Gemini 4 Argon is generally available: $2/$10 until December, and the 1M-token limit is shared

Thirteen days after announcing it for a closed group of cyber defenders, Google made Gemini 4 Argon generally available on October 13. The model is now in the paid tier of the Gemini API through Google AI Studio and the Gemini platform, in Google Antigravity for Google AI Ultra subscribers, and in the Gemini app and AI Mode in Search for Ultra and Pro subscribers, according to Google's launch post. Android developers get it through Android Studio, per Google's developer announcement.

Promptea covered the September 30 announcement when access was limited to Google's Fairwind Program. What changed today is access, the final shape of the token limits, and a firm date for the price increase.

Pricing: a two-phase schedule

Google's pricing page lists Argon at $2 per million input tokens and $10 per million output tokens through November 30, 2026, rising to $4 and $20 from December 1. The output price includes thinking tokens. Cached input costs $0.10 per million during the introductory period and $0.20 afterwards, and the Batch API halves the input rate to $1.00 for now. Argon is not available on the free API tier.

Two details affect cost planning. First, the page says Argon supports implicit caching only: explicit context caching is not supported, so you cannot pin a cache yourself. You get cache discounts only when your requests happen to share a prefix. Second, VentureBeat reports that Argon at the introductory rate undercuts Claude Opus 5.5 ($4/$20) and GPT-6 Astra ($10/$50), but from December 1 its list price matches Opus 5.5. A cost comparison you run in October will not hold in December.

The 1M-token limit, clarified

The September post described an output limit of 1M tokens, up from 64K. Today's documentation is more precise. The model page lists a unified context window of 1,048,576 tokens shared across input, thinking and output, and a default maximum output of 65,536 tokens. Going beyond that requires what Google calls long decode.

Long decode works like this: you create an interaction with background execution turned on and set max_output_tokens above 65,536. The server generates the response in 64K-token hops, each continuing where the last stopped, and you poll or stream the interaction and join the output steps yourself. According to the docs, a synchronous request (streaming or not) that asks for more than 65,536 output tokens returns a 400 error, and long decode for that mode is listed as coming soon.

So an Argon call cannot take 1M tokens in and send 1M tokens out. Whatever you send in reduces what can come back. It also means our earlier headline phrase, "1M-token output budget", only holds for an asynchronous call with an almost empty prompt.

Benchmarks: whose numbers

Most of the headline figures come from Google. It reports 77.9% on DeepSWE v1.1, 51.3% on Zapier's AutomationBench, 87% on the ERQA embodied reasoning benchmark, a 0.7% attack success rate on Gray Swan's indirect prompt injection benchmark, and shortcut attempts on coding tasks down to 2%. None of these has been independently replicated yet.

VentureBeat cites outside leaderboards that give a mixed picture. On the Vals Index, which Google highlights, Argon ranks first at 68.90%, narrowly ahead of Claude Sonnet 5.5 (67.04%) and Claude Opus 5.5 (66.97%). On the Artificial Analysis Intelligence Index, it scores 53, tied with GPT-6 Astra and Claude Fable 5.1 and behind Opus 5.5 (58), Sonnet 5.5 (56) and Meta's Muse Wonder Full (56). VentureBeat also notes Argon trails Opus 5.5 on Terminal-Bench 4.0 (57.4% vs. 66.4%). Promptea could not open the Vals or Artificial Analysis pages from our research environment, so these leaderboard numbers are attributed to VentureBeat's reporting.

Overall, Argon is competitive at the top and leads some aggregate and domain evaluations. Calling it the best model overall would overstate the evidence.

What else is in the release

• Model facts. The model card describes a sparse mixture-of-experts model trained on TPUs. Google says it is a new model, not a fine-tune of an earlier one, and lists its knowledge cutoff as March 2026.
• Features. Function calling, structured outputs, code execution, search and Maps grounding, URL context and file search are supported. Computer use is in preview. Thinking can be set to low, medium or high.
• Cyber access. Google says the Fairwind Program, which gives vetted defenders a version without cyber guardrails, is expanding to hundreds of additional partners.
• Science program. Google is offering 10,000 scientists the highest Antigravity usage tier free for six months, subject to eligibility and capacity.

What it changes for developers

Model selection. Through November 30, Argon's list price is below Claude Opus 5.5 and GPT-6 Astra, per VentureBeat's comparison. If your workload is long-horizon coding or document-heavy analysis, the introductory window is a low-cost time to run your own evaluations. Budget for the doubled price when deciding whether to switch.

Agent frameworks. Google's long-decode docs warn that a low max_output_tokens value set in an agent framework or evaluation harness, such as 8,192, can stop Argon before it finishes thinking. Because thinking tokens count toward output, check the defaults in whatever tooling wraps the model before judging its results.

Long outputs cost real money. A long decode that uses most of the 1M window would cost around $10 at the introductory output rate and about $20 after December 1, before any input charges. Long decode is for the rare task that needs it, and Google's docs say most requests finish within 64K tokens.

Prompt structure. With implicit caching only, the practical way to benefit from the 95% cache discount is to keep long, stable content (system instructions, reference documents, tool definitions) at the start of the prompt and put the variable parts at the end.