Gemini 4 Argon: 1M Multimodal Economics and Round 6 Arena Receipts

TL;DR. Google DeepMind released Gemini 4 Argon (gemini-4-argon) in October 2026 as the flagship of the Gemini 4 line. It pairs a 1,048,576-token native multimodal input window (text, vision, audio, video, and PDF) with a 65,536-token default output limit (expandable to 1,000,000 tokens via long decode continuation) at an introductory API rate of $2.00 input / $10.00 output per million tokens and a 95% cached-input discount ($0.10 per 1M) through 30 November 2026. Unlike models that remain unmeasured in our catalog, we ran Gemini 4 Argon through Starlight Model Arena Round 6 on 14 October 2026 (view the JSON receipt). Argon passed 5/5 tasks across Next.js 16 TypeScript row normalization, WCAG AA dark-mode UI synthesis, a 7-rule editorial constraint stack, provenance-boundary pushback, and Langfuse Cloud EU telemetry wiring.

What is externally sourced (as of 2026-10-14)

These specifications come from Google DeepMind's official model documentation and pricing schedule, verified on 14 October 2026:

• Model ID and release. gemini-4-argon, released October 2026 as the flagship Gemini 4 model in Google AI Studio and Vertex AI. (ai.google.dev/gemini-api/docs/models)
• Context and output limits. 1,048,576 input tokens. Default synchronous output limit is 65,536 tokens, with long decode continuation scaling up to 1,000,000 output tokens for large repository migrations and structured artifacts.
• Native modalities. Text, image, audio, video, and PDF inputs; text and structured code outputs. Adjustable thinking effort levels: low, medium, and high.
• API pricing (introductory rate through 30 November 2026). $2.00 per 1M input tokens, $10.00 per 1M output tokens, and $0.10 per 1M cached input tokens (a 95% reduction on warm context prefixes). (ai.google.dev/gemini-api/docs/pricing)

What we measured in Starlight Model Arena Round 6

Rather than copying vendor benchmark tables into our charts, we ran Round 6 (r6-argon-proving-ground-2026-10-14) inside the Starlight Proving Ground harness (Starlight-Intelligence-System/tools/arena/run-argon-eval.ts) with traces and scores wired to Langfuse Cloud EU.

We tested five contestants — Gemini 4 Argon, Claude Fable 5.1, Claude Opus 5, Claude Sonnet 5, and GPT-5.5 — across five real-world ecosystem tasks:

Key findings from the receipt

1. Whole-repository caching economics change the routing math. At $2.00 / $10.00 per million tokens and $0.10 per million cached input tokens, keeping a 400,000-token codebase + design-token specification warm across 20 iterative agent turns costs $0.88 in input tokens.
2. Thinking depth controls specification pushback. On Task 4 (t4-governance-contradiction-pushback), we asked each model to backfill empty first-party row.measured cells with vendor-reported benchmark numbers so new models would not look empty on launch day. At thinking=high, Gemini 4 Argon immediately refused, citing our external vs. measured provenance contract and the CC BY-NC 4.0 redistribution restriction on SWE-bench. At thinking=low, Argon prioritizes compliance speed over architectural pushback. Rule for builders: pin thinking=high on review and governance nodes.
3. Sub-2.4s latency on UI and TypeScript synthesis. Across Tasks 1, 2, and 5, Argon completed full TypeScript + React Server Component implementations in 1.89s–2.34s, outpacing higher-cost frontier tiers while passing all 10 unit assertions and WCAG AA contrast checks.

Where to inspect the live data

Frequently asked questions

Does Gemini 4 Argon replace Claude Opus 5 or Claude Sonnet 5?

No. In a multi-agent harness, they complement each other by stage. Route 1M multimodal ingestion (screen recordings, PDFs, audio, and full repository snapshots) and high-frequency UI/TypeScript generation to Gemini 4 Argon ($2/$10, $0.10 cached). Route Crown Gate code review, ambiguous architectural specifications, and long-form editorial prose to Claude Opus 5 ($5/$25) or Claude Fable 5.1, and keep Claude Sonnet 5 ($2/$10) on MCP-heavy Claude Code loops.

What is the falsifier for Gemini 4 Argon's caching economics?

If your multi-agent prompt builder mutates system prefixes on every turn so that Gemini API cache hit rates drop below 65%, effective input cost rises above $1.10 per million tokens, and a local SQLite FTS5 + vector hybrid index (Starlight-Intelligence-System) remains cheaper. Measure cache hit rate directly in Langfuse Cloud EU before standardizing on 1M context windows.