Claude AI logo via Wikimedia Commons
A new model, and a new tier in the lineup
Anthropic released Claude Opus 5.5 on September 22, 2026, and it introduces something the company hadn't done before: a "5.5" family sitting between two existing models rather than replacing either of them. Opus 5.5 slots in above Opus 5 and below Fable 5.1, which remains Anthropic's flagship. That positioning matters more than it might sound. Instead of a straight upgrade that retires the older model, Anthropic is now offering a middle option aimed at a specific trade-off: performance close to the top model, at a meaningfully lower price and with faster responses.
For anyone who has been following model releases over the past couple of years, this is a familiar shape of announcement, but the specifics are new. Opus 5.5 is not a repackaged Opus 5 with a version bump. Anthropic says it performs at roughly the level of Fable 5.1 on most tasks, while costing 40% less to run than Opus 5 and generating output more than 30% faster. If that holds up in practice, it changes the calculus for a lot of teams that were choosing between "cheap and fast" and "capable but expensive."
What the cost and speed numbers mean day to day
Numbers like "40% cheaper" and "30% faster" are easy to skim past, so it's worth translating them into what they actually change for someone building on the model. A lower per-token cost means applications that call the model repeatedly, think coding assistants, document processors, or customer support bots, get materially cheaper to run at the same volume. That's the kind of change that lets a team either lower their own prices, run the model on tasks they'd previously reserved for a cheaper but weaker model, or simply extend their existing budget further.
Faster output has a different, more immediate effect: it shows up directly in how a product feels to use. A chat response that streams in noticeably quicker, or an agentic coding session that finishes a multi-step task sooner, changes the user's willingness to let the model work autonomously for longer stretches. Speed and cost together are why Anthropic is pitching Opus 5.5 less as a capability leap and more as an efficiency one. The ceiling on what's possible didn't move much. What moved is how much of that ceiling is now affordable to use routinely.
The benchmark that got the most attention
The specific example Anthropic has been highlighting is a 200,000-line code audit. According to the company, a review of that size used to take more than 20 hours and now takes under 3 hours with Opus 5.5. That's not a synthetic benchmark score; it's a description of a real class of task, auditing a large, unfamiliar codebase for issues, that plenty of engineering teams actually have to do and usually dread.
On the benchmarks Anthropic shared more broadly, Opus 5.5 surpassed Fable 5.1 in several areas: agentic coding, knowledge work, computer use, visual chart recognition, and multidisciplinary reasoning. Beating the more expensive flagship model on a subset of benchmarks while costing less and running faster is the core of Anthropic's pitch here. It doesn't mean Opus 5.5 is better than Fable 5.1 across the board, Anthropic's own framing is that it performs at "roughly" the same level on most work, but it does mean the gap between the mid-tier and top-tier model has narrowed in a way that's unusual for how these releases typically go.
Pricing, in plain numbers
Opus 5.5 is priced at $4 per million input tokens and $20 per million output tokens, a 20% reduction from what Opus 5 cost. Cached reads are priced at $0.20 per million tokens, a 60% decrease from before. For developers who aren't deep in token economics, the practical takeaway is this: reading text into the model (input) is cheap, generating text out of it (output) costs five times as much per token, and if your application repeatedly reuses the same context, like a long conversation history or a large reference document, cached reads make that reuse dramatically cheaper than treating every request as fresh input.
That caching discount in particular tends to matter more than it first appears. Applications that keep a large document, codebase, or knowledge base in context across many requests, rather than sending it fresh each time, benefit from the lower cache-read price on every subsequent call. Combined with the general 20% price cut, Opus 5.5 ends up meaningfully cheaper than Opus 5 for exactly the kind of long, context-heavy sessions that agentic coding and document analysis tend to involve.
Preserved thinking and stronger cyber capabilities, explained plainly
Two details in the release are worth understanding even though neither one is something an everyday user will directly interact with. The first is "preserved thinking," an anti-distillation safeguard that Anthropic first introduced with Fable 5.1 and has now carried over to Opus 5.5. In plain terms, distillation is a technique where someone trains a separate, smaller model to imitate a larger one by studying its outputs, including its intermediate reasoning steps. Preserved thinking is a measure aimed at making that kind of copying harder. It's a defensive, behind-the-scenes safety feature, not a setting users toggle or a change to how the model talks to you.
The second detail is that Opus 5.5 includes what Anthropic describes as the strongest cyber capabilities of any Anthropic model released so far. That's documented in a 230-page system card, the technical report Anthropic publishes alongside major releases to lay out a model's capabilities and risks in detail. Stronger cyber capabilities cut both ways: the same skills that help a security researcher find and fix a vulnerability can, in principle, help someone look for one to exploit. This is a real and specifically flagged capability increase, not a marketing line, but it's also not a controversy in the way some other AI safety stories have been. Anthropic's response has been to document it thoroughly and restrict the more sensitive research uses to vetted programs, which is covered in the next section.
Who's using it for coding
Coding is where Opus 5.5 is getting the most real-world use, and the use cases go well beyond autocomplete-style suggestions. Teams are pointing it at long-running, agentic work: building out a feature across multiple files, debugging an issue that spans several services, or refactoring a section of a codebase without breaking what depends on it. The kind of work that used to require a developer to hold a lot of context in their head is now something the model can carry across a longer session.
Large code migrations are another common case, moving a codebase from one framework or language version to another, where the model needs to understand both the old patterns and the new ones and apply that understanding consistently across thousands of files. Multi-repository debugging follows a similar shape: tracing a bug that only shows up when two or more codebases interact, which requires holding context across boundaries that used to sit outside a single developer's usual view. The 200,000-line audit example fits squarely into this category, and it's a useful proxy for how much faster this class of work is going now that a model can be reasonably trusted to do a first, thorough pass on its own.
Who's using it for business and knowledge work
Outside of engineering, the use cases cluster around document-heavy work. Business and operations teams are using Opus 5.5 to analyze large sets of internal documents, contracts, reports, policy files, and pull out what's relevant without someone reading every page manually. Drafting structured reports is another common pattern: giving the model source material and a format, and having it produce a first draft that follows a consistent structure rather than starting from a blank page each time.
Comparing requirements against each other, say, checking a proposal against a set of stated criteria, or reconciling two versions of a specification, is the kind of task that's tedious for a person and well suited to a model that can hold both documents in context at once. Preparing implementation plans, taking a decision or a set of requirements and turning it into a concrete, ordered plan for a team to execute, rounds out the picture. None of this is new in kind. What's changed is that the lower cost makes it practical to run these tasks routinely rather than saving them for occasions that justify the expense.
Who's using it for research, including the gated programs
Research use cases split into two groups: general scientific research that's open to anyone building on the model, and two specific areas where Anthropic has put access controls in place because the capabilities involved are more sensitive. Anthropic runs a Life Sciences Verification Program that lets vetted organizations use Opus 5.5 for biology research, and separately allows verified cybersecurity practitioners to use it for security work.
The word "verification" here is doing real work. It means access to the model's full capabilities in these two domains isn't open by default, an organization or practitioner has to go through a vetting process first. That's a direct reflection of the stronger cyber capabilities mentioned in the system card, and of the fact that biology research tools can, in the wrong hands, assist with things far more dangerous than the vast majority of intended use cases. Gating access rather than restricting the model outright is Anthropic's way of trying to keep the tool available to the researchers who need it while adding a checkpoint against misuse. It's a narrower and more specific measure than a blanket policy, and it only applies to these two domains.
Long tool-using and agentic workflows
A theme that runs through most of the use cases above is that they're not single-turn interactions. Opus 5.5 is being used for workflows where the model calls tools, reads results, decides what to do next, and keeps going across many steps, sometimes without a person checking in after every action. Financial models that pull data, run calculations, and adjust based on results. Sourced research briefings that involve searching, reading, and synthesizing across multiple documents or sites. Computer-use tasks where the model operates software the way a person would, clicking through an interface rather than calling an API.
This is where the speed and cost improvements compound. A workflow that runs 20 tool calls in sequence pays the token cost and the latency of each one. A model that's both cheaper per token and faster to generate output makes that kind of extended, autonomous session more practical to run regularly rather than as an occasional experiment. It's also where the coding benchmark gains matter most in practice, since agentic coding is itself a long tool-using workflow: read a file, make an edit, run a test, read the result, repeat.
Part of a broader pattern in 2026
Opus 5.5 fits a pattern that's become more visible across the AI industry this year: models converging on similar capability tiers while competing harder on cost and speed than on raw benchmark scores. Rather than every release chasing a new capability ceiling, several labs have been shipping mid-tier models that close most of the gap to their flagship while cutting the price sharply. Anthropic's own framing, that Opus 5.5 performs at roughly Fable 5.1's level for a fraction of the cost, is a clear example of that shift.
For people actually building products, this matters more than another headline benchmark number would. It's the difference between a capability being demonstrated in a lab and being affordable enough to build into something people use every day. Anthropic is positioning Opus 5.5 as the strongest default starting point in its current lineup for complex coding, research, writing, and tool-using work, not because it's the most capable model the company has ever released, but because it's the one most people will actually be able to run at the scale their work requires.
Frequently asked questions
Is Opus 5.5 replacing Opus 5 or Fable 5.1?
No. It sits between the two as a new mid-tier option. Opus 5 and Fable 5.1 remain part of Anthropic's lineup alongside it.
Is Opus 5.5 more capable than Fable 5.1?
Not across the board. Anthropic says it performs at roughly Fable 5.1's level on most work, and actually surpassed it on the specific benchmarks the company shared: agentic coding, knowledge work, computer use, visual chart recognition, and multidisciplinary reasoning.
How much cheaper is Opus 5.5 than Opus 5?
Anthropic says it costs 40% less to run than Opus 5 overall, with list pricing of $4 per million input tokens and $20 per million output tokens, a 20% cut from Opus 5's pricing. Cached reads are $0.20 per million tokens, 60% cheaper than before.
What is "preserved thinking"?
It's an anti-distillation safeguard, first used with Fable 5.1, meant to make it harder for a third party to train a copycat model by studying Opus 5.5's outputs and reasoning steps. It's a background safety measure, not a user-facing feature.
Can anyone use Opus 5.5 for biology or cybersecurity research?
General use of the model is open, but Anthropic's Life Sciences Verification Program and its cybersecurity verification process gate the more sensitive research applications in those two areas to vetted organizations and practitioners.
What's the single most-cited example of its coding ability?
Anthropic's own example is a 200,000-line code audit that used to take more than 20 hours and now takes under 3 hours with Opus 5.5.