Gemini 4 Argon's $2/$10 API intro price doubles Dec 1
The model topping Vals and ERQA is cheap only until Nov 30 — re-run your evals now. Plus Anthropic's distillation lockdown now breaks some reasoning calls.
Copy markdown
The price, plainly
gemini-4-argon costs $2 per 1M input and $10 per 1M output through Nov 30, then doubles to $4/$20 on Dec 1 — and thinking tokens bill as output. Cached input stays $0.10 (a 95% discount) and the Batch API is half price; any paid API account can call it, but it's off the free and Pro tiers.
What the money buys
Argon tops the Vals Index at 68.90% (ahead of Claude Sonnet 5.5's 67.04 and Opus 5.5's 66.97), hits 87% on ERQA embodied reasoning and 77.9% on DeepSWE. It ships a 1,048,576-token window shared across input, thinking and output, with a 65,536-token default output cap you can raise toward the full window.
Do this before the hike
Set max_output_tokens on every call — Argon stops once thinking plus output reach the cap, so it also caps your bill — and swap explicit caching for implicit, which is all Argon supports. Re-run evals on anything pinned to Gemini 3.5 or 3.7 Flash: those requests now route to 3.6 and 3.8.
A sharper deadline: Oct 23
Google retires the deep-research-pro-preview-12-2025 agent on Oct 23 — move pipelines to deep-research-preview-04-2026 or the -max variant before it goes dark. It's an easy one to miss with the Dec 1 price change dominating the headlines.
Anthropic's distillation lockdown can 400 your calls
Anthropic says it cut extracted reasoning tokens by 90-99% — and since Sep 24 it bills benign requests its classifier wrongly blocks (a sub-0.1% false-positive rate). On Claude Fable 5.1 and accounts created from Aug 31, returning a reasoning block after you change the system prompt, tools or a prior message now throws a 400, and cross-account Sonnet 5.5 reasoning blocks are stripped before the model sees them.