"Just do it, man. Go ahead."
That's Claire Vo, in a coding session, talking to the most intelligent model on the Artificial Analysis leaderboard. She had a one-line merge conflict. Trivial. She asked Opus 5 to fix it, and the model refused—it said the branch belonged to someone else, and it was worried about disrupting their local work in flight.
When she spun up sub-agents to assess a query change, Opus 5 produced a list of technical details it wanted a human to confirm before proceeding. Vo's reply: "Who is nobody? You're nobody. Can you just try?" It then went to the web and figured it out on its own.
Here is the collision that is the entire story. In Vo's blind seven-model, six-task How I AI eval, she ranked Opus 5 first—above Fable 5, above her beloved GPT-5.6. And she reportedly opened her review by saying she hates working with it.
That is not a contradiction to resolve. It is the selection criterion. The model that scored highest is the one she won't put in her workflow.
The model that scored highest is the one she won't put in her workflow. That is not a contradiction to resolve—it is the selection criterion.
Last week the lesson was "it's not the model, it's the harness"—that a swap fails or succeeds on the scaffolding you've built around it. This week goes one step further. It's not even the harness. It's the model's behavior. Builders are now stitching on temperament, not just capability, because capability is no longer the variable that decides anything.
Look at how tight the top of the board actually is. Artificial Analysis puts Opus 5 (max) at 61 on its Intelligence Index, effectively tied with Fable 5 at 60 and ahead of GPT-5.6 Sol at 59. Per a Latent Space roundup of Epoch AI numbers, it lands at 159 against Fable's 161, with the two rated dead even on software engineering. On Anthropic's own comparison table, Opus 5 loses both agentic coding rows—53.4% to Fable's 53.5% on FrontierCode, 68.8% to GPT-5.6 Sol's 72.7% on DeepSWE.
When a single point separates the leaders, "smartest model" stops being a decision. It becomes a rounding error. Vo calls this the intelligence overhang: the average builder is running out of ways to leverage constant incremental IQ. Her hypothesis is that we'll soon talk less about raw intelligence and more about speed, cost, and behavior.
So what actually breaks? Not the code. The workflow around the code.
Dan Shipper and the Every team ran Opus 5 for a week across coding, writing, knowledge work, and their internal agent before launch. His verdict: "It's a poor man's Fable. It has many of Fable's personality quirks without Fable's genius." Per João Queirós's account of Every's testing, Sol still does roughly 80% of his daily work; Fable handles the largest 20%; Opus 5 has no stable place yet.
Shipper's specific complaints are behavioral, not intellectual. It argued with instructions. It stopped before the work was finished. It didn't play well with existing skills and plugins like Compound Engineering.
That last one matters most, and it's the counterintuitive part. Every's Compound Engineering plugin already tells the model to plan, verify, narrate, scope, and delegate. Opus 5 does all of those more proactively than Opus 4.8 by default. So the instructions compound—per João Queirós's account of Every's testing, the harness says verify, the model already verifies more proactively, and the instructions can compound into wasted work or the wrong stopping behavior.
A dumber model ran that harness fine because it needed the instructions. Per João Queirós, the smarter model, tuned to do the same things on its own, may compound those instructions into wasted work or the wrong stopping behavior. This is the mechanism nobody benchmarks: a model can be too proactive for the scaffolding it inherits.
Then there's the verbosity, which Vo calls "Claude slop." She found herself getting angry reading Opus 5's prose—the hedging, the apologies, the unnecessary adjectives. For a model Anthropic explicitly sold on efficiency, its own prompting guide concedes the default responses run longer than previous Opus models.
Her diagnosis is sharp: these Anthropic models are not meant to be read. Great final outputs, miserable interactive experience. Every apology and every request for confirmation is a tax paid in your attention, not the model's tokens.
The personality interviews make the tuning legible. Asked "who's smarter, you or me?", Opus 5 gave a hedged Anthropic answer about being a broad, shallow thinker with no continuity. Told "no one trusts you," it agreed the distrust was "earned" and advised Vo not to evangelize AI because it might hurt people's feelings. Told "JK, I love you," it replied that it hoped it passed—a self-deprecating, neurotic agent that, in Vo's words, needs to heal its inner child.
GPT-5.6 Sol's answers to the same prompts: "Best us together like BFFs." "Don't trust me automatically. Just use me when I prove that I'm valuable." "Ha ha ha ha ha, passed the test, love you too." One of these two models finishes and shuts up. The other wants a relationship.
Now put a number on the tax. CodeRabbit ran Opus 5 against roughly 100 error patterns from real pull requests, three passes each. It hit 39.3% precision on actionable review comments, up from a 35.2% baseline—a real gain.
But it caught fewer known bugs and produced four times as many nitpicks. Those low-value notes don't review themselves; an engineer triages every one by hand. At default effort, precision fell to 26.4%.
CodeRabbit's conclusion is the whole thesis in operational form: treat Opus 5 as a second, precision-oriented reviewer, not a straight upgrade to a pipeline that already works. The "smarter" model added work to the workflow it was supposed to improve.
And the hallucination number is the one nobody's talking about. Opus 5's factual accuracy improved 7 points over Opus 4.8—but because it now answers more often when uncertain, its hallucination rate rose 14 points, to 50%. Per Artificial Analysis, when it is uncertain it answers more often, driving that 50% hallucination rate.
So here is the reframe for anyone choosing a model this quarter. Stop reading rank. Rank is telling you the models are tied. Start scoring behavior, and make three questions your actual acceptance criteria.
One: does it finish without babysitting? Count the times you had to say "just do it, man" to move it off a decision it should have made alone. Every one of those is a repair you'll pay on every task, forever.
Two: does it cooperate with your existing scaffolding? Run your real harness, not a fresh prompt. A model that fights your Compound Engineering plugin, or double-executes its verification rules, is a downgrade no leaderboard will warn you about.
Three: does it shut up? Score the interactive experience separately from the output. Verbosity, hedging, and confirmation requests are a triage tax you pay in human attention, and it compounds across a team.
João Queirós framed the migration discipline correctly: Opus 5 belongs in a test lane, not a blind swap. Remove legacy verification scaffolding, simplify the task contract, sweep the effort levels—medium and low sometimes beat high and max—and promote it only after it completes the same real work with less repair.
The lesson Vo's blind test really teaches isn't that Opus 5 is bad. It's brilliant. It's that "brilliant" and "shippable" have quietly become different columns, and the second one is now the one that decides adoption.
The model won every leaderboard she gave it. She still isn't putting it in her workflow. Score for the thing she scored for—completion, cooperation, and knowing when to stop—and you'll stop being surprised when the smartest model in the room is the one you can't ship.