Access status: Fairwind preview vs public API GA
This is the decisive difference for shipping code this week. Google announced Gemini 4 Argon on 2026-09-30 and is rolling it out first to trusted cyber defenders via the Fairwind Program while it strengthens frontier safeguards; public AI Studio and general Gemini API access are not available yet, with broader release planned to start with paid API customers and Google AI Ultra. Claude Sonnet 5.5 shipped 2026-09-28 as API id claude-sonnet-5-5 on the Claude Platform, Amazon Bedrock, Google Cloud, and Microsoft Azure — developer GA today. Do not treat Argon as interchangeable with a production Claude default until Google opens the paid API path.
Official benchmarks (different suites — no fake head-to-head)
Vendors publish different evals; we only cite each lab's own numbers. Gemini 4 Argon (Google blog): DeepSWE v1.1 long-horizon SWE 77.9% SOTA; Zapier AutomationBench #1 at 51.3%; LVBench long-video 91.7%; CWE-bench v1 tied first at 68%. Claude Sonnet 5.5 (Anthropic): Terminal-Bench 4.0 70.6% (vs Sonnet 5 10.3%); CursorBench 4.0 55.5%; GDPval-AA 1844 (near Opus 5.5's 1846). These are not the same tests — use them as directional signals for long-horizon SWE / automation / cyber defense (Argon) versus agentic terminal and IDE-session coding (Sonnet 5.5), not as a single scoreboard winner.
Pricing: same intro band, different post-intro path
Claude Sonnet 5.5 matches Sonnet 5 at $2 input / $10 output per million tokens (cache reads $0.20); Anthropic says it often needs fewer tokens and can cost up to ~30% less per task, with 30%+ faster output. Gemini 4 Argon introductory price is also $2 / $10 per million input/output (cached input ~95% off); after the intro period Google lists $4 / $20. Argon list price after intro is therefore 2× Sonnet 5.5's current public API rate — but Argon is still Fairwind-gated, so most teams cannot burn that budget in production APIs yet.
Context and output headroom
Claude Sonnet 5.5 ships with a native 1M-token context window and up to 128K max output on the Message API (per Anthropic developer docs / launch materials) — no beta header required for 1M context. Gemini 4 Argon expands max output to 1M tokens (from ~64K), which Google positions for single-trajectory deep reasoning on hard problems; full public input-context specs for the developer API remain limited while access is Fairwind-only. For monorepo agent sessions you can run today, Sonnet 5.5's GA 1M context is the practical ceiling; Argon's 1M output is the headline capacity to watch once API GA lands.
Who should pick which (and how to follow the news)
Default to Claude Sonnet 5.5 for daily coding agents, Claude Code sonnet alias (2.1.284+), Cursor, and any CI that needs a stable public model id now. Shortlist Gemini 4 Argon if you are a Fairwind-eligible defender, need Google's cyber-defense / long-horizon SWE posture, or are planning a paid-API migration once Ultra / paid tiers open — keep a dual-stack design so Sonnet 5.5 remains the fallback. Track our related stories for access and changelog updates rather than third-party fully-GA rumors.