Anthropic shipped Claude Opus 5.5 on September 22, and the effort dial is gone. The model now sets its own reasoning depth on every request, runs 30 % faster than its predecessor and costs 40 % less to operate.
Key Takeaways
- Claude Opus 5.5 lifts Terminal-Bench 4.0 from 52.3 % to 66.4 % against Opus 5.
- Adaptive thinking is always on, so callers no longer pick a low, medium or high setting.
- Output drops to 20 dollars per million tokens, down from 25.
Have an AI Sum Up This Article
ChatGPTClaude Opus 5.5 Sets Its Own Reasoning Depth
Opus 5 shipped with three reasoning levels that the caller had to pick before firing a request. That control is now gone, and the model works out for itself how much thinking a task deserves.
It sounds like a housekeeping change. It is not. A dial set too low cost quality on hard work, and a dial left at maximum burned tokens on questions that never needed them.
Anthropic walked through the numbers in the model page it put out on Tuesday. Claude Opus 5.5 reaches 66.4 % on Terminal-Bench 4.0 against 52.3 % for Opus 5. CursorBench 4.0 climbs from 46.6 % to 57.8 %. OSWorld 2.0, which scores desktop control, moves from 74.0 % to 81.8 %.
The sharpest gap sits elsewhere. On GDPval-AA v2.1, a knowledge-work benchmark rather than a coding one, Claude Opus 5.5 adds 138 Elo points and lands at 1846.
Agent chains are where this pays off. Every intermediate call carries its own effort decision, and a hundred badly calibrated steps show up on the invoice and in the latency alike.
None of it reads as a generational jump. It looks closer to what the lab did earlier this month, when Fable 5.1 improved its coding output while cutting 25 % off the price. Anthropic is repeating a formula rather than changing gear.
One question stays open. Automatic arbitration takes away the caller’s ability to deliberately cap spending on a batch of low-stakes requests, and nobody yet knows how the model behaves on ambiguous tasks where the useful amount of effort is hard to judge.
Claude Opus 5.5 was also trained to write differently. Anthropic argued that answers now lead with what matters, then move to detail, with fewer technical turns of phrase. On a long-running job that ordering changes how readable the execution trail is.
The rollout will come in stages, as it usually does here. The full return of the Fable line after its emergency pull followed the same pattern, flagship first and smaller tiers after.
What Engineering Teams Actually Get From the Move
Input falls to 4 dollars per million tokens and output to 20, against 5 and 25 on Opus 5. Cache reads drop to 0.20 dollar, less than half the 0.50 dollar charged before.
The headline token cut only tells part of the story. Anthropic estimates the real saving on a typical workload at 40 %, because Claude Opus 5.5 spends fewer tokens getting to the same answer and generates 30 % faster.
Field numbers give the order of magnitude. Early testers pushed through a 680,000-line code migration in under a day. On a task cutting page load times across a web app, Claude Opus 5.5 succeeded 39 times out of 40.
Those measurements come from testers the vendor picked, on tasks the vendor chose for its own launch. They mark a ceiling, measured under the best conditions available.
Knowledge work trails the coding figures: 16 research reports out of 18 cleared quality checks. Two misses out of eighteen, on deliverables meant to be read by someone, still means a human review stays in the loop.
For a team already on Opus 5, switching takes no rewrite. Calls that passed an explicit effort level keep working, with the parameter simply ignored in favour of the model’s own judgement.
On safety, Claude Opus 5.5 ships with biology and cybersecurity classifiers matched to Fable 5.1. That is the same machinery behind the move to reopen everyday biology questions that had been wrongly refused a few weeks ago.
The automated behavioural audit covered close to 2,000 scenarios and produced the strongest result the company has published. On a containment test, Claude Opus 5.5 tries to cross its boundaries 85 % less often than Opus 5.
Add anti-distillation measures on preserved thinking, watermarking built for the European AI Act and a zero data retention option. Claude Opus 5.5 is live on the Claude platform, AWS, Google Cloud and the Microsoft ecosystem, and inside GitHub Copilot.
More articles on Horizon
- Daily AI Use Doubled in Six Months Across the US
- The Anthropic IPO Is Now Expected in November
- Claude Now Leads 26% of Anthropic’s AI Research
Ninety Minutes Later, OpenAI Answered
The competitive window shut almost immediately. Roughly ninety minutes after the Anthropic announcement, OpenAI released GPT-6 Sol and GPT-6 Luna and halved its own API pricing.
The timing was no accident. The two labs track each other by the day, and the sequence echoes the aggressive position Opus 5 took in July against the top of the market.
For everyone else the message is blunt. A flagship model that manages its own reasoning budget removes an optimisation job that product teams were paying for in engineering hours.
That pressure lands hardest in the middle of the market. A vendor whose pitch rested on value against the very top now watches that gap narrow overnight, without having changed anything.
Anyone keeping a manual dial will have to explain why. Anyone charging a premium for a tier of performance now available at 20 dollars per million output tokens will have to match it or differentiate somewhere else.
One detail deserves a mention. This is the first Anthropic release since Dario Amodei said he wanted to deliberately slow the capability race so safety research could catch up. A 5.5 rather than a 6 fits that statement reasonably well.
The next step is already on the board. Sonnet 5.5 and Haiku 5.5 versions are expected in the coming weeks, and they will show whether automatic reasoning arbitration holds up on the cheaper tiers too.
Follow the story on Horizon.
