Two days after the GPT-5.5 launch, the first independent benchmark comparisons are here — and the results are more nuanced than the headlines suggest. On the ten benchmarks both providers report, Opus 4.7 leads on six, GPT-5.5 leads on four. Margins range from 2 to 13 points.
Where GPT-5.5 Leads
OpenAI’s new model shines at autonomous work. On Terminal-Bench 2.0, GPT-5.5 scores 82.7% according to VentureBeat — Opus 4.7 manages 69.4%. GPT-5.5 also leads on BrowseComp, OSWorld-Verified, and CyberGym. The pattern is clear: wherever a model needs to operate tools independently and work autonomously over extended periods, OpenAI wins.
Token efficiency is also notable. According to Artificial Analysis, GPT-5.5 uses about 40% fewer tokens for comparable tasks. With doubled API prices ($5/$30 per million tokens), that brings the net price increase down to roughly 20%.
Where Claude Opus 4.7 Leads
Anthropic’s flagship dominates tasks that require deep thinking. SWE-bench Pro: 64.3% vs. 58.6%. MCP-Atlas, MMLU, and HLE reasoning without tools — Opus 4.7 leads across the board. The pattern fits: for code reviews, complex reasoning, and scientific questions, Claude remains the stronger model.
On pricing, Opus 4.7 is also 17% cheaper at $25 per million output tokens compared to GPT-5.5’s $30. But when you factor in token efficiency, GPT-5.5 can actually end up cheaper in practice.
And What About Mythos?
The benchmarks also show that GPT-5.5 narrowly beats Anthropic’s secretive Mythos Preview model on Terminal-Bench 2.0. However, Mythos is still a research preview and not generally available — so direct comparisons should be taken with a grain of salt.
My Take
The era of one model dominating everything is over. GPT-5.5 is the better model for autonomous agents and computer use. Opus 4.7 is the better model for code quality and complex reasoning. For most of us, that means: it depends on the use case. And that’s actually good news — competition makes everything better.
Sources: