What is Claude Opus 5?

Claude Opus 5 is the new state of the art on ARC-AGI 3 with a score of 30.2%, dwarfing the previous best of 7.8% set by GPT-5.6 Soul.

“Opus 5 is the new state of the art on ARC-AGI 3 with a score of 30.2%. This absolutely demolished all the other models. The previous high score was GPT-5.6 Soul at 7.8%.”

System design, not raw model capability, is the primary lever for unlocking frontier-level long-horizon AI performance, as demonstrated by a 30-to-100% ARC-AGI 3 improvement through orchestration alone.

“Reaching 100% on Arc-AGI 3 from a 30% model baseline with Claude Opus 5, their conclusion was that system design, rather than model capability alone, can unlock frontier-level long-horizon performance.”

NVIDIA Evo achieved 100% on ARC-AGI 3 starting from a 30% model baseline using Claude Opus 5, attributing the gain to system design rather than raw model capability.

“Reaching 100% on Arc-AGI 3 from a 30% model baseline with Claude Opus 5, their conclusion was that system design, rather than model capability alone, can unlock frontier-level long-horizon performance.”

Anthropic expects 85% fewer refusals from Claude Opus 5 compared to Fable 5, attributing the difference to a distinct set of guardrails on Opus.

“As a result, Anthropic is using a different set of guardrails on Opus than they do on Fable 5, which they believe will lead to 85% fewer refusals.”

Claude Opus 5 scored 43.3% on Frontier Bench, roughly 10 points above Fable 5 and 9 points above GPT-5-6-Soul on the same benchmark.

“It scored a 43.3% on that one that I just mentioned, Frontier Bench, which is a more difficult version of Terminal Bench, which is about 10 points higher than Fable 5 and around 9 points higher than GPT-5-6-Soul.”

Anthropic removed 80% of the system prompt for Claude Opus 5, Fable 5, and Claude Code with zero change to coding benchmark results.

“Anthropic had removed 80% of the system prompt for Opus 5 and Fable 5 and Claude Code. He said this resulted in zero change to their coding benchmarks.”

Claude Opus 5 (Anthropic) is worth watching as the first model to autonomously build its own computer vision pipeline to complete a task when denied image access, a capability no rival including Mythos demonstrated.

“Opus 5 created its own computer vision pipeline to view the image before successfully recreating the part. No other model, including Mythos, was able to complete this task.”

Anthropic intentionally did not train Claude Opus 5 on cyber tasks, yet the model still achieved meaningful improvement in finding code vulnerabilities without that training.

“Anthropic pointed out that Opus is intentionally not trained on cyber tasks. The model has still achieved solid improvement on finding vulnerabilities in code, making it similar to Mythos 5 in that aspect.”

Claude Opus 5 beats Fable 5 on many benchmarks but is described as nowhere near Fable 5 in practical use, illustrating a widening gap between benchmark rankings and real-world utility.

“Opus 5 is nowhere near Fable in practical use, not even close. Anyone who's used it meaningfully can tell this very quickly after a few tasks. Yet, Opus beats Fable on many.”

Nathaniel Whittemore on the practical gap between Claude Opus 5 benchmark scores and real-world performance.

“Opus 5 is nowhere near Fable in practical use, not even close. Anyone who's used it meaningfully can tell this very quickly after a few tasks. Yet, Opus beats Fable on many.”