Google opened Gemini 4 Argon to every paid Gemini API customer on Tuesday 13 October, two weeks after announcing it for "trusted cyber defenders" only. The model ID is gemini-4-argon, the context window is 1,048,576 tokens shared between input, thinking and output, and the introductory price is $2 per million input tokens and $10 per million output. That is the same as GPT-6.1 Sol, half of Claude Opus 5.5, and a fifth of GPT-6 Astra and Claude Fable 5.1.
Artificial Analysis scored Argon at its high thinking level at 53 on its Intelligence Index, the same as GPT-6 Astra, and measured $1.99 per index task against Astra's $3.26. The distance between a fifth and 60% is appetite: Argon averaged 62,000 output tokens per task, Astra 27,000. On 1 December the price doubles to $4 and $20, and the same measurement becomes $3.98 per task, about 1.2 times Astra for the same score.
The migration guide carries three changes that will bite before the price does. max_output_tokens now stops the response "once thinking and output together reach it", and Google says to raise any existing limit to at least 32,768. temperature, top_p and top_k are accepted and ignored. Explicit caching is gone, so every cached_content reference has to come out of your requests.
What Google shipped
The first announcement came on Wednesday 30 September, under the byline of Koray Kavukcuoglu, SVP of Google DeepMind. Argon was "rolling out to a set of trusted cyber defenders" through the Fairwind Program, because "safely releasing frontier capabilities at this level requires a phased approach." The same post already gave both prices, $2 and $10 "at an introductory price" and $4 and $20 after, with no date for the switch. The Hacker News thread that day reached 1,705 points. The 13 October post supplied the date, 30 November, and a second thread reached 1,217 points and 858 comments.
The pricing page is less coy than either post. Input is "$2.00 through November 30, 2026" and "$4.00 starting December 1, 2026." Output, which includes thinking tokens, goes from $10 to $20. Cached input goes from $0.10 to $0.20. Batch and Flex are half of standard both before and after; Priority is $3.60 and $18 now, $7.20 and $36 from December. A footnote adds that "Gemini 4 Argon is not available on the Free tier or on the Google AI Pro plan."
The model card, published the same day, calls Argon a sparse mixture-of-experts model that is "not a modification or a fine-tune of a prior model," with a knowledge cutoff of March 2026 "but in some domains the model's knowledge may be less up to date." Google says it "cannot rule out" the model reaching its critical capability levels for CBRN and cybersecurity and that it applied safeguards in response. The public version ships with cyber guardrails; Fairwind partners can apply for one without.
Consumers got a smaller release than the headline. Google AI Pro subscribers have Argon in the Gemini app and in AI Mode in Search. Google Antigravity, the agent IDE, needs the Ultra plan, and Pro users who tried it on launch day saw "Unavailable · Gemini 4 Argon is currently available to Ultra plan subscribers." One Ultra subscriber in the thread saw the same message.
Per token and per task
Artificial Analysis published its numbers on 30 September, the day of the restricted announcement. "Gemini 4 Argon (high) scores 53 on the Artificial Analysis Intelligence Index, matching GPT-6 Astra (max, 53)" and "1 point ahead of GPT-6.1 Sol (max, 52)." It "costs $1.99 per Intelligence Index task, 60% of GPT-6 Astra (max)" and "2.7x GPT-6.1 Sol (max)." Then the sentence that matters: "This cost efficiency is driven by lower token prices, rather than reduced token use," with Argon "averaging 62k output tokens per task, compared with 27k for GPT-6 Astra (max)."
They also did the December arithmetic. "The 50% discount is an initial promotion. After the discount ends, this will rise to $3.98 (~1.2x GPT-6 Astra (max))." Sol charges the same per token as Argon does today and finished the index at a third of the cost per task. From December that ratio is more than five to one, for a one-point difference in score.
Simon Willison ran his pelican SVG prompt at all three thinking levels on launch day: 10.185 cents at low, 11.098 at medium, 17.166 at high. His explanation for the narrow spread: "it used more output tokens than reasoning tokens for all three pelicans, even for high." Lowering the thinking level saves money only on tasks where thinking is the expensive part.
A developer in the Hacker News thread ran 11 small bug fixes from their own projects through Argon, Sonnet 5.5 and GPT-6.1 Sol. Argon solved 10. A median run cost $0.52 against $0.18 for Sonnet and $0.09 for Sol, and took about four minutes against 32 seconds and a minute and a half. Argon produced about 24,000 output tokens per run against Sonnet's 3,000. The eleventh it lost to a ten-minute timeout, in the developer's words by "thinking its way into" it. Google's own guide says "Hard prompts can take several minutes" and tells you to set HTTP timeouts to at least 15 minutes or use background execution.
The migration checklist
The output limit is the change most likely to produce a quiet failure. The guide: "max_output_tokens is a hard limit: the response stops once thinking and output together reach it," and "Gemini 4 Argon can spend more of the limit on thinking than earlier models." The instruction is "Raise max_output_tokens to at least 32,768." The thinking page describes what happens if you do not: when the model hits the limit while reasoning it stops with status incomplete and "returns truncated or empty output (while still billing for any thinking tokens generated)." An app that set 8,192 for Gemini 3 and never reads the finish status can get empty answers and a full bill.
The default is 65,536 for a synchronous call, and "Setting max_output_tokens > 65536 with background=False returns a 400 error." Past that you are in long decode, which runs as a background interaction in 64K hops up to a million tokens minus your input. "You pay for every token in every hop, including hops that contain only thinking." Past the first 64K, "Long streams can drop, so your client must reconnect." The default thinking level is medium, where Gemini 3.1 Pro defaulted to high, and there is no off setting. The guide's advice is to "Run your evaluations at all three levels, then pick the lowest level that meets your quality bar."
The sampling change is silent. "Remove temperature, top_p, and top_k from your requests. If provided, they are ignored." No error, no warning in the response. A classification pipeline that set temperature to 0 for repeatability will keep running, at whatever the model's default sampling is, and the only way to find out is to measure repeatability again.
The caching change is loud by comparison, because the old calls stop working. "Explicit caching (cachedContents / caches.create) is not supported on Gemini 4 Argon." You are told to "Remove calls that create or reference cached_content from your requests." What remains is implicit caching, "enabled by default for prompts of at least 4,096 tokens," at a 95% discount on hits. Opus 5.5 charges $5 per million to write a cache and $0.20 to read one; Argon charges nothing to write and $0.10 on a hit. The difference is that you cannot pin a prefix. A miss on a 100,000-token system prompt bills at $2 today and $4 from December, and the API does not promise when it will miss.
What to do with this
Price the work at $4 and $20. The introductory rate lasts 48 days from launch, and anything you build this quarter runs mostly at the standard one. If your case only closes at $2, it closes on 30 November.
Count output tokens per task before you compare per-token prices. Run your own eval at low, medium and high, log total_thought_tokens and the output count, and multiply. Artificial Analysis found a model at a fifth of Astra's price costing 60% of Astra per task; the pelicans show that output, not thinking, can be the bulk of the bill, so the thinking level is a smaller lever than it looks.
Set max_output_tokens to 32,768 or higher, read the status on every response, and treat incomplete as a failure you are paying for. Move anything that needs more than 65,536 tokens to background execution and give your client 15 minutes.
Grep for temperature, top_p, top_k and cached_content before you change the model ID. The first three will not error, which is the problem. If a pipeline depended on temperature 0, test it for drift on day one, because the knob you used to hold it still has been removed.