Gemini 4 Argon hits the API: $2/$10, up to 1M-token outputs
Google's frontier model is GA for developers and Pro users, leads agentic-coding benchmarks, and undercuts Opus 5.5 — until intro pricing ends Nov 30.
Copy markdown
It's live in the API today
Gemini 4 Argon is generally available on the Gemini API paid tier (AI Studio, Android Studio, Vertex), inside Antigravity, and to Google AI Pro/Ultra subscribers. It's proprietary and API-only — no open weights — with a March 2026 knowledge cutoff and text, image, audio, and video input.
Cheaper than Opus 5.5 — but a clock is running
Intro pricing is $2 per million input tokens and $10 output, with cached input 95% off ($0.10/M). That beats Claude Opus 5.5 ($4/$20) and GPT-6 Astra ($10/$50). On Nov 30 it jumps to $4/$20, matching Opus — so lock in batch jobs and evals before then.
One response, up to a million tokens out
Max output climbs from 64K to as much as 1M tokens in a single response via a new continuation protocol, inside a 1M-token window shared by input and output. Whole-repo refactors, long transcriptions, and book-length generations stop needing manual chunking.
It leads agentic coding — and trails on terminals
Argon tops DeepSWE v1.1 at 77.9% (Opus 5.5 74.2%, GPT-6 Astra 74.1%) and GraphWalks long-context at 84.2%, but lags on FrontierSWE v2 (55% vs Astra's 65.5%) and Terminal-Bench 4.0 (57.4% vs Opus's 66.4%). Strong for long-horizon agents; test it before trusting it on shell work.
In Antigravity, +35% code submissions
Google says Argon raised agentic code submissions over 35% and cut rollbacks nearly a third in its own Antigravity IDE, and dropped indirect prompt-injection success to 0.7%. If you're on Antigravity or the Agent Development Kit, the agent loop is the real upgrade, not the chat.