricing page lists the higher rates "starting December 1, 2026."
• The context window is 1,048,576 tokens, shared by input, thinking and output, per Argon's API model page. Requests default to 65,536 output tokens; longer outputs need background execution, which Google calls long decode.
• The knowledge cutoff is March 2026, per the model card.

The public model also isn't the one Fairwind partners got. Google Cloud says general-release Argon "has cyber mitigations that block offensive tasks such as pentesting and exploitation." Unmitigated access stays with vetted defenders, and Google says it is expanding Fairwind to "hundreds of additional" partners, with rolling applications.

How much does Gemini 4 Argon cost, and who gets it?

Last verified: October 14, 2026 (ai.google.dev pricing, models and Google AI plans pages; gemini.google subscriptions page; Google and Google Cloud posts of October 13)

Output prices include thinking tokens. Google's model page lists thinking levels low, medium (the default) and high, so the same prompt can cost very different amounts. Our reasoning models explainer covers why. Our guide on how to use Gemini 4 Argon has the setup steps and a worked cost example.

What does Google claim Argon can do?

These are Google's claims unless noted. Per Google's methodology document, most competitor figures in its table come from those vendors or public leaderboards, with settings that differ by model.

• ERQA (embodied reasoning for robots): 87%, which Google says beats all other frontier models. Vendor-reported.
• CWE-bench v1 (fixing security vulnerabilities): 68%, tied with OpenAI's GPT-6 Astra in Google's table. Vendor-reported.
• DeepSWE v1.1 (long-horizon coding): 77.9%, against 74.2% for Claude Opus 5.5 and 74.1% for GPT-6 Astra, on Google DeepMind's model page. This number dates from September 30. Vendor-reported.
• Gray Swan Indirect Prompt Injection: an attack success rate of 0.7% across 15 adaptive attempts. Google's methodology document says the result was "sourced from Gray Swan," but Google doesn't publish the configuration or scenario count. Lower is better. Our prompt injection explainer explains why even a low rate isn't a guarantee.

The same Google table still shows Argon behind Claude Opus 5.5 on Terminal-bench 4.0 (57.4% vs. 66.4%). Our three-way comparison puts these numbers side by side.

The independent result. Vals AI's Vals Index, a weighted average of mostly private agentic tasks in finance, coding, legal and tax work, listed Argon first at 68.90% on October 13, ahead of Claude Sonnet 5.5 (67.04%) and Opus 5.5 (66.97%). Vals calls its GDP weighting a "vast oversimplification," and most of its tasks aren't public.

What are enterprise customers saying?

Google Cloud's launch post quotes early-access customers. These are statements Google published, not independent results:

• ServiceNow says Argon delivered "roughly 20% improved task success" over Gemini 3.8 Flash on its NowAI-Bench.
• Harvey says it "saw substantial improvements over Gemini 3.1 Pro" across its legal agent evaluations.
• HubX says Argon identified "100% of top-priority vulnerabilities with 0% fabricated vulnerabilities across verified claims," without publishing a sample size.

What about safety?

In an October 13 safety post, Google says Argon solved all 12 of its simulated end-to-end cyberattack scenarios, up from five for the previous generation, and that the public model sits behind three layers of real-time protection. The model card says Google "cannot rule out" that Argon has reached its Frontier Safety Framework critical capability levels for biological information risk and for cyber. Google's safety cases assess both risks as below those levels. Our model card report goes through the details.

On government testing, Google's posts mention only its participation in the US government's voluntary pre-release access process. On October 12, AFP reported that UK AI minister Kanishka Narayan said Britain's AI Security Institute "had been granted access to Google's most powerful AI model, Gemini 4 Argon." Google's October 13 posts don't mention the UK institute.

What's confirmed, reported and uncertain

• Confirmed (Google docs and posts): October 13 general availability, the gemini-4-argon ID, prices and the November 30 cutoff, no free API tier, the 1,048,576-token unified window, plan access, the March 2026 knowledge cutoff, cyber mitigations on the public model, and the Fairwind expansion.
• Vendor-reported: ERQA, CWE-bench, DeepSWE, Gray Swan and every internal and customer result above.
• Third-party: the Vals Index ranking (Vals AI, October 13).
• Reported: UK AI Security Institute access (AFP, October 12, citing Narayan).
• Uncertain: when, or whether, free Gemini app users get Argon; whether Google AI Plus ($4.99) includes it, since Google's plan pages show an "Includes Gemini 4 Argon" line at the end of the Plus card while the launch post and release notes name only Pro and Ultra; numeric Gemini app usage limits for Pro and Ultra; Argon's per-model API rate limits, which Google shows only inside AI Studio; and independent replication of Google's coding and security scores.