Gemini 4 Argon is out; its API price doubles on 1 December

Google has released Gemini 4 Argon, its new frontier model, to developers on the paid Gemini API and to paying subscribers in the Gemini app [1]. Through the API it costs $2 per million input tokens and $10 per million output tokens until November 30, 2026, then $4 and $20 from December 1 [2]. In the app it is limited to the AI Pro and AI Ultra plans, and in the Antigravity coding tool to AI Ultra [1]. The free tier does not get it [2].

If you want to try a new frontier model cheaply, the next seven weeks are the window. If you plan to build on it, budget at the December price.

What was announced

Google’s launch post, by Koray Kavukcuoglu of Google DeepMind, says Gemini 4 Argon is globally available today [1]. It lists five routes in. Developers get it through the Gemini API in AI Studio and the Gemini platform [1]. Enterprises get it in Gemini for business, the API and Google Antigravity [1]. Everyday users get it in the Gemini app and AI Mode in Search, but only on Google AI Ultra or Pro [1]. Antigravity gets it for AI Ultra subscribers only [1]. Scientists can apply for Antigravity’s highest usage tier free of charge for 6 months; Google says it is offering this to 10,000 scientists, subject to eligibility, capacity and selection [1]. Cybersecurity defenders can apply to the Fairwind Program for access to Argon without cyber guardrails [1].

According to [3], AI Pro costs $20 a month and AI Ultra starts at $99 a month. Ars also reports that Pro subscribers can generate code with Argon in the Gemini app’s Canvas mode, and that Google is keeping Gemini 3.1 Pro available for now; Google declined to say when, or if, Argon will come to Antigravity for other tiers [3]. According to [4], it is not clear when, or if, Argon will reach Plus subscribers or free users.

The API price is the clearest number in the launch. Google’s pricing page lists gemini-4-argon with an input price of $2.00 per million tokens through November 30, 2026, and $4.00 starting December 1, 2026 [2]. Output, including thinking tokens, is $10.00 through November 30 and $20.00 from December 1 [2]. Context caching costs $0.10 per million tokens, then $0.20, and only implicit caching is supported; explicit caching is not [2]. The same page says Argon is not available on the free tier [2]. The launch post describes the cached-input price as 95% off, and says the higher price applies after the introductory period expires [1]. Google says Argon has a 1M unified input-output token limit through the Gemini API, and points developers to documentation on a new continuation protocol [1].

On performance, the claims are Google’s. The company says Argon ranks first on the Vals Index, which covers finance, coding, legal and tax work [1]. It says Argon scores 87% on the Embodied Reasoning (ERQA) benchmark, ahead of every other frontier model [1]. On private evaluations by Gamow Labs, it reports 57% on RareBench and 52.2% on LabBench [1]. According to [4], Google says Argon outperforms GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5 across a range of benchmarks, including coding, cybersecurity, science, finance and law. Google also says Argon answers with less hedging, less repetition and fewer clichés [1].

Google’s own use cases are specific. It says Argon in Antigravity helped teams increase agentic code submissions by more than 35% while reducing code rollbacks by nearly a third [1]. It says Argon found optimisations in its Agent Development Kit for Python that cut in-memory session prep time by 40% in long sessions and sped up cold imports by 21% [1]. According to [3], Argon found redundant emoji rendering in YouTube’s comment system whose removal cut server load by 45 percent, and found bugs in Wing’s drone physics simulator that had been in the code since 2018.

Safety gets its own section. Google says external bio red teaming could unlock less than 1% of the bio sub-tasks it tests for [1]. It says it trained Argon to respect system boundaries, report negative findings honestly and pause for human confirmation before irreversible steps, and that shortcut attempts on coding benchmarks dropped to 2% [1]. It also runs an automated misalignment monitor that reads the agent’s reasoning and actions to block unintended workarounds in real time [1]. On Gray Swan’s indirect prompt injection benchmark, Google says Argon cut attack success to 0.7% across 15 adaptive attempts [1]. According to [3], Google first revealed Argon a few weeks ago and said it needed to move cautiously toward a full release; according to [4], the public rollout follows a limited release to “trusted cyber defenders”.

What changed, and for whom

For API users, another frontier-class model is now on the price list, and for seven weeks it is the cheapest of the ones the reports compare. According to [4], the introductory price is around a fifth of the cost of OpenAI’s GPT-6 Astra and Anthropic’s Fable 5 models, and Argon would still be significantly cheaper per token after the price doubles. The Verge adds the caveat that matters: models can use different numbers of tokens for the same task, so a lower price per token does not always mean a lower bill [4]. The output price includes thinking tokens [2], so a model that thinks longer costs more on the same task.

The doubling date is the practical trap. A workflow that is cheap in October can cost twice as much in December without any change on your side [2]. If you test Argon now, record the token counts, not just the bill, and multiply by the December prices before you decide.

For Gemini app subscribers, the plan you are on decides what you get. AI Pro gets Argon in the app and in AI Mode in Search, but not in Antigravity [1][3]. AI Ultra gets it everywhere [1]. Free users do not get it; this follows Google’s recent change that left free Gemini app users with Flash-Lite only, according to [3]. If you are on AI Plus, neither Google nor the reports say when Argon will reach you [4].

For teams that run agents, the safety claims are worth reading but not taking on trust. The 0.7% prompt injection figure is from one benchmark under Google’s test conditions [1]. Pausing before irreversible steps and an automated monitor are useful designs, and they are also controls you can and should keep in your own setup, whatever the model.

For robotics and science teams, Google says Argon built simulation code that was transferred to a humanoid robot in an internal test, and that it is available to build with through the paid Gemini API and, for AI Ultra subscribers, in AI Studio and Antigravity [1].

Who it is for — and not

Argon is for developers and teams already paying for the Gemini API, and for AI Pro and Ultra subscribers who want Google’s strongest model in the app. It is also for anyone who wants to test a new frontier model cheaply before December 1. It is not for free users, who cannot use it, and not yet for Antigravity users below AI Ultra [1][2][3]. If your work already runs well on a cheaper model, an introductory price is not a reason to switch.

This page will be updated if Google changes Argon’s prices or plan availability, or extends it to AI Plus, free users or Antigravity on AI Pro.

∴ Argon is cheap until November 30; budget at $4 and $20 per million tokens, and test Google’s claims yourself.