Claude Opus 5 just scored 4x higher than GPT-5.6 Sol on novel problem-solving

Anthropic dropped a new flagship model that reshapes the price-performance conversation. Claude Opus 5 doesn’t just inch past the competition on a few benchmarks—it posts a genuinely weird result on ARC-AGI-3, a test explicitly designed to measure reasoning without memorized patterns. Opus 5 scores 30.2 percent. GPT-5.6 Sol manages 7.8 percent. That’s not a gap, that’s a different weight class.

The model achieves this while costing the same as its predecessor: $5 per million input tokens and $25 per million output tokens. That’s half the price of Anthropic’s own Claude Fable 5, which Opus 5 beats in agentic coding (43.3% vs 33.7% on Frontier-Bench) and knowledge work. It’s a direct shot at pricing pressure from OpenAI and Chinese labs, making Opus 5 the default on Claude Max and the best model available on Pro plans.

But token rates don’t tell the whole story. Anthropic’s own prompting guide warns that higher effort settings don’t always mean better results. On two benchmarks, Opus 5 actually performs slightly worse at max effort than at the second-highest setting, while burning more tokens. It’s a repeat of a pattern we saw with Opus 4.7, where per-task costs ended up 30 to 40 percent higher despite identical base rates. The company is nudging users toward “low” and “medium” settings for most work.

The most revealing detail isn’t a benchmark number. It’s a specific test where Opus 5 had to recreate a 3D model from a drawing it couldn’t directly view. The model wrote its own computer vision pipeline from scratch to extract geometry from raw pixels, then reconstructed the part. No other model solved it in five attempts. An engineer at a trading firm reportedly shipped a market data feed in one session—something previous models couldn’t do with detailed plans. Benchmarks are useful, but that kind of adaptive tool-building hints at something benchmarks miss.