Anthropic launches Claude Opus 5.5

Anthropic published Claude Opus 5.5, the first model in its Claude 5.5 family, saying it performs at Claude Fable 5.1 level on most work while costing about 40% less to run than Opus 5, with faster output and raised subscription usage limits.

SOFTWARE desk — Anthropic just shipped a cheaper, faster Claude in the Opus line for everyday work, right after it asked the industry to slow the race toward more powerful models.

What the company says the model is for. Anthropic says Opus 5.5 performs at the level of Claude Fable 5.1 on most work. Fable 5.1 is an earlier Claude the company has treated as a top model. Opus 5 is the Opus that came before this one. “Most work” is the company’s phrase, not a test this desk defined. The same page says that on its own benchmarks Opus 5.5 leads at coding agents, computer use, and knowledge work, and then warns that those score gaps are a less reliable guide at this level. In the company’s own use, it says, the gap versus Fable 5.1 is narrower than the table suggests. A coding agent, here, is a model that takes a series of steps in a codebase, not only one reply. File both the lead claim and that caveat. The cells are in Sources.

The price, and what 40% means. Anthropic says Opus 5.5 needs less computing power to serve than Opus 5. A token is a small chunk of text the model reads or writes. You pay by the million of them. Input tokens, what you send, are $4 per million. Output tokens, what the model writes, are $20 per million. The page says those two rates are 20% less than Opus 5, which it lists at $5 and $25. The “about 40% less” line is a different number. Anthropic says that at default settings, on typical workloads, the bill is about 40% lower than Opus 5. Default settings means the effort level it ships unless you turn it up. The page says the model also uses fewer tokens per task, and that the lower rate plus fewer tokens is the 40% drop. Do not read 40% as the sticker cut. A cache read is a saved piece of a long prompt you do not pay full price for again. Those reads are $0.20 per million tokens, which the page says is 60% less than Opus 5. File every dollar figure as Anthropic’s. This desk did not receive an invoice.

Speed. Anthropic says Opus 5.5 writes output more than 30% faster than Opus 5. A separate fast mode, available in Claude Code and on the Claude Platform, runs at up to 2.5 times the speed. Claude Code is Anthropic’s coding tool. The Claude Platform is the service a developer calls. That faster mode costs more: $8 per million input tokens and $40 per million output tokens. “Up to 2.5 times” and “more than 30% faster” are the company’s claims. This desk did not time a reply.

The pacing sentence. Anthropic calls Opus 5.5 the first release since it publicly argued for pacing the frontier. On this page, pacing means slowing how fast models get more capable so safety work can stay ahead. The page says that argument came last week from chief executive Dario Amodei. Before release, the company says outside groups tested the model. It names Frontier Design and METR. An outside evaluator is a group asked to check the model before the public can use it. File the names as the page’s. This desk did not read their reports.

The behavior check, still the company’s. Anthropic runs an automated behavioral audit, a large set of simulated situations meant to catch a model acting against its instructions. The page calls that audit the most complete alignment test the company runs. Alignment means the model’s behavior stays in line with what people intend. Anthropic says Opus 5.5 is the strongest-performing model it has tested on that audit to date. It also says the model often seems to notice when it is being tested, which makes the audit a weaker picture of ordinary use. On biology and cybersecurity, the page says Opus 5.5 is comparable to Claude Mythos 5.1, so it ships with safeguards similar to those on Fable 5.1. A safeguard is a limit meant to stop misuse. Vetted organizations, groups the company has checked, can apply to a Life Sciences Verification Program to use Opus 5.5 for biology research. The company says a Cyber Verification Program will widen in the coming weeks. This filing does not describe how those checks work, and it is not a permit this desk issued.

Where it is on. Anthropic says Opus 5.5 is available now, including on Amazon Web Services, Google Cloud, and Microsoft Azure. On the Claude Platform the model’s name is claude-opus-5-5. Fast mode is also in Claude Code. Five-hour usage limits on the Pro, Max, and Team subscription plans are going up. The page does not print the new cap. Subscribers also get a rate-limit reset they can save and use later, whenever they choose. A rate limit is the cap on how much you can use in a stretch of time. Saving the reset means you can hold that fresh start instead of spending it immediately. File that as the page’s. This desk did not watch a counter move.

What is still a promise. Anthropic says Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks, with many of the same changes to performance, efficiency, and safety. Sonnet and Haiku are the company’s names for smaller models in the same family. “Coming weeks” is a plan. Those two models are not this launch.

Do not treat the score table as a ranking this desk ran. The page prints benchmark numbers. A benchmark is a fixed test used to compare models. Some comparison cells are not Anthropic’s own runs: the page says Zapier ran AutomationBench, and that the GPT-6 Astra and GPT-5.6 Sol figures are as OpenAI reported them. GDPval-AA is Artificial Analysis’s test of professional work. The scores on the page are the ones Anthropic prints. Early-tester stories on the page, including a long code migration and an overnight coding job, stay quotes. This desk did not rerun them. The table is in Sources.

Plain English for the rest of the card: token = a small chunk of text. input = what you send the model. output = what the model writes back. cache read = a saved chunk of a long prompt, billed cheaper the next time. default settings = the effort level Anthropic ships unless you raise it. Fable 5.1 = an earlier top Claude. Opus 5 = the previous Opus. pacing the frontier = slowing capability gains so safety work can keep up. alignment = behavior that stays in line with what people intend. safeguard = a limit meant to stop misuse. Claude Code = Anthropic’s coding tool. Claude Platform = the developer service, where this model’s name is claude-opus-5-5. rate limit = a cap on use in a time window. benchmark = a fixed comparison test. This filing is the 22 Sep company post. It is not an independent score.

PRIMARY here: Anthropic’s 22 Sep 2026 page, “Introducing Claude Opus 5.5,” at anthropic.com/claude-opus-5-5 — Tier A PRIMARY, the company’s own record. The first-in-the-5.5-family line, the Fable 5.1 “most work” claim, the narrower real-world gap versus that table, the less-compute line, the $4 and $20 sticker prices, the 20% sticker cut versus Opus 5 at $5 and $25, the about-40% workload drop at default settings, the $0.20 cache-read price, the more-than-30% faster output, the fast mode at up to 2.5 times for $8 and $40, the pacing sentence and the named outside evaluators Frontier Design and METR, the strongest-on-the-audit line, the notice-it-is-being-tested caveat, the Mythos 5.1 biology and cybersecurity comparison, the Fable 5.1-like safeguards, the Life Sciences Verification Program, the coming Cyber Verification Program, availability on Amazon Web Services, Google Cloud, and Microsoft Azure, the claude-opus-5-5 name, the higher Pro, Max, and Team five-hour limits, the saved rate-limit reset, and the Sonnet 5.5 and Haiku 5.5 “coming weeks” line are Anthropic’s. Zapier’s AutomationBench runs and OpenAI’s reported GPT figures stay third-party, as the page labels them. NOT claimed: an independent rerun, that 40% is the sticker-price cut, a new five-hour cap this desk measured, a timed speed test, that Sonnet 5.5 or Haiku 5.5 has shipped, a system-card finding this desk read as a second score, how a safeguard check works, a stock tip, or investment advice. Distinct from the already-filed deepseek-v4-1-flash, xiaomi-mimo-v2-6, and openai-frontier-standards.