July 25, 2026/1271 words/6 min read

Opus 5 Might Win Me Back

Anthropic shipped their best model in years.

In April I wrote that Claude had gotten really dumb and that Anthropic had spent too much of my trust. Three months later, Anthropic has released Opus 5, and apparently public complaints now have the shelf life of yoghurt.

I should be delighted! I am mostly annoyed that the new model addresses nearly every reason I moved away.

The pitch is unusually sensible for a frontier model release. Opus 5 gets close to Claude Fable 5 on coding and knowledge work, including agentic tasks, while keeping Opus 4.8's API price of $5 per million input tokens and $25 per million output tokens. Fable costs twice as much. Opus 5 retains the full one-million-token context window and supports up to 128,000 output tokens.

For Claude subscribers, it becomes the default on Max and is the strongest model included with Pro. Anthropic is selling a model people can use every day. A practical enterprise detail is buried near the bottom of the announcement too. Unlike Fable 5, Opus 5 has no special data-retention requirement for general access. Fable requires 30-day retention. For a company with a zero-data-retention agreement, that makes the comparison very short.

And yes, the benchmark sheet is silly in the familiar way all model benchmark sheets are silly. Every company picks the tests and effort settings that flatter the model it released five minutes ago. Anthropic's charts still make a stronger case than I expected.

Fable beats Opus 5 by only 0.8 percentage points on SWE-bench Pro, 80% to 79.2%. On Cognition's FrontierCode main set, the gap shrinks to 0.1 points, 53.5% to 53.4%. Those gaps are basically lint.

The direction flips on FrontierBench, where Opus 5 scores 43.3% against Fable's 33.8% and GPT 5.6 Codex's 34.4%. It also wins on OSWorld 2.0 with 70.6%, ahead of Fable at 66.1% and GPT 5.6 Sol at 62.6%. Opus 5 tops the GDPval-AA and AA-Briefcase leaderboards for professional and long-running knowledge work as well.

Fable still wins in places. It edges Opus 5 on DeepSWE and FrontierCode, and Anthropic still recommends it for the hardest long-running autonomous work. Mythos remains stronger in cybersecurity and biology. GPT 5.6 Sol also leads the DeepSWE result in Anthropic's own system card. There is enough mess in the table to make it more credible. Opus 5 is very close to Fable, sometimes ahead and sometimes behind, which is exactly what Anthropic said it would be.

The cost claim needs the same amount of caution. Opus 5 is half the token price of Fable and priced exactly like Opus 4.8, but it is still an expensive model. Artificial Analysis currently ranks the max-effort version first on its Intelligence Index. The same evaluation found it slower than most models in its class and much more verbose. It generated 100 million output tokens against a class average of 63 million, making it one of the more expensive models in the group. Max effort can apparently turn a good bargain into a very clever furnace.

Fortunately, max may also be the wrong setting.

The most useful detail in the system card is that Opus 5 reaches its best FrontierCode score at medium effort. Above high, it starts changing code outside the request. Nearby refactors and "improvements" drag the score down because nobody wants a bug fix to arrive with an unsolicited renovation of the kitchen.

Anthropic's own prompting guide spends a surprising amount of time on this behaviour. Opus 5 may keep working past the task's boundaries, either by widening the scope or verifying its work more than once. All that effort comes with too much narration, and its documents run longer than those from earlier models. Old prompts that explicitly demand final verification or a separate review pass can make it repeat work it would already have done. Anthropic recommends removing that scaffolding and setting clear boundaries on narrow tasks. It also suggests trying low or medium effort before spending more tokens.

That explains some of the mixed first impressions. People are dropping Opus 5 into workflows built around the shortcomings of Opus 4.8, then discovering that the compensations have become instructions to waste time. Workflows that shout "check your work" three different ways now make it repeat itself. A plugin that forces a reviewer into every task causes the same problem when the model already delegates well. More intelligence can make an old workflow worse for a while, which is funny and slightly infuriating.

It also tracks with the early reports that Opus 5 can feel literal or occasionally argumentative. People are also getting better results at medium or low effort than max. The model sometimes looks brilliant after they remove half the machinery built to make previous Claude models reliable. That is not exactly a drop-in upgrade, though it may be a healthy reason to delete several pages of agent instructions.

GPT 5.6 offers a useful comparison. What I liked about it was its willingness to finish the boring parts around the interesting problem. It keeps using the repository and the available tools until the checks pass. When it notices a missing state along the way, it handles that too rather than stopping at the main implementation. The model became better at completing work at the same time that Codex became better at giving it a computer to work with.

Anthropic says Opus 5 now finishes full tasks rather than leaving stubs. During review, it catches real bugs without burying them in false positives.

Across longer tasks, Opus 5 stays coherent throughout the one-million-token context window, and its subagents can work without trampling each other's files. The launch demos are suitably ridiculous. In one FrontierBench task, the model had to rebuild a part from a drawing it could not directly view. It solved that by writing a computer-vision pipeline to inspect the pixels before opening FreeCAD. Nobody needs that on a Tuesday evening, but it says something useful about whether the model gives up when the obvious path is missing.

The guardrail story also looks less awful than it did in April. Anthropic's system card reports a 0.47% over-refusal rate on benign requests in Claude.ai, slightly better than Opus 4.8 and Fable 5. On FrontierBench, Opus 5's cyber classifiers triggered in 4% of trials, compared with 26% for Fable. Flagged requests can still fall back to Opus 4.8, so the product can still change models underneath a task when the classifier gets nervous. I would rather it did not! But anyone whose normal work contains words that frighten the safety layer should hit that fallback 85% less often than with Fable.

One launch day is not enough to decide whether a model is trustworthy. We have been through the cycle too many times. Every model spends its first day being called astonishing, usually beside a screenshot of a 3D game built in one prompt. The rate limits catch up, and within three weeks half the same people are convinced it has been secretly nerfed.

Opus finally has a coherent place in Anthropic's lineup. It costs enough less than Fable that long agent runs no longer feel irresponsible, while retaining enough of Fable's judgment for difficult work. That leaves Fable for the small number of jobs where the last bit of capability is worth twice the price. Sonnet remains the better fit when speed and volume matter more.

Back in April, I wrote that Claude still had flashes of brilliance but I no longer trusted the product around them. Repairing that trust will take longer than a day. The value proposition is coherent again, though, and medium effort is the setting I want to test rather than max.