Gemini 4 Pro / Flash: benchmarks, hands-on tests and leaks

A chart shows Argon leading outright on 13 of 19 results, scoring 84.2% on GraphWalks at 256K-1M tokens and 19.6% on Harvey's legal benchmark, nearly three times the next model, while still trailing Astra and Opus 5.5 on FrontierSWE and Terminal-Bench.

A French user says the Gemini 4 benchmarks look good but is too wary to let it navigate their codebase without hearing from the community first.

Google released Gemini 4 Argon with a 1M token output limit, and early reviews put it ahead of OpenAI and Anthropic on most benchmarks.

Google says Gemini 4 Argon sets a record in real software engineering and ties GPT-6 Astra and Grok 4.7 for first in cybersecurity, and it is not public yet, starting with trusted security partners.

An evaluation says Gemini 4 Argon ranks fourth on the RSI index, generalizes very well and does not plateau on KSP Bench, joining Fable 5.1, GPT-6 Astra and Opus 5.5 among the few models that can beat the game.

A Spanish-speaking user says Google has released Gemini 4 Argon with impressive benchmark results and is eager to try it.

Linus Ekenstam posts a chart of Gemini 4 Argon results, calling it comparable to GPT-6 Astra and Fable and declaring Google back in the race.

Gemini 4 Argon raises the output limit to 1M tokens, can autonomously find and patch vulnerabilities, and is priced at $2 per 1M input and $10 per 1M output tokens, starting with select security teams.

Google's Gemini 4 Argon targets software engineering, enterprise work and cyber defense, beats GPT-6 Astra, Opus 5.5 and Fable 5.1 on several benchmarks, and is rolling out to paid API users and AI Ultra subscribers.

LuminaBench says Gemini 4 Argon took first place on the Arena text leaderboard with a score of 1533, well ahead of Opus 5.5.

Bloomberg reports that Gemini 4 scores well on industry benchmarks but employees with direct access say it struggles with certain coding tasks, fueling internal skepticism ahead of launch.

A poster says Gemini 4 Argon outperforms GPT-6 Astra, Fable 5.1 and Opus 5.5 on SWE and is priced low, with comparison data in the image.