Evaluating Raven Max: #1 on Terminal Bench 4.0 and Agent's Last Exam

Raven Max is our flagship reasoning model, built for complex engineering, long-horizon planning, and decisions where errors are costly. This is its full empirical evaluation against the current frontier: Fable 5, GPT-5.6 Sol, Gemini 3.8 Flash, Muse Spark 1.3, and Kimi K3. All evaluations run with maximum reasoning effort enabled.

#1 on Terminal Bench 4.0

Terminal Bench 4.0 measures autonomous software engineering in a live terminal. Raven Max scores 42.8, first globally among the models compared:

#1 on Agent's Last Exam

On Agent's Last Exam, a long-horizon agentic reasoning benchmark, Raven Max scores 39.5, outperforming the nearest peer, Fable 5, by nearly 8 points. Kimi K3 places fourth among the peers at 27.6:

Headline results

Why Max leads on multi-step work

The pattern across these benchmarks is multi-step dependency resolution: Raven Max plans across long chains of tool calls, keeps intermediate state coherent, and streams its reasoning before executing actions. On Terminal Bench and Agent's Last Exam, that architectural choice shows up as the widest margins in the suite. On GDPVal-AA, which measures real economic task completion, Raven Max leads at 1,771.

Evaluate it yourself

Benchmarks are the floor, not the ceiling. Run Raven Max against your own hardest tasks in the playground, or through the API at $1.25 per 1M input and $6.25 per 1M output tokens.

All evaluations run with maximum reasoning effort enabled. Full methodology and additional benchmarks appear on the models page.