Gemini 4 Argon Lands With 1M-Token Output and a Safety-Gated Rollout
Google announced Gemini 4 Argon on September 30 with a 1 million token output limit, a DeepSWE state of the art at 77.9%, and defensive-cyber focus, but most users cannot touch it yet: rollout runs through a trusted-defenders program first.
Google announced Gemini 4 Argon on September 30, closing out the strangest release month in memory, and its two headline numbers point in opposite directions. The first is capability: a 1 million token output limit (up from 64K), a DeepSWE v1.1 state of the art at 77.9%, a 51.3% top score on Zapier’s AutomationBench, and the lead on Wiz’s black-box penetration-testing benchmark. The second is access: Argon is currently rolling out to trusted cyber defenders through the Fairwind Program, with paid API customers, Ultra subscribers, and consumers later, an unusual sequencing Google attributes to its Frontier Safety Framework, since the model is explicitly designed for autonomous offensive-and-defensive security work. Introductory pricing is $2/$10 per million tokens, rising to $4/$20 after the intro period, which lands it exactly on the GPT-6.1 Sol and Sonnet 5.5 price line the market settled at this week.
The 1M Output Limit Is the Underappreciated Number
Every frontier lab now ships million-token context windows; Argon ships a million-token output, which is a different and harder engineering claim. Sixty-four thousand tokens of output was the practical ceiling that forced agents to checkpoint, summarize, and lose state mid-task. A million tokens of continuous generation means an agent can produce an entire codebase migration, a full audit, or a complete formal proof in one uninterrupted run, which is precisely what Google says it used Argon for internally: C-to-Rust migrations of up to 800,000 lines including the Fuchsia Zircon kernel, and a libgav1 port that replaced 32K lines of SIMD code with a memory-safe decoder running 2.7 times faster. For agent builders, output length has been the quiet constraint that shaped every architecture; Argon makes it a budget problem instead.
The Cyber Focus Explains the Gated Rollout
Argon’s positioning as a defensive-cyber model explains why Google is doing something no lab has done with a flagship: shipping it to security professionals before developers. The model leads CWE-bench v1 at 68% on vulnerability remediation, discovered a critical healthcare-software flaw through Wiz’s Scan for Good program, and is designed to refuse cyber and CBRN misuse per Google’s Frontier Safety Framework, with chain-of-thought monitoring that can halt execution on detected misalignment and hard sandboxes for high-risk runs. Google also engaged the US government’s voluntary pre-release access process. The sequencing inverts the usual launch: the riskiest capability set ships to the most vetted users first, and the general public gets it after the incident track record exists. It is the model-level counterpart to Nvidia’s network-silicon containment: trust earned incrementally rather than assumed at launch.
A Month of Releases, Read Together
September’s releases now form a coherent spectrum of access philosophies. OpenAI shipped GPT-6.1 Sol and Dots to paying customers immediately; Anthropic shipped Opus and Sonnet 5.5 broadly with safeguards built in; Google ships its strongest model to vetted defenders first and everyone else later. The Argon bet is that regulatory and enterprise buyers will pay for a model whose deployment story includes gatekeeping, which matters the same week the FTC opened an investigation into AI product risks and Australia’s agent-breach review is still running. Pricing tells the same story: $2/$10 introductory is competitive, but the post-intro $4/$20 is a premium for the governance package, not just the model.
What to Watch
Three things. First, whether the Fairwind Program produces a public incident record (vulnerabilities found, misuse attempts refused) that justifies the phased rollout, because that record is the product Google is really building. Second, DeepSWE 77.9% under independent verification, since a seven-point jump over the prior SWE state of the art inside a month would be remarkable. Third, the intro-period clock: when Argon’s price doubles to $4/$20, its competitiveness against Sol 6.1 and Sonnet 5.5 at $2/$10 becomes the first real test of whether safety-gated access carries a price premium in this market.
Get Insanely Good at AI
The book for developers who want to understand how AI actually works. LLMs, prompt engineering, RAG, AI agents, and production systems.
Keep Reading
How to Use Symbolic Execution for Automated BPF Analysis
Learn how Cloudflare uses the Z3 theorem prover to instantly generate magic packets and reverse-engineer BPF bytecode for security research.
Gemini 4 Carbon reaches the API: $2/$10 until November 30, then double, and a safety report that calls it no material change
Google opened Gemini 4 Carbon to paying developers and AI Ultra subscribers on November 10, with a 1M-token window and a price rise to $4/$20 on November 30. Its safety report carries Argon's risk findings forward, while showing a 12.8-point jump in exploit success.
Gemini 4 Carbon nine days on: top of the index, a third of Claude's task cost, and still no API
Google's October 28 frontier model leads Artificial Analysis at 59 and costs about $16 per GDPx task against Opus 5.5's $51. Developers still cannot call it, its hallucination rate rose on one test, and its linked safety report reportedly covers Argon.
Gemini 4 Hydrogen replaces Flash at $1.50/$7.50: faster agents, triple the Terminal-Bench score, and a higher bill per token
Google released Gemini 4 Hydrogen on November 12 as its everyday model for coding and agents. It scores 51 on the Artificial Analysis index, beats Argon, costs twice Flash's current price, and its model card admits more unjustified refusals.
Claude Haiku 5.5 Collapses Small-Model Prices to $0.10 Per Million Input Tokens
Anthropic released Claude Haiku 5.5 on October 7 with GDPval 1620 against Haiku 4.5's 735, an adjustable effort setting, and pricing of $0.10/$0.50 per million tokens for prompts under 100k, roughly 90% cheaper than its predecessor.