On October 13, Google moved Gemini 4 Argon to general availability, thirteen days after a defenders-first release. The first model to carry the Gemini 4 name costs $2 per million input tokens and $10 per million output tokens through November 30, then $4 and $20, and it can return a single answer of nearly a million tokens. The price doubles in seven weeks. The answer length is what engineers will still be planning around after it does.
Thirteen days from defenders to everyone who pays
Google announced Argon on September 30 as a staged release: first to trusted defenders in its Fairwind Program, to internal teams and to Wiz's Scan for Good effort, with paid API customers and Google AI Ultra subscribers promised next. Koray Kavukcuoglu, Google's chief AI architect, wrote that shipping capability at this level "requires a phased approach." Defenders and Google's own teams received a version without cyber guardrails; the public version refuses harmful cyber and CBRN requests. The model card says Google "cannot rule out" that Argon reaches its critical capability level for cyber.
The October 13 phase opened the API's paid tier in AI Studio and Android Studio, brought Argon to the Gemini app for Google AI Pro and Ultra subscribers, and widened Fairwind to hundreds more partners. The API has no free tier. The notable exception is an application program offering 10,000 scientists six months of free access to Antigravity's highest usage tier.
Defenders-first is no longer a Google habit; it is the template. Anthropic held Claude Mythos Preview to 11 organizations under Project Glasswing in April, and OpenAI followed within about a week with GPT-5.4-Cyber for verified defenders. Anthropic's Opus 5.5 system card promises verified security users the same model with fewer restrictions. The people who patch now get the model before the people who probe.
A discount with an expiry date
The pricing page states its own expiry: $2 input, $10 output and $0.10 cached input through November 30, then $4, $20 and $0.20 from December 1. Batch and Flex run at half those rates; Priority costs 1.8 times standard. When Argon first appeared, neither Artificial Analysis nor The Batch could say when the discount would end. The date arrived with general availability.
Output pricing includes thinking, and Argon thinks at length. Google says the model "works longer and harder" on complex problems, and the token counts agree: Artificial Analysis recorded 116,000 output tokens per index task, against 68,000 for GPT-6.1 Sol. So will the invoice.
Against the incumbents, the price story is tidy. The introductory rate matches GPT-6.1 Sol at $2 and $10. The December rate matches Claude Opus 5.5 at $4 and $20. GPT-6 Astra sits at $10 and $50. On December 1, Argon moves from Sol's price to Opus's.
One window, one very long answer
Input, thinking and output share a single 1,048,576-token window, and max_output_tokens is a hard cap on thinking and output together. The default is 65,536; Google suggests 131,072 or 262,144 for complex work. Beyond 65,536 you are in long decode: the request runs in the background with store: true, the model writes in hops of up to 64K tokens, and the answer is stitched together from the steps. GPT-6 Astra and GPT-6.1 Sol stop at 128,000 output tokens; Opus 5.5 stops at 128K by DataCamp's count.
That moves the unit of budgeting from the request to the job. A 900,000-token answer costs about $9 in output at the introductory rate and $18 from December 1. Google's long-decode docs say the first 64K tokens take about ten minutes, so a near-million-token answer is measured in hours; background runs may last 24 hours before they are marked failed, and failed or cancelled runs are billed for the work already done. Each hop re-reads its context at a 95% cached discount, which keeps input cheap but inflates total_input_tokens; the docs warn against using that counter to trigger compaction.
Two lines belong in any runbook. The model does not follow token-length or word-count targets, so the cap is the only budget control. And Google steers long decode away from chat, summarization, fixed-length writing and agents making many short tool calls, toward proofs, research and the hardest coding tasks. It is a tool for the overnight job.
Where Argon leads, and where the board has already moved
At launch the third-party numbers were strong. Argon ranked first of 43 on the Vals Index at 68.9%, first on Arena's text leaderboard at 1,525 Elo by The Batch's account (ahead of Claude Opus 4.6 at 1,505 and Opus 5.5 at 1,504), and 53 on Artificial Analysis's v4.3 index, tied with GPT-6 Astra. Its 15% hallucination rate on AA-Omniscience was the lowest of any model scoring 45 or more; Astra's was 51%.
Agentic coding is the soft spot, and Google's own model card shows it: 57.4% on Terminal-Bench 4.0 against 66.4% for Opus 5.5, a figure Anthropic reports at xhigh effort. The Batch's verdict was that Argon "still trails Anthropic's newest models on agentic coding." Independent runs narrow the gap: Artificial Analysis measures 57% for Argon and 60% for Opus 5.5. Even vendor tables disagree at the margin; Google lists GPT-6 Astra at 58.2% on Terminal-Bench 4.0, while Anthropic, citing OpenAI, lists 57.9%.
Leaderboards also move. On Vals Index v2.2, updated October 27, Argon scores 69.24% and ranks fourth, behind Claude Sonnet 5.5, Opus 5.5 and a Gemini 4 Carbon checkpoint reportedly in partner testing. Artificial Analysis's v5 index, released the same day with harder agentic tests, puts Argon at 50, behind a top five of Anthropic and OpenAI models led by Opus 5.5 at 58, with GPT-6.1 Sol at 55 and GPT-6 Astra at 57. At the December price, Argon's cost per index task is $7.68, against $10.04 for Opus 5.5 and $1.27 for Sol.
The 30,000-foot read
Strip away the choreography and Argon is three bets. That defenders-first staging is now the cost of shipping frontier capability at all, which Google has accepted as fully as its rivals. That a model which thinks long can be sold on how long it answers, a dimension where its main rivals stop at 128,000 tokens. And, more quietly, that seven weeks priced like GPT-6.1 Sol will earn it the right to be paid like Opus 5.5, while Artificial Analysis's new index ranks it below both and Vals places it between them.
None of that is in the engineer's control; the cap, the queue and the monitoring are. Treat a million-token answer the way you treat a batch pipeline: bounded, observed and costed before it starts. Teams that skip that step will learn the price of an abandoned run from an invoice. Watch December 1, and watch Carbon. The leaderboard is not waiting for the discount to expire.
Sources: Gemini API changelog (October 13 general availability); Gemini 4 Argon: our next era of frontier intelligence (September 30 announcement, Fairwind, Wiz and internal rollout, guardrail-free defender version, Kavukcuoglu quote); Gemini 4 Argon: Built to solve your hardest problems (general availability surfaces, Fairwind expansion, scientists program, November 30 end of introductory pricing, "works longer and harder"); Gemini 4 Argon is now available for developers (paid tier in AI Studio and Android Studio); Gemini 4 Argon is now available in the Gemini app (Google AI Pro and Ultra); Gemini API pricing (standard, cached, Batch, Flex and Priority rates before and after December 1, no free tier); What's new in Gemini 4 Argon (shared 1,048,576-token window, output cap defaults, thinking billed as output, no free tier); Gemini API long decode guide (64K hops, background execution, ten-minute first hop, 24-hour limit, billing of failed runs, cached re-reads, length targets, recommended uses); Gemini 4 Argon model card (Google's Terminal-Bench 4.0 and Vals table, cyber capability statement); The Batch, issue 374 (Vals v2.1 rank, Arena Elo, Artificial Analysis v4.3 score, public-version refusals, undisclosed discount end, "still trails" quote); Artificial Analysis launch analysis (AA-Omniscience hallucination rates, discount described as open-ended at preview); Artificial Analysis Intelligence Index v5 (v5 scores and ranking); Artificial Analysis: Argon vs Opus 5.5 (independent Terminal-Bench 4.0, cost per task); Artificial Analysis: Argon vs GPT-6.1 Sol (output tokens and cost per task); Artificial Analysis: Argon vs GPT-6 Astra (Astra's v5 score); Vals Index (v2.2 standings, October 27); Introducing Claude Opus 5.5 (66.4% at xhigh, GPT-6 Astra at 57.9% as reported by OpenAI); Claude API pricing (Opus 5.5 rates); GPT-6.1 Sol model page and GPT-6 Astra model page (prices, 128,000-token output limits); Claude Opus 5.5 system card (verified cyber users); The Next Web (Glasswing and GPT-5.4-Cyber, secondary); DataCamp (Opus 5.5 output limit, secondary); TestingCatalog (Gemini 4 Carbon testing, secondary). CAVEATS: Terminal-Bench 4.0 figures from Google's model card and Anthropic's page are vendor-reported; Vals, Artificial Analysis and Arena are third-party. The Arena Elo figures are single-sourced to The Batch because Arena's leaderboard did not load for us; Arena was formerly LM Arena. Artificial Analysis scores from v4.3 (53 at launch) and v5.0 (50, October 27) are not comparable across versions, and its cost per task uses the $4/$20 December price. Leaderboard standings after October 13 are reported in hindsight. Opus 5.5's 128K output limit is from DataCamp, since Anthropic's model page does not state it. The $9 and $18 figures are our arithmetic from Google's rates. Gemini 4 Carbon is reported by TestingCatalog and listed by Vals; we found no Google announcement of it. Google's separate gemini-4-availability post returned a server error and is not used. CORRECTIONS to our leads: Google did not label the September 30 release a preview, and access also went to internal teams and Wiz; The Batch carries no Terminal-Bench figure for Argon (the 57.4% vs 66.4% comparison comes from Google's model card and Anthropic's page); cached input also doubles, to $0.20, on December 1; "paid tier only" holds for the API and consumer app, with a free application program for 10,000 scientists as the exception.