Opus 5.5, Day One. The Endurance Gap Just Got a Lot Smaller.

Yesterday Anthropic shipped Claude Opus 5.5. I said in May that I don't write about every model release, and I meant it. But this one lands squarely on the distinction I've been using all year — endurance versus intelligence — so it's worth a short note.

What they're claiming

Three things stood out in the announcement, and I'm quoting the framing rather than the benchmarks:

• Roughly Fable 5.1's level on most work. That's their line, not mine. Fable is the long-horizon model I've been using for the connected, easy-to-corrupt chains — like the entire company-formation project this series is about.
• About 40% cheaper to run than Opus 5, with output more than 30% faster.
• A 1M-token context window, adaptive thinking always on, and Sonnet 5.5 and Haiku 5.5 following in the coming weeks.

If the first claim holds, the practical question I've been asking operators — what shape is your task? — gets a simpler answer for a lot more of the day.

What I did with it on day one

I didn't run a benchmark. I ran my Tuesday.

The Clarent build has a long tail of connected work right now: the books, the CRM structure, the brand system, a pile of post-formation paperwork from the state. Every one of those depends on decisions made two weeks ago. The classic failure mode is the model quietly re-litigating something already settled — treating the entity type or the bookkeeping approach as an open question when it isn't.

Day-one impressions, with the caveat that one day is an anecdote:

• It held the thread on the long stuff. Working through the ledger and the CRM pipeline in one sitting, it kept the earlier decisions locked. Where Opus 4.8 wobbled on my attribution problem in May, 5.5 didn't visibly wobble on this. I'll keep watching.
• It's faster in a way you feel. Not "benchmark faster." Waiting-less faster. Drafting and revising felt closer to a conversation than a queue.
• The writing is more direct. Anthropic called this out and it's real. Fewer throat-clearing sentences. For someone who edits a lot of AI-drafted prose, that's hours a month.

The question was never "which model is smartest." It was "which model can hold the whole chain." As of yesterday, the everyday model holds a lot more of it.

What didn't change

I still think about task shape. Most of my work is short and self-contained — answer this, draft that, decide this — and any strong general model handles it. The long connected chains are a minority of my hours and a majority of my risk. What changed is that the everyday model now covers more of that minority, so I reach for the specialist less often.

And I still verify. A model that holds the thread beautifully can hold a wrong thread beautifully. The ledger reconciles against the bank feed, not against the model's memory. The formation documents get read by a human. Better endurance doesn't relax that; it just means fewer silent mid-chain corruptions to catch.

The operator's takeaway

If you've been paying for two tiers of model — one for daily work, one for the long stuff — it's worth re-running your own hardest chain on 5.5 this week and seeing whether the gap is still there for you. Mine narrowed. Yours might not, depending on what your long chains look like.

I'll say more once I've had a few weeks with it. In the meantime: if you've hit a task where 5.5 either impressed you or fell over, tell me in the comments. Real failure cases are more useful to me than praise, and I'll share what I find.