Google announces Gemini 4 Argon with Fairwind-first access, 1M output tokens and $2/$10 intro pricing

What happened

Google unveiled Gemini 4 Argon on September 30, 2026 as its next frontier model, but almost nobody can call it yet. According to a dev.to write-up by Toronto-based AI engineer YongBo Yu, the model is initially live only for participants in Google's Fairwind Program for vetted cyber defenders. Broader access is described only as coming soon, with paid API customers and Google AI Ultra subscribers first in line, and no public model identifier exists — so outside Fairwind there is no way to build against Argon today.

Capability and positioning

The post frames Argon as a model built for extended, multi-step software engineering, enterprise knowledge work, and defensive cybersecurity, with Google saying thousands of its own staff already use it internally for specialized coding, research, and writing. Notably, Google intends to give trusted defenders and its internal teams a build with cyber guardrails removed so they can use the model's full defensive capabilities. For everyone else, safeguards covering CBRN and cyber misuse, prompt injection, misalignment monitoring of chain-of-thought and actions, and hardened sandboxes are still being finalized while Google participates in the U.S. government's voluntary pre-release process.

Pricing doubles after the intro window

Google's stated rates begin at $2 per million input tokens and $10 per million output tokens during an introductory period, then rise to $4 and $20. Cached input carries a 95 percent discount off the input price. The dev.to author's advice is to model both columns now: workloads that look cheap at launch rates can double in cost when the promotion ends, and long agent loops that repeatedly resend the same system prompt, repository summary, or tool schema will hinge on cache hit rates once the API is available.

The competitive contrast is sharp. OpenAI shipped GPT-6.1 Sol on September 29 at the same $2/$10 band, and per the OpenAI and GitHub changelogs cited in the post it is already callable via API and live in GitHub Copilot.

A 1M token output ceiling

Argon's output limit expands from 64K to 1M tokens, which Google presents as industry-leading generation headroom. The practical consequence, per the post, is that a single agent trajectory could think, call tools, revise, and emit a large artifact — a migration plan, a multi-file patch set, a research memo — without mid-flight truncation or artificial "continue" handoffs. The flip side is runaway-spend risk, so the recommendation is to add budgets, stop conditions, and per-step logging before ever obtaining a key.

Vendor-reported benchmarks

Google's own numbers, relayed by the post and flagged as vendor-reported rather than independently verified, include 77.9 percent on DeepSWE v1.1 for long-horizon software engineering — with 9to5Google cited putting Claude Opus 5.5 at 74.2 percent and GPT-6 Astra at 74.1 percent. Argon also posts 51.3 percent on Zapier's AutomationBench, 91.7 percent on the long-video LVBench, and 68 percent on CWE-bench v1 for vulnerability remediation, with Google claiming first place or a tie on each. The company additionally points to leading placement on the Vals Index and strong results on Vals Finance Agent v2 and Harvey's Legal Agent Benchmark.

Internal results Google cites

Google's announcement also includes internal anecdotes: quantum subroutine optimization 40 percent better than a published baseline in one example, fleet memory work that freed over 300 TiB, CRust migrations covering more than 800,000 lines of re2 and libgav1 code for Fuchsia's Zircon kernel, and a libgav1 Rust SIMD rewrite claimed 2.7 times faster with identical video output. On security, Wiz's Scan for Good is credited with surfacing a critical healthcare exposure that earlier frontier models missed. The post stresses these are vendor examples, not service-level guarantees.

What developers should do now

The post's guidance: keep shipping on models callable today such as GPT-6.1 Sol, Claude Opus and Sonnet 5.5, or Grok where it fits; build long-horizon software engineering evals covering multi-file edits, migrations, and flaky tests so a harness is ready when a model ID appears; price any Argon ROI case against the post-intro $4/$20 rates and assume cache misses until measured; and strictly isolate any guardrail-free build, since Fairwind is a defense capability, not the public chatbot's sandbox.

Why it matters

Frontier launches now split into two products: the capability announcement and the availability calendar. Argon is a substantial capability claim — longer output, cyber defense focus, headline benchmark wins — paired with a Fairwind-first calendar that leaves most developers unable to touch it. Teams that budget against introductory pricing or assume day-one API access will get burned. Until a public model ID and billing path exist, Argon is a planning signal, not a shipping surface, and the durable move is having evals, spend controls, and rollback paths ready for the day it lands.