When the obvious route closes, it finds another

In Anthropic’s FreeCAD evaluation, Opus 5 was asked to rebuild a machine part from a drawing it could not directly see. Rather than stop there, Anthropic says the model built a computer-vision pipeline, extracted the geometry and completed the model.

That episode captures the launch proposition neatly: Opus 5 is meant to find another route, inspect what it produces and keep working towards a finished handoff. It is Anthropic’s evaluation, not independent proof, but it gives us a more useful question than whether a model can produce an impressive first response.

Token price tells us what generation costs. It does not tell us what the task costs once we include failed approaches, repeated prompts, manual corrections and the time spent checking whether the result is actually usable.

That is why Anthropic’s emphasis on verification matters. If Opus 5 completes more of the loop itself, the saving may appear in fewer repairs rather than a cheaper token. Whether that happens consistently is something we still need to test on our own work.

Anthropic has kept Opus 5 at the same base API price as Opus 4.8: $5 per million input tokens and $25 per million output tokens. The effort setting then lets us trade faster, cheaper runs for deeper reasoning without changing models.

Higher effort can consume more tokens, so it is not automatically the economical choice. The useful measure is the cost to finish the job: model spend, elapsed time, corrections and human intervention together. Fast mode adds another trade-off, with around 2.5 times the speed at twice the base price.

Anthropic describes Opus 5 as a thoughtful, proactive model that approaches Fable 5 intelligence at a lower task cost in selected comparisons. That is narrower than saying it is universally half the price, and more useful because the claim can be tested against named workloads.

The supporting material combines Anthropic-run evaluations with reports from early-access customers. It can show us where to look, but it cannot establish how often the same gains will survive different prompts, tools, data and review standards.

Claude Opus 5 is much stronger at verifying its work and iterating carefully until it succeeds.

Anthropic — Introducing Claude Opus 5

The strongest reading is specific: on Anthropic’s Frontier-Bench v0.1 presentation, Opus 5 moves beyond Opus 4.8 on both performance and cost per completed task. CursorBench 3.2 adds a similar supplier-published signal at higher effort settings.

Neither chart gives us a universal productivity rate. They are task-bound comparisons, and the Frontier-Bench presentation includes Anthropic’s own run conditions and fallback handling. We can treat the result as a reason to test long coding work, not as a settled verdict on every workload.

The FreeCAD example gives this loop a concrete shape. Faced with a drawing it could not directly inspect, the model created another way to extract the information it needed. In a separate package-manager evaluation, Anthropic says Opus 5 traced a bug to its root cause and found an edge case missed by an existing patch.

These are reported episodes rather than guarantees. Their value is in the behaviour they illustrate: the model does not have to treat a blocked route as the end of the task, and it can use inspection and revision as part of the work rather than waiting for us to notice every gap.

We can start with one task that normally needs several tools, decisions and checks. Before the run, we define what finished means. During it, we record the effort setting, model spend, elapsed time, failed routes and each correction we still have to make.

The final review remains ours. Cyber safeguards may restrict or reroute some requests, including fallback to Opus 4.8 in covered cases, and a confident handoff can still be wrong. The point of the trial is to see whether Opus 5 reduces repair work without hiding cost or risk.

The useful model is not the one that looks smartest at the start. It is the one that helps us reach a checked finish with less repair.

Creator Broadcast synthesis, based on Anthropic’s published evidence

Anthropic is not presenting Opus 5 as its absolute frontier. It is presenting a model close to that frontier, available at everyday Opus pricing and designed to spend effort where a longer task demands it.

That proposition becomes valuable only if the whole session improves. Completion, corrections, latency, safeguards and total cost belong in the same comparison. A real, repeatable task will tell us more than a headline score alone.