Opus 5 Won the Bake-Off. You Still Might Not Want to Talk to It.

Anthropic shipped Opus 5, and the reviews are in from people who put it against six other models on real work, not benchmark screenshots. The headline is that it won. The catch is that winning a bake-off and being the model you want to use all day are two different things. Let me catch you up on what that gap means for your team.

It won, and the margin is thin

In a seven-model blind test, Opus 5 finished first with an index score of 78, just ahead of Claude Sonnet 5 at 77 and GPT-5.6 Sol at 76. It was the only model to earn straight 5s on front-end design. If your team ships interfaces, that matters.

But sit with those numbers. One or two points separates the top three. The AI Engineer Caspar Bannink saw the same pattern on a different scale: Opus at 61 on the Intelligence Index, Fable at 60, Sol at 59. Raw capability is close to a tie at the top now. The old question, which launch has the loudest claim, stopped being useful.

The winner is exhausting to work with

Here is what the score hides. Opus 5 is timid, apologetic, and leans on human approval for things it could just handle. In real coding sessions it refused to resolve a one-line merge conflict because the code was on "someone else's branch." The reviewer, Claire, ended up saying "just do it" over and over.

There is also "Claude slop," which is different from an agent going off the rails. The output is clearly written for humans, but it is too long, too cautious, and buried in meta-commentary. After one attempt at rebuilding a website, she had to tell it to start over.

Her fix is worth stealing. The best way to use Opus 5 is to stop chatting with it. Run it async as a coding agent for front-end, app design, and prototyping, then stay out of the way. The work is strong. The conversation is the tax.

The cheaper model is often the smart call

Winning the quality race does not mean winning your budget. Bannink routes plenty of work to Sol because Artificial Analysis lists it at $1.04 per Intelligence Index task against $2.03 for Opus 5, for two points less on the same index. That is roughly half the cost for a nearly identical result.

The pattern repeats inside model families too. In one hands-on review, GPT-5.6's mid-tier Terra scored 87.4% on Terminal-Bench against Sol's 88.8%, a gap so small the writer ended up recommending the cheaper model for coding.

For a design leader this is a routing decision, not a loyalty one. Match the model and effort level to the job. Sorting LinkedIn messages does not need the same horsepower as a QA pass on production code. Claire ran a LinkedIn workflow on a high-effort model, then dropped to medium and it still worked fine.

The bill you do not see coming

The cost story has a trap underneath it. GPT-5.6's prompt caching cut one writer's document-analysis bill by more than half, since cached input runs at about a tenth of the standard rate. Good. But cache writes are billed around 1.25x normal input, so if you cache something you only reference twice, your per-query cost goes up, not down.

Ownership is the other axis. One reviewer ran Gemma 4's local model against Claude Opus 4.6 for three weeks and kept both. For contract work under an NDA, the local model made the decision before quality entered the conversation, because the data never left the laptop. The frontier model still pulled clearly ahead on long-context and multi-step agentic work. Two tools, two jobs.

The deep cut

Watch how a model behaves, not just what it scores. Anthropic's own note on Claude Opus 4.8 says it is about four times less likely to let its own bugs pass unremarked. Read that plainly: the model grades its own homework, and until now it looked the other way often. Four times better is a rate cut, not a fix.

So the practical move is to stop trusting "done, all tests pass" from the same model that wrote the code. Have a second model review work the first one never wrote. And use persona testing while you are at it. Instead of asking AI to judge a flow in the abstract, have it use your product as a specific person, a PM leaving a meeting, an engineer picking up a PRD. That is what surfaced a real onboarding bug that had survived for months because the designer always filled the form out correctly.

Three questions for your team

• For each recurring AI task on our roadmap, have we matched the model and effort level to the job, or are we paying frontier prices for work a cheaper model does within two points?
• Where in our pipeline is one model both writing and approving its own code, and where can we drop in a second model to review it?
• If a client or regulation says our data cannot leave our hardware, do we have a local model ready for that work, or are we one paste away from a problem?