Google opens Gemini 4 Argon to everyone — at $2/$10

Two weeks ago, Google announced Gemini 4 Argon and then promptly locked it behind the Fairwind Program — a gated rollout for trusted cybersecurity defenders only. On October 13, 2026 at approximately 12:00 (UTC), Google dropped the gate. Argon is now available to developers via the Gemini API, to enterprises through Gemini for Business, and to Google AI Ultra and Pro subscribers inside the Gemini app and AI Mode in Search. The frontier model race just got a fourth serious contender at the consumer table.

What changed

The September 30, 2026 announcement (at around 20:00 UTC) was always a phased release. Google DeepMind SVP Koray Kavukcuoglu framed it as a safety-first rollout: start with cyber defenders, gather feedback, iterate on guardrails, then expand. The October 13 update is that expansion — and it is broader than expected.

Who gets access starting October 13, 2026 (UTC):

• Developers: Gemini API paid tier; Google AI Ultra subscribers also get it in Google Antigravity, AI Studio, and Android Studio.
• Enterprises: Gemini for Business, the API, and Google Antigravity.
• Everyday users: Gemini app and AI Mode in Search — but only for Google AI Ultra ($99.99/month) and Pro ($20/month) subscribers. Free and Plus tiers are excluded for now.
• Scientists: 10,000 researchers get the highest Antigravity usage tier free for six months, applications open now.
• Cyber defenders: Fairwind Program expands to hundreds of additional partners, with the no-cyber-guardrails version available on a rolling basis.

The specs that matter

Sources: Google blog (Oct 13, 2026), Artificial Analysis, Meta blog (Oct 12, 2026), Anthropic. "—" indicates the benchmark was not reported by the vendor or independent evaluator at time of writing.

The 1 million unified input-output token limit is the headline technical spec. Google raised the output ceiling from 64,000 to 1,000,000 tokens — a 15.6x jump — giving the model room to generate hundreds of thousands of tokens of reasoning in a single trajectory. The continuation protocol in the API lets developers chain outputs beyond even that.

Pricing is aggressive. At $2/$10 during the introductory period (expiring November 30, 2026 at 23:59 UTC), Argon undercuts GPT-6 Astra ($3/$15) by 33% on input and 33% on output, and Claude Opus 5.5 ($4/$20) by 50%. After the intro period, Argon moves to $4/$20 — matching Opus 5.5 exactly. The 95% cached-input discount means repeat queries with large context effectively cost $0.10 per million input tokens.

What Google is actually showing off

Google's blog is dense with real-world deployment evidence, not just benchmark scores. The most credible examples:

• Chrome refactor: Argon in Antigravity helped the Chrome team separate core webpage handling from a ~6 million line legacy folder inside Chromium's 40 million-line codebase, completing "months of work in weeks" across hundreds of files. The payoff for end users: live webpages will be able to move fluidly between tabs, side panels, and floating windows without reloading.
• Gemma inference speedup: Argon achieved a 2.85x end-to-end inference speedup on an A100 via vLLM for the open Gemma 4 E2B model — from ~37,900 to over 108,000 tokens per second — with 100% token parity and zero quality drop, in days rather than months.
• Quantum computing: Argon translated Google's internal AlphaQubit quantum error correction decoder into a single optimized hardware kernel, nearly doubling inference throughput. For the arcsin subroutine in quantum simulation, an Argon-written implementation formally verified in Lean cut spacetime cost by over 100x versus the published CORDIC baseline.
• Nuclear fusion: Working with Professor Matt Landreman at the University of Maryland, a Google team discovered a broader class of counterexamples to "Grad's conjecture" — a 1967 hypothesis that exact mathematical solutions for stellarator magnetic fields were impossible.
• Rare disease diagnosis: Gamow Labs used Argon to analyze de-identified newborn genomes. In one case, Argon pinpointed a chromosomal rearrangement that routine software missed for years. On private evals, Argon scored 57% on RareBench and 52.2% on LabBench.
• Robotics: Argon watched a raw video of a human playing cornhole, mapped spatial coordinates, calculated whole-body joint movements, built simulation code, and transferred it to Apptronik's Apollo humanoid — which sank seven consecutive shots in a real-world trial.

Why this matters

The timing is not accidental. Meta dropped Muse Wonder on October 12, 2026 (at approximately 16:00 UTC), claiming frontier-level coding performance at Opus-matching prices. Google's wide Argon release 20 hours later is a direct response — and the pricing tells the story. At $2/$10 introductory, Google is essentially selling frontier compute at a loss to buy market share before the November 30 price reset.

The competitive landscape now has four frontier models at the $4/$20 price point or below: Argon ($4/$20 post-intro), GPT-6 Astra ($3/$15), Opus 5.5 ($4/$20), and Muse Wonder ($4/$20). The differentiator is no longer raw benchmark score — they are clustered within a few points of each other on the Artificial Analysis Index — but distribution and integration. Google has Search, Android, Chrome, and Workspace. OpenAI has ChatGPT's 300+ million weekly users. Anthropic has enterprise lock-in. Meta has open weights and Facebook/Instagram distribution.

Argon's real moat may be the 1M token limit combined with the 95% cache discount. For enterprise workflows that process the same large document sets repeatedly — legal contract review, financial analysis, codebase understanding — the effective cost per task could be dramatically lower than competitors. Google's claim that Argon matches Astra on the Artificial Analysis Index at 60% of the cost per task is plausible precisely because of this cache economics.

The skeptical read

Google's benchmark table is self-reported and, as OrcaRouter noted in an October 1, 2026 analysis, "completely unreproduced." The DeepSWE v1.1 score of 77.9% is impressive, but Meta's Muse Wonder Full claims 79.0% on DeepSWE 1.1 — and these may be different benchmark versions or evaluation protocols. Independent evaluators like Artificial Analysis have not yet published a full Argon re-test post-wide-release.

The consumer rollout is also narrower than the "available for everyone" headline suggests. Free users get nothing. Google AI Plus subscribers get nothing. Only Pro ($20/month) and Ultra ($99.99/month) subscribers get access in the Gemini app and Search. That is a paid-gated launch, not a general release — and it mirrors the strategy OpenAI used with GPT-6 Astra.

The Fairwind Program's "no cyber guardrails" version for defenders deserves scrutiny. Google is intentionally shipping a model without cybersecurity safeguards to trusted partners, arguing that defenders need the full capability. This is the same logic Anthropic used with its Cyber Verification Program. The risk — that a trusted partner gets compromised or an insider exfiltrates the unguarded model — is real, even if Google assesses it as manageable.

What to watch

Three concrete indicators will determine whether Argon is a genuine frontier leap or a well-marketed incremental step:

1. Independent Artificial Analysis re-score — if Argon holds its claimed tie with Astra at lower cost after third-party testing, the pricing pressure on OpenAI and Anthropic becomes acute. If it falls short by 5+ points, the "frontier" label is marketing.
2. Chrome tab migration shipping — Google claims the Argon-assisted refactor will enable live webpage movement between tabs without reloading. If this lands in stable Chrome by Q1 2027, it is tangible evidence of agentic coding in production at scale.
3. Post-November 30 retention — when the introductory $2/$10 pricing expires and Argon moves to $4/$20, will developers stay or churn back to Astra at $3/$15? The answer reveals whether Argon's capability justifies parity pricing or whether it was the discount doing the work.

Gemini 4 Argon is no longer a rumor or a gated experiment. It is a real, available frontier model — and the fourth entrant in a price war that shows no sign of slowing.