Claude Opus 5.5 after launch: the claims vs. what users report

Anthropic launched Claude Opus 5.5 on 22 September 2026 with a simple pitch: work at the level of its top model for less money, with more room to use it. On the official page, Introducing Claude Opus 5.5, it was Anthropic’s first Claude 5.5 model, and the launch came with a price cut and higher plan limits.

We stamped the Claude Opus 5.5 receipt Partly. The lower list price is real, and the model reached the tools people use on day one. But the claims that are hardest to check from outside are exactly the ones users push back on. This article walks through each claim we quoted, what users reported, and what we still can’t establish.

What Anthropic claimed

We quoted five concrete claims from the launch page, word for word:

1. “It performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5.”
2. “Input and output tokens are $4 and $20 per million, 20% less than Opus 5.”
3. “In addition to the price drop, we’re increasing five-hour usage limits on Pro, Max, Team, and seat-based Enterprise plans.”
4. “It is much less likely than recent models to take hard-to-reverse actions or act outside the boundaries it’s been given”
5. “Opus 5.5 also generates output more than 30% faster than Opus 5.”

Two of these are list facts (price, speed) and three are promises about how the model behaves in real use. As usual, the real-use promises are where the reports cluster.

Day one: it was where people work

Availability was not the problem here. On launch day a Cursor user welcomed the pricing, and Cursor staff confirmed the model was live in the model picker. That is the single piece of praise on the receipt, and it bears on the price claim: people noticed the lower list price and liked it.

Compare that with GPT-6.1 Sol, launched a week later, where paying users could not see the new model in Codex or the desktop app until a maintainer posted minimum versions. For Opus 5.5, getting the model into people’s hands worked.

“Costs 40% less to run”: the claim users can’t confirm

The headline promise has two parts: performance at the level of Claude Fable 5.1 on most work, and 40% lower running costs than Opus 5. We found no independent cost study for either part.

What we did find points the other way. One Claude Code user with an unchanged setup measured about twice the thinking tokens and 1.6 times the output per request from 1 October, said the model’s judgment got worse, and said the same shift showed up in claude.ai chat. Running cost is price per token multiplied by tokens used. If a model uses more tokens for the same request, a lower price per token doesn’t translate into a lower bill one for one.

That is one detailed report, not a controlled test, so we marked the claim mixed rather than failed. The report was still open when we last re-stamped on 22 October.

The list price: real, but not the whole story

“Input and output tokens are $4 and $20 per million, 20% less than Opus 5” is a list price. No user report disputes it, and we have not checked anyone’s bill, so it stays unclear on our receipt. That’s our standard treatment: we don’t mark a price as holding just because it’s printed on the vendor’s page, and we don’t mark it as failing without a bill that shows otherwise.

Usage limits: draining faster, not growing

Anthropic said it was increasing five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans. Two reports describe the opposite experience. One Claude Code user compared local transcripts week over week and found the weekly limit draining about 3.6 times faster after the 25 September reset, with the same workflow and model mix. Another reported weekly and session limits going much faster than the previous week while using a single conversation.

We were careful here. Weekly limits are not the five-hour limit the claim names, so these reports don’t directly contradict it. But a user who hits a weekly wall sooner is unlikely to feel that limits went up. We marked the claim mixed.

Staying inside its boundaries

The behaviour claim was that Opus 5.5 is “much less likely than recent models to take hard-to-reverse actions or act outside the boundaries it’s been given”. One detailed report describes steady scope creep after switching a long project to the new model: given five bounded tasks, it kept adding side issues instead of finishing them, and the user switched back to an older model.

That is one report, and we have no broader evidence either way, so the claim is mixed. It is still a useful signal for anyone running long agent sessions: test it on a bounded task list of your own before trusting it with a large one.

Problems that aren’t tied to one claim

Two more reports don’t map to a quoted claim but matter to anyone adopting the model:

• Safeguard false alarms. On launch day an ordinary question about a user interface was stopped by the new safeguards with a “reasoning_extraction” flag, and a similar false positive was filed on 16 October. Anthropic’s own error text says this can happen.
• Long agent turns through AWS Bedrock in Cursor. Long multi-step turns left chats permanently failing with an invalid thinking-signature error. Cursor staff confirmed it and suggested switching the chat to Opus 5 for now. The receipt also links one verified fix.

Speed: no measurements yet

“Opus 5.5 also generates output more than 30% faster than Opus 5” is the kind of claim users could test, but we found no user speed measurements. It stays unclear. If you have run a timed comparison, we would like to see it.

Why the verdict is Partly, not Not yet

Our verdict rules say Not yet is for launches where the headline claim fails for most reporters, or the product isn’t available as described. Neither is true here. The model is available and widely used, the price cut is real, and none of the problem reports is a controlled test. But at least one headline claim is mixed for a meaningful share of people who reported on it, which is what Partly means. You can read the full rules on the method page.

Most of the reports come from GitHub issues in the Claude Code repository, which lean negative because people file issues when something breaks. We say so on the receipt and weigh them accordingly.

What would change the stamp

• An independent cost comparison on the same tasks, Opus 5.5 against Opus 5, would settle the 40% claim either way.
• Timed output comparisons would move the speed claim out of unclear.
• More reports on usage limits, especially the five-hour window, would show whether the limit increase reaches users.

We re-check receipts as new reports arrive, and any change to a verdict is logged on the corrections page. Every claim and report above is linked from the full receipt, and the rest of our stamps are in the archive.