Gemini 4 Argon benchmarks: 77.9% on DeepSWE, ahead of Opus and Astra
Output limit rises to 1M tokens, Terminal-Bench 4.0 scores 57.4%, introductory pricing is $2/$10, and access is currently limited to trusted testers.
gemini 4 argon is here. and google is back at the frontier. some benchmarks: > deepswe v1.1: 77.9%, new state of the art. ahead of opus (74.2%) and astra (74.1%) > terminal-bench 4.0: it scored 57.4% it's not a clean sweep, but it's right there with the best. > it has 1M output tokens up from 64k. it can think and write for hundreds of thousands of tokens in one go. > google's own agents used it to free up 300+ TiB of memory across their data centers price: > $2/$10 introductory price. it's same as gpt-6.1 sol and sonnet 5.5. goes up to $4/$20 later. but you can't use it yet. it's rolling out to trusted defenders first. paid API and AI ultra come after.
This post quotes another post: x.com/ananth7e/status/2105392874779185299