The reasoning-effort dial gets a reality check

Anthropic asks users to flag Opus 4.8 over-thinking the same week a top practitioner finds high-effort reasoning fails at contradictions — and Simon Willison fact-checks Uber's AI spend in real time.

Two posts on the same May 28 timeline tell a story about where frontier-model design is actually stuck. Anthropic's Alex Albert asked users to flag Opus 4.8 over- or under-thinking after release — and the next day, practitioner jxnlco observed that high-thinking models "kind of just lose their mind in contradictions," while low-thinking variants stay robust. The reasoning-effort knob, marketed as a 2026 power tool, is harder to tune than the launch posts admit.

The pattern is consistent across the May 28-29 signal set. Albert's request for examples of over- and under-thinking is not a closing-the-loop post — it is a vendor saying, in public, that even with internal calibration the production envelope is unclear. jxnlco's observation cuts harder: when a prompt contains contradictory information, the model that thinks longer is the one that fails. That inverts the implicit benchmark logic of the past two model generations, where more reasoning compute meant better answers across the board.

The accountability theme bleeds into the enterprise narrative the same week. Simon Willison spent May 29 pushing back on a quoted figure that Uber's CTO had spent $300m on tokens, repeatedly asking interlocutors to cite a credible source before deciding "what's acceptable and set a cap." A separate follow-up noted the CTO comment had been edited out of a transcript because it was "harder to edit into non-speaking-prose." When practitioners are fact-checking executive AI-spend numbers in real time, the gap between vendor calibration and buyer scrutiny is widening, not closing.

For builders shipping with Opus 4.8 (released May 28, 2026) or any frontier reasoning model, the practical move is to test both ends of the thinking budget on contradiction-heavy prompts before defaulting to the highest setting. For buyers, the Willison posture — name the source, name the number, refuse to set policy on hearsay — is the right one for any "we spent $X on AI" pitch landing in your inbox this quarter. And for platform vendors selling enterprise control planes, observability into when models over-think matters as much as the dashboards showing how much they think.

The bottom line: This year's competitive edge isn't the highest reasoning ceiling — it's knowing when to dial it down.

Sources