Gemini 4 Argon launched at half price until 30 November. Use the window to test, not to commit

In brief

Google released Gemini 4 Argon, its new flagship model, on 13 October 2026. It is available worldwide through the Gemini API, Gemini for business, the Gemini app and AI Mode in Search for Pro and Ultra subscribers, and the Antigravity coding tool for Ultra subscribers and enterprises.

The launch price is $2 per million input tokens and $10 per million output tokens on the Standard paid tier (Argon isn't on the free tier, and thinking tokens are billed as output). Caching is implicit only; explicit context caching isn't supported. Cached input tokens are 95% cheaper. This introductory rate runs through 30 November 2026; from 1 December 2026 the price is $4 and $20. Reuters said the release came after months of delays and a leadership shakeup at Google's DeepMind lab.

What is confirmed, and what is still a claim

• Confirmed by Google and Reuters. Global availability, both prices, and an expansion to hundreds more partners of the Fairwind Program, which gives vetted cyber defenders a version of Argon without its cyber guardrails. Google hasn't disclosed Fairwind pricing. Google is offering 10,000 scientists six months of free access to Antigravity's highest usage tier; Google says they must apply and are accepted subject to eligibility, capacity and selection.
• Vendor claims. Google's benchmark figures, such as a first-place Vals Index ranking, are unverified. Treat them as hypotheses to test.
• A design change to plan for. Google's documentation confirms that input, thinking and output share one 1,048,576-token window. A single request returns up to 65,536 output tokens (32,768 for direct non-streaming REST calls). Longer outputs need "long decode", which repeats the request with a continuation token. Current Google GenAI SDK versions (Python 2.30.1 or later, JavaScript 2.29.1 or later) do this automatically; REST callers must write their own loop. Either way, max_output_tokens must be set above 65,536 for long decode to apply.

Decisions to make now

1. Cost every pilot at the December rate. Anything piloted through 30 November costs half what it will from December, and as we noted when frontier AI prices fell in September, volume and agentic loops drive most of the bill. Model $4 and $20 per million tokens. Because caching is implicit only, state your cache-hit assumption rather than relying on the 95% discount. In long-decode requests, Google says the cache expires after 15 to 20 minutes without use, so a long pause between continuation hops is charged at the full input rate.
2. Test on your own work. Compare Argon with your current model on real tasks, by quality, latency and cost per completed task.
3. Check long-context jobs against the shared limit. Measure what your largest prompts leave for output, and set max_output_tokens deliberately.
4. Ask security suppliers whether they use Fairwind. A tier without cyber guardrails is a supplier-risk question. If they plan to run it on your estate, agree in writing the scope, logging and who sees the findings.

Sources and scope

Availability (including Antigravity for enterprises), pricing, the 30 November date, the Fairwind expansion, the scientist programme's application and eligibility terms and the benchmark figures are from Google's announcement, Gemini 4 Argon: Built to solve your hardest problems, by Koray Kavukcuoglu, published 13 October 2026. Availability, pricing, the Fairwind expansion, the scientist programme, and the delays and DeepMind leadership shakeup were cross-checked against Reuters' launch report, carried by Yahoo Tech, published 13 October 2026. The token limits, long decode and the 15-to-20-minute cache expiry for long-decode requests are from Google's Gemini 4 Argon model page and its long decode guide. The paid-tier pricing, the 1 December 2026 price change, free-tier exclusion, billing of thinking tokens and implicit-only caching are from the Gemini Developer API pricing page. The announcement and Reuters report were first checked on 14 October 2026, and all sources were checked on 27 October 2026. The Reuters text refers to a release "on Wednesday" although it is dated Tuesday 13 October; Google's dated announcement is used for the release date. Benchmark figures are Google's claims and have not been verified by ByteNib. The recommendations are ByteNib editorial analysis. This article was researched and drafted with AI assistance, including Anthropic's Claude, and edited under ByteNib's editorial standards.

Update, 27 October 2026: added Google's cache-expiry figure for long-decode requests.

Drafted with AI assistance and reviewed by the editor before publication.

Continue exploring: AI & Machine Learning analysis, the related implementation tutorial, and the structured learning path.