Anthropic's Claude Opus 5 Tested in New Post-Training Delivery Benchmark

Post-training is turning into a service business: a customer hands an operator its data and a goal, and expects back a fine-tuned, evaluated, deployed model, on a budget, with a human sign-off step along the way. The new benchmark asks a sharper question than most agent evaluations bother with. It is not whether an agent can push a metric higher. It is whether the agent can be trusted to actually deliver, end to end, without anybody having to check its homework.

The authors built what they call a governed delivery plane. An agent drives ten stages of the job, and an oracle scores each stage from facts recorded by the platform itself rather than from the agent’s own account of what it did. The failure mode the paper is built around is what it names TBDL, for “trains but doesn’t learn”: loss falls, every dashboard stays green, and the delivered model performs no better than the untouched base. That is the silent failure a customer paying for the work would otherwise never catch. An acceptance gate run by the operator is shown catching every such run before payment goes out, and a separate detector, calibrated on runs the researchers deliberately corrupted, flags severe corruption while a job is still running.

Four frontier agents were run through the full pipeline: Claude Opus 5, GPT-5.6-luna, Gemini 3.7 Flash and DeepSeek V4-Pro, working on open base models ranging from 8B to 70B parameters on metered L40S, A100 and H200 GPUs. A human forward-deployed engineer was run through the identical oracle as a comparison point.

The numbers

The benchmark’s headline figures are about the test rig rather than a leaderboard: ten governed stages per delivery job, four agents including Claude Opus 5 compared against a human engineer, and base models spanning 8B to 70B parameters across three GPU types. Anthropic’s own published pricing for Claude Opus 5, $5 per million input tokens and $25 per million output tokens with cache reads at $0.50, sets the per-run cost a customer would actually pay before any rerun is added on.

The paper does not publish which of the four agents produced the fewest TBDL runs, only that an acceptance gate caught every one before payment. For anyone evaluating a post-training vendor, that gate, not the headline metric, is the detail worth asking about first.