Google Launches Gemini 4 Argon API With a 1M Context Budget and $2/$10 Intro Pricing

Gemini 4 Argon is now broadly available

Google changed Gemini 4 Argon from a narrow Fairwind early-access model into a broadly available product on October 13, 2026. Paid Gemini API customers can call it in Google AI Studio under the model ID gemini-4-argon. Google also names Android Studio and its Gemini platform for developers and enterprises, Antigravity for AI Ultra subscribers, and the Gemini app and AI Mode in Search for eligible subscribers.

The eligibility details differ by surface. The developer notice describes the paid API tier; the consumer notice says AI Pro and Ultra subscribers can use Argon and tasks in the Gemini app; Antigravity access is described for Ultra. “Globally available” therefore does not mean every free account receives every route. Confirm that the exact endpoint or product appears in the intended account before committing a migration.

The API ID is gemini-4-argon and the introductory price expires

Google’s developer guide identifies gemini-4-argon as the API model ID. It lists $2 per million input tokens and $10 per million output tokens through November 30, 2026, with cached input at a 95% discount. That makes the stated cached-input rate $0.10 per million tokens during the introductory period.

The same table says standard pricing begins December 1 at $4 per million input tokens and $20 per million output tokens. A pilot priced only at the launch rate can understate steady-state spend by half. Budget with the December rates, include reasoning and retry tokens, and measure cost per accepted task rather than cost per single call.

The 1M figure is a shared input-and-output budget

Google documents a unified 1-million-token context window for input and output tokens. That wording corrects an important ambiguity in the September announcement, which emphasized a larger output limit. The API guide does not describe separate one-million-token allowances for both sides of a request.

For long codebases, document collections or agent traces, teams should explicitly reserve output capacity and observe truncation, latency and cost as prompts grow. A large nominal budget does not establish reliable recall across every position, successful million-token generations or acceptable economics. Those are workload-specific questions that require measured trials.

Thinking controls trade depth for latency and cost

Argon exposes low, medium and high values through thinking_level, with medium as the default. Google recommends high for multi-file debugging, vulnerability remediation, formal math and architecture work; medium for interactive coding and agentic tool use; and low for routing, classification and targeted extraction.

Those descriptions are starting points, not an automatic quality ladder. Evaluate every level against the same cases and stop conditions. Record accepted outcomes, severe failures, end-to-end latency, input and output tokens, retries and reviewer time. If low or medium already clears the acceptance threshold, higher reasoning can add cost without changing the decision.

Google reports production and benchmark gains

Google says Argon ranks first on the Vals Index and reports results across coding, legal, finance, tax, long-video and cyber evaluations. It also says internal use of Argon in Antigravity increased agentic code submissions by more than 35% while reducing code rollbacks by nearly one third, and describes a Chrome refactor across hundreds of files. These are Google-reported results and case studies; AccessAllGPT did not inspect the underlying traces or reproduce them.

The launch evidence is useful for deciding what to test, not for assuming transfer to another repository or workflow. A fair comparison should hold tools, permissions, time budgets, test suites and acceptance checks constant. Score completed, accepted work and damaging failures—not only benchmark rank or generated code volume.

Higher cyber capability raises the control requirement

Google says Argon solved all 12 of its simulated end-to-end cyberattack scenarios, compared with five for the previous generation, and showed stronger ability to troubleshoot complex biology procedures. Google says it responded with input screening, activation probes that inspect internal signals and continuous monitoring that can stop a response mid-stream. It also reports fewer harmful-response violations, less than 1% unlock of tested biology subtasks in external jailbreak work and no covert-action behavior in its described honeypot tests.

Those results and controls are vendor-reported. They do not remove the application owner’s responsibility to constrain tool authority, isolate execution, require approval for consequential writes, retain independent logs and test rollback. Teams handling code execution, security data or scientific workflows should review the linked full safety report and their contract-specific controls before enabling autonomous actions.

What builders and buyers should do next

API teams can begin a bounded evaluation now. Use gemini-4-argon in a non-production project, freeze representative cases, cap tool permissions and spend, and compare low, medium and high thinking against the current model. Price the decision using the December 1 rates as well as the launch period, and preserve enough of the unified context budget for the expected response.

Consumer and enterprise buyers should distinguish app, tasks and Antigravity eligibility. Verify the plan, region, connected-app permissions, data controls and administrator policy on the actual account. Deploy only after the trial clears explicit quality, failure, latency, cost and authority thresholds; otherwise retain the existing model and recheck when Google publishes the missing account-specific operational details.

Disclosures

AccessAllGPT did not receive Google access, credits, a subscription, briefing, demo, review or compensation for this article. Google did not sponsor, review or endorse it. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with Google, Google DeepMind or organizations cited. Publication-wide relationships are listed on the disclosures page.