Gemini 4 Argon general availability
B · Useful, limited Google's launch post for Gemini 4 Argon with prices, availability and named internal and partner results; useful facts, but the performance claims are vendor-reported showcases.
Summary based on the full text·Original: about 12 minutes to read
In short
Koray Kavukcuoglu's post (13 October 2026) opens Gemini 4 Argon to developers, enterprises and paying consumers, after a limited preview for cyber defenders. Google prices it at $2 and $10 per million input and output tokens until 30 November, then $4 and $20. Most of the post is a list of results from Google teams and partners: code refactoring, performance tuning, security finds and science projects. Google claims first place on the Vals Index of economically relevant tasks.
Key numbers
- $2 / $10introductory price per million input / output tokens; cached input 95% off
- $4 / $20price per million input / output tokens after the introductory period ends on 30 November 2026
- more than 35%rise in agentic code submissions at Google with Argon in Antigravity, with rollbacks down nearly a third
- 2.85xend-to-end inference speedup Argon found for Gemma 4 E2B on an A100 via vLLM
- 87%Argon's score on the Embodied Reasoning (ERQA) benchmark
As reported by the source.
English·Web page·opens blog.google
What it reports
- Available via API in AI Studio and the Gemini platform, in Antigravity for Ultra subscribers, and in the Gemini app and AI Mode for Ultra and Pro subscribers. 1M unified input-output token limit.
- Internal engineering: agentic code submissions up more than 35% with rollbacks down nearly a third; Chrome separated a ~6 million line legacy folder within a 40 million line codebase.
- Google's ADK: in-memory session prep time down 40% in long sessions, cold imports 21% faster.
- Open-source xTB quantum chemistry: up to 2.3x faster structure prediction, 1.8x faster conformer search.
- Gemma 4 E2B inference: 2.85x faster on an A100 (about 37,900 to over 108,000 tokens per second) with token parity.
- Partner Gamow Labs reports 57% on RareBench and 52.2% on LabBench (private evals); ERQA 87%.
- 10,000 scientists offered the top Antigravity tier free for 6 months; the Fairwind cyber-defender programme expands to hundreds of partners.
- Safety summary: shortcut attempts on coding benchmarks at 2%; prompt-injection attack success 0.7%.
How they got it
A product launch post. Results come from Google's own teams and early-access partners, with links to some pull requests and repositories. No evaluation method is given for the benchmark claims.
Limits
- Productivity figures (submissions, rollbacks) are internal, with no baseline period or definition.
- Several results are single showcases selected by the vendor; partner views are flagged as their own.
- The introductory price ends after seven weeks, so cost comparisons should use the $4/$20 rate.
Why it matters for companies
The price doubles on 1 December, so any business case built in the trial window needs the later rate. The verifiable items, merged pull requests in public repositories, are a better guide than benchmark rankings. The coding productivity claim is worth testing in-house with your own rollback and review data rather than taking at face value.
Summary based on the full text. Read the original for the full argument and data.Read the original on blog.google
- Tags
- Model releases, Pricing