Sign in Subscribe
news

Gemini 4 Argon

Sebastien Dubois

Sebastien Dubois

— 8 min read

This is a note from my public notes. View the canonical version: Gemini 4 Argon.

Gemini 4 Argon is Google DeepMind's frontier Gemini model, announced on 30 September 2026. It's the first model of the Gemini 4 generation, and according to Artificial Analysis it's Google's first proprietary model above the Flash class in more than seven months. Google built it for long, multi-step work: real-world software engineering, enterprise knowledge work (legal, finance, tax), and cybersecurity defense.

Why does it matter? For most of 2026, Google competed with cheap and fast Flash models (see Gemini 3.6 Flash) while Anthropic and OpenAI held the top of the leaderboards. Argon is the model that puts Google back in that top group: Artificial Analysis scores it level with GPT-6 Astra on its Intelligence Index.

My launch take, with the community reaction, is in 2026-10-03 Gemini 4 Argon - Google is back at the top, but you can't use it yet.

Key facts

  • Announced by Koray Kavukcuoglu (SVP at Google DeepMind and Chief AI Architect at Google) on the Google blog, and by Demis Hassabis on X
  • Context window: 1M tokens (per Artificial Analysis). The API model page gives the exact figure: 1,048,576 tokens, shared by input, thinking and output
  • Output limit: up to 1M output tokens, up from 64K on previous Gemini models. Google's argument: with that headroom, the model can generate hundreds of thousands of tokens in a single trajectory and solve hard problems in one go. A new Gemini API feature, Long Decode Continuation, pauses long responses and resumes them across follow-up calls, so reasoning can run that long without request timeouts. In the API docs (October 2026), a response stops at 65,536 tokens by default; to go further you turn on background execution and raise max_output_tokens, and Google's servers chain continuation hops of up to 64K tokens each until the model finishes
  • Modalities: text, image, video and speech input; text output (per Artificial Analysis). The API model page also lists PDF input
  • Reasoning: Google ran its own benchmarks at the highest thinking setting. Artificial Analysis tested "high", the highest setting available to them. The API exposes three thinking levels (low, medium, high), and medium is the default
  • Pricing: introductory price of $2 per million input tokens and $10 per million output tokens; cached input is 95% off ($0.10 per million). After the introductory period: $4 / $20. Google hasn't said when the promotion ends (Artificial Analysis says "at least one month"). Update (October 2026): the Gemini API pricing page now dates it. The intro price runs through 30 November 2026, and $4 / $20 applies from 1 December 2026. Batch and Flex cost half the standard price; Priority costs $3.60 / $18 during the intro period. There's no free tier, and caching is implicit only
  • Availability: restricted. First to trusted cyber defenders through Google's Fairwind Program, while Google takes part in the U.S. government's voluntary pre-release model access process. Next in line: paid API customers and Google AI Ultra subscribers, then developers, enterprises and consumers. No date. Update (13 October 2026): generally available on the Gemini API and in Google AI Studio as gemini-4-argon, paid tier only. Google also added it to the Antigravity agent in Gemini Managed Agents

What it's good at

Knowledge work. In Google's comparison table, Argon leads on the Vals Index (68.9% vs 67.0% for Claude Opus 5.5), Vals Finance Agent v2 (65.4% vs 58.9% for Claude Fable 5.1), Zapier's AutomationBench (51.3% vs 42.5% for Opus 5.5), and Harvey's Legal Agent Benchmark. That last one is the outlier: 19.6% vs 6.7% for the next best model, roughly three times higher. Artificial Analysis independently ranks it #1 on its own AutomationBench-AA (78%).

Long context and vision. GraphWalks at 256K to 1M tokens: 84.2% vs 71.8% for GPT-6 Astra. LVBench (long video understanding): 91.7%, state of the art. Chartography (chart analysis): 71.6%.

Coding at scale. DeepSWE v1.1 (long-horizon software engineering): 77.9%, state of the art per Google. Vibe Code Bench: 91.9%. Inside Google, Argon agents migrate C/C++ codebases to Rust, from tens of thousands of lines (re2, libgav1) up to 800K+ lines for the Fuchsia Zircon kernel. For libgav1, they replaced 32K lines of SIMD code with safe Rust the compiler vectorizes automatically; the result runs 2.7x faster than the existing Rust port, with identical output. Google says these rewrites still go through automated and manual audits before production.

Cyber defense. Google trained it to find, validate and patch vulnerabilities on its own. It ties for first on CWE-bench v1 (68%, with GPT-6 Astra) and beats Gemini 3.8 Flash Cyber on Google's internal vulnerability benchmark (20 programming languages) and on Wiz's black-box penetration testing benchmark. Trusted defenders and Google's internal teams get it without cyber guardrails. Through Wiz's Scan for Good program, it found a critical vulnerability exposing personal data in healthcare software used by hospitals worldwide, which previous frontier models had missed.

Knowing what it doesn't know. This is the most interesting number to me. On AA-Omniscience, Argon hallucinates 15% of the time, the lowest of any model scoring 45+ on the Artificial Analysis Intelligence Index (GPT-6 Astra: 51%, GPT-6.1 Sol: 54%). Its accuracy is 50%, lower than Astra's 63%. So it doesn't know more; it says "I don't know" instead of guessing.

Where it trails

Google's own table shows five benchmarks where Argon loses:

Benchmark Gemini 4 Argon Best competitor
FrontierSWE v2 55.0% 65.5% (GPT-6 Astra)
Terminal-Bench 4.0 57.4% 66.4% (Claude Opus 5.5)
PostTrainBench (ML engineering) 45.3% 49.3% (Claude Opus 5.5)
Terminal-Bench Science 0.1 57.6% 68.1% (GPT-6 Astra)
OSWorld-2.0 (computer use) 69.2% 72.6% (GPT-6 Astra)

Artificial Analysis confirms the terminal weakness: 57% on Terminal-Bench 4, behind Claude Sonnet 5.5 (max, 64%), Claude Opus 5.5 (max, 60%) and GPT-6 Astra (59%). It's still a 53-point jump over Gemini 3.1 Pro Preview.

It's also token hungry: 62k output tokens per Artificial Analysis task, vs 27k for GPT-6 Astra (max). Its cost advantage comes from the token price, not from efficiency.

Positioning vs other models

Artificial Analysis Intelligence Index and cost per task (30 September 2026):

Model Intelligence Index Cost per task
Gemini 4 Argon (high) 53 $1.99 at intro price, $3.98 at standard price
GPT-6 Astra (max) 53 $3.26
GPT-6.1 Sol (max) 52 about 2.7x cheaper than Argon
Gemini 3.8 Flash (high) 41
Gemini 3.1 Pro Preview 30
  • vs GPT-6 Astra: same score, 60% of the cost per task at the intro price, about 1.2x the cost at the standard price
  • vs GPT-6.1 Sol: one point higher, 2.7x more expensive per task
  • vs Claude Opus 5.5: in Google's table, ahead on knowledge work, long context, multimodal and science (LABBench 2: 88.8% vs 73.1%), behind on terminal coding and ML engineering
  • vs previous Gemini models: +23 points over Gemini 3.1 Pro Preview and +12 over Gemini 3.8 Flash. Gemini 3.5 Pro had been announced at Google I/O in May 2026 (see Gemini 3.5 Pro), but Artificial Analysis compares against 3.1 Pro Preview as Google's previous non-Flash model

StarCraft bot-writing (October 2026)

In the 15 October 2026 edition of StarSkirmish Bench (Kai McPheeters), each model gets one hour to write a StarCraft: Brood War bot in C++, and that bot then plays a fixed roster of human-written bots. The models write bots; they don't play the game live.

Argon, running in Antigravity CLI, placed 3rd of nine with 39.3, effectively tied with Claude Opus 5.5 (38.6) and Claude Fable 5.1 (37.2), at about $3.63 per run. GPT-6 Astra (57.3) and GPT-6.1 Sol (45.8) are ahead, and Argon is the least consistent flagship (18 points between its best and worst run).

It's the kind of independent long-horizon coding eval my launch take said to wait for. Method, full results and the Hacker News critiques are in the StarSkirmish Bench note.

Safety approach

Google holds the broad release back while it strengthens safeguards in four areas:

  • Misuse: refuse cyber and CBRN (chemical, biological, radiological, nuclear) attack requests while keeping legitimate dual-use research, per Google's Frontier Safety Framework. Includes monitoring the model's internal activations to spot misuse, tested by internal and external red teams
  • Prompt injection: Google calls it its most resilient model against indirect prompt injection, leading on Gray Swan's IPI benchmark
  • Misalignment: monitors watch Argon's chain-of-thought and actions and stop execution when needed. Google used a similar system during training, without feeding the findings back into training (so it wouldn't teach the model to hide its reasoning from the monitor), and asks the rest of the industry to keep reasoning transparent
  • Hardened sandboxes: isolated, sealed environments before high-risk training or evaluations

See AI Safety.

Caveats

  • Benchmarks are Google's setup. Competitor numbers are mostly self-reported at their maximum reasoning setting. Several Argon numbers are self-computed (DeepSWE with a mini-swe-agent harness, Terminal-Bench 4.0, OSWorld-2.0 maxed over three runs). The two Terminal-Bench figures for Opus 5.5 (66.4% in Google's table, 60% at Artificial Analysis) show how much the test setup matters
  • You can't test it yet. As of 3 October 2026, it isn't listed on the Gemini API models or pricing pages. Artificial Analysis is the only independent evaluation so far. Update (19 October 2026): Argon has been on the Gemini API since 13 October (see Availability), and StarSkirmish Bench's 15 October edition added a second independent result (see above)
  • The price is temporary. Budget with $4 / $20, not $2 / $10. The intro price ends on 30 November 2026
  • Open questions: the API model ID, a release date, and whether other Gemini 4 sizes follow. The first two are answered (gemini-4-argon, 13 October 2026 on the Gemini API); the third is still open

References

About Sébastien

Ready to get to the next level?

Found this valuable? Share it with someone who needs it.

Join 6,000+ readers. Get practical systems for knowledge & AI. Free.

Subscribe ✨

Free: Knowledge System Checklist

A clear roadmap to building your own knowledge system. Subscribe and get it straight to your inbox.

6,000+ readers. No spam. Unsubscribe anytime.

Subscribe