Google opens Gemini 4 Argon to developers with introductory API pricing
Google opened Gemini 4 Argon to developers on October 13, bringing the model to the paid Gemini API and expanding access across its coding tools and consumer products. The rollout gives software teams a route to use Argon beyond the limited testing and cybersecurity access that accompanied its September announcement.
Google first announced Argon on September 30, when access was rolling out to selected defenders in its Fairwind Program. The company said broader availability would follow further testing. Today’s announcement marks that wider release, while access to the version without cyber guardrails still runs through Fairwind.
Where developers can use Argon
Paid API customers in supported regions can build with Argon in Google AI Studio and Android Studio. Google AI Ultra subscribers can use it in AI Studio and Antigravity. Consumer access follows a separate route, with Google AI Pro and Ultra subscribers gaining Argon in the Gemini app and AI Mode in Search.
For Android teams, Google’s integration instructions distinguish the regular API route from Antigravity inside Android Studio Canary. The latter requires an Ultra subscription. The Android CLI plugin connects agents to local build tools, and Google describes a feedback loop that sends Gradle and Kotlin errors back to the model so it can revise unsuccessful changes.
Introductory pricing ends in November
Standard Gemini API pricing is $2 per million input tokens and $10 per million output tokens through November 30, 2026. Thinking tokens count toward output charges. On December 1, those rates rise to $4 and $20 respectively, according to Google’s pricing page. Argon has no free API tier.
Cached input is priced at $0.10 per million tokens during the introductory period and $0.20 afterward. Argon supports automatic implicit caching, but developers cannot create explicit caches for it. Batch and Flex requests have separate, lower rates, so the standard figures are not a universal price for every execution mode.
Longer reasoning needs a different request setup
The model reference lists a shared capacity of 1,048,576 tokens for input, thinking and output. Argon accepts text, images, audio, video and PDFs, and generates text. Its default synchronous output limit is 65,536 tokens. The larger ceiling requires long decode, and any input already occupying the window reduces the space left for generation.
Google’s long decode guide explains that the shared window also includes thinking. On the Interactions API, developers must enable background execution and request more than 65,536 output tokens to continue beyond the default limit. Google then generates successive portions and returns them under one interaction. This path requires server storage and can take several minutes.
The billing implications matter for evaluation and production budgets. Each continuation processes the original prompt and the material generated so far again. Repeated input receives a caching discount, but every portion remains billable, including portions spent entirely on thinking. Work completed before a request fails or is cancelled is also charged.
The release still needs practical testing
Google’s model card warns that Argon can produce incorrect information and may respond slowly or time out. It lists a March 2026 knowledge cutoff, with some subjects less current. Those limits make workload testing and review of generated changes important even when a task fits inside the expanded context window.
Illustrative coding workspace photographed by Virendra Kumar Yadav. CC0 via the WordPress Photo Directory.
