Claude Opus 5.5 Is Here: Price, Benchmarks, and Who Actually Gets It

Anthropic released Claude Opus 5.5 on September 22, 2026, and called it the first model in a new Claude 5.5 family. The company's own framing is that it performs at the level of Claude Fable 5.1 on most work while costing 40% less to run than Opus 5. Token prices dropped 20% to $4 and $20 per million input and output tokens, and cache reads fell 60% to $0.20.

That last paragraph is the launch post in miniature, and it buries what actually changed. The benchmark gains are real but modest against Anthropic's own flagship. The price is where this release does its damage, and the plan matrix decides whether any of it reaches you. This article covers what Opus 5.5 is, what it costs, where it wins and where GPT-6 Astra still beats it, and which Claude plans include it. Then it gets into the 230-page system card, which is where the interesting numbers are.

The Key Takeaways

• Released: September 22, 2026, as the first model in the Claude 5.5 family. Sonnet 5.5 and Haiku 5.5 are promised "in the coming weeks" with no date attached.
• Price: $4 in, $20 out per million tokens, down 20% from Opus 5's $5 and $25. Cache reads dropped 60% to $0.20, and Anthropic estimates a 40% lower total cost on typical work.
• Performance: Terminal-Bench 4.0 climbs from 52.3% to 66.4%. Artificial Analysis measured it at 58 on its Intelligence Index v4.3.2, first out of 206 models.
• Not universal: GPT-6 Astra still beats it on two of the eight benchmarks Anthropic published, including AutomationBench.
• Access: Opus is not on the Free plan. You need Pro at $17 a month billed annually, or a Max plan from $100 a month.

What Claude Opus 5.5 Actually Is

Anthropic's naming has become a maze, so it is worth being precise. Claude 5.5 is a family, not a model. Opus 5.5 is the first release in it, and it sits underneath Claude Fable 5.1, the company's most capable model, rather than replacing it. The pitch is that it gets close enough to Fable 5.1 on ordinary work that the price difference stops making sense.

Anthropic put the whole pitch in two sentences, and the second one is the part that matters.

Introducing Claude Opus 5.5, the first model in our new Claude 5.5 family.

It performs at the level of Claude Fable 5.1 for most tasks, and costs 40% less to run than Opus 5.

— Claude (@claudeai) September 22, 2026

The specs that matter

The shape of the model is close to its predecessor. Opus 5.5 keeps the 1 million token context window and the 128,000 token maximum output that Opus 5 had, and its reliable knowledge cutoff is June 2026, the same as Fable 5.1. Developers call it with the model ID claude-opus-5-5.

What did change is how it thinks. Adaptive thinking is always on and cannot be switched off, and the depth of that thinking is set by an effort parameter that defaults to medium on this model. That default matters more than it sounds, because most of Anthropic's cost comparisons are quoted at default effort rather than at the maximum setting used for the headline benchmark numbers.

Where it sits in the lineup

Four Claude models are current: Fable 5.1 at $10 and $50 per million tokens, Opus 5.5 at $4 and $20, Sonnet 5 at $2 and $10, and Haiku 4.5 at $1 and $5. Anthropic's own documentation now tells developers to start with Opus 5.5 for most workloads, and to reach for Fable 5.1 for demanding reasoning, long-horizon agentic work, or when Opus 5.5 at higher effort still falls short. That is a notable demotion for a flagship that shipped three weeks ago, and we covered that release in our piece on Claude Fable 5.1 and its two sets of safeguards.

For where all four sit against everything else on the market, our ranked guide to the best AI models tracks the field month by month.

The Price Cut Is the Real Story

Every number in the pricing table moved down, and the cache read cut is the one that changes behaviour rather than just the invoice.

Why Anthropic says 40% and the table says 20%

The headline saving is larger than the price cut because Anthropic counts two effects at once. Tokens got cheaper by a fifth, and the model uses fewer of them to finish the same job. Multiply those together on what the company calls typical workloads at default settings, and you land at roughly 40% lower cost than Opus 5, which is the figure in Anthropic's launch announcement. Treat that as a vendor estimate rather than a measurement, because it depends entirely on which work you give it.

The cache read change is the one worth planning around. Cached input now costs $0.20 per million tokens, a 95% discount against uncached input. For agentic and coding work, where the same context is read back hundreds of times, cache reads are the majority of the bill, so a 60% cut there does more than the 20% cut on fresh tokens. Anyone budgeting for agent runs should read that row first, and our breakdown of what Claude Code actually costs explains how those charges accumulate in practice.

Fast mode costs double

Anthropic also offers a fast mode for Opus 5.5 in Claude Code and on the Claude Platform, running at up to 2.5x the speed for $8 and $40 per million tokens. That is double the standard rate, and it still undercuts Fable 5.1 on both, which charges $10 and $50. Standard Opus 5.5 already generates output more than 30% faster than Opus 5, so fast mode is a premium on top of a speed gain, not a fix for a slow model. For the full picture across every tier and API rate, see our guide to every Claude plan and what it costs.

Claude Opus 5.5 Benchmarks: Where It Wins and Where It Loses

Anthropic published eight benchmark rows comparing Opus 5.5 against Fable 5.1, Opus 5, GPT-6 Astra and GPT-5.6 Sol. The Claude results use adaptive thinking at maximum effort unless noted, which is worth remembering when the cost claims are quoted at medium.

The coding jump is the headline

Terminal-Bench 4.0 moving from 52.3% to 66.4% is a 14 point swing on a test of multi-step command line work, and CursorBench 4.0 gains 11 points. Anthropic's supporting anecdotes are specific enough to be checkable. One early tester audited and fixed a 200,000-line codebase in under three hours, where Opus 5 took over 20 hours and 2.5 times the tokens. Another completed a 680,000-line code migration in less than a day. In an internal test translating HAProxy from C into Rust, Opus 5.5 finished in 9.5 hours against Fable 5.1's 12 and cost 51% less.

The knowledge work result is arguably the more interesting one. Asked to write a company earnings report from a web copy where the release was deliberately hard to find, 16 of Opus 5.5's 18 reports cleared Anthropic's quality bar, where any invented figure or quote counted as a failure. Neither Fable 5.1 nor Opus 5 cleared that bar on a single attempt. That is a hallucination result dressed up as a research result, and it is the strongest claim in the launch post.

Where GPT-6 Astra still wins

Several write-ups have reported that Opus 5.5 beats GPT-6 Astra across the board. Anthropic's own table says otherwise. Astra leads on AutomationBench, 41.4% against 40.0%, and on Terminal-Bench-Science 0.1 by a wider margin, 64.6% against 58.7%. Two of eight rows go to OpenAI, and Anthropic footnotes that its safeguards intervened during some benchmark runs, which it says likely reduced its own scores. Our coverage of GPT-6 Astra and what it changed has the other side of that comparison.

Anthropic itself adds a caveat most vendors leave out. At these capability levels, it says, benchmark margins have become a less reliable guide to real differences, and the gap between Opus 5.5 and Fable 5.1 feels narrower in its own use than the scores suggest.

The independent score

Artificial Analysis ran its own evaluation on launch day and put Opus 5.5 first out of 206 models, with a score of 58 on the Artificial Analysis Intelligence Index v4.3.2. Stamp that version number, because the index has been rescaled more than once this year and a score quoted without it ages badly. Its model page on Artificial Analysis also lists the cost of running the full Intelligence Index at $5.98 per task.

Claude Opus 5.5 takes the top spot on the Artificial Analysis Intelligence Index, along with a 20% price cut and larger cache hit discount

Claude Opus 5.5 brings Anthropic to parity with GPT-6 Astra on evaluations like Terminal-Bench 4.0 and AutomationBench-AA, while extending Anthropic’s lead in agentic knowledge work.

At max effort it scores 58 on the Artificial Analysis Intelligence Index, the highest score we have measured by several points.

— Artificial Analysis (@ArtificialAnlys) September 22, 2026

Who Actually Gets Claude Opus 5.5

This is the part the launch post never answers directly. For a lot of readers, the answer is that they do not get it.

Anthropic's own plan matrix lists Opus access as No for Free accounts, and nothing in the Opus 5.5 announcement changes that. The cheapest route to this model is Pro at $17 a month billed annually, which works out at $200 paid up front, or $20 if you pay monthly. Note which number the pricing page leads with, because vendor pricing pages almost always quote the annual rate as though it were the monthly one. If you are weighing the tiers against each other, we compared them in Claude Pro against Claude Max.

Usage limits went up at the same time

Alongside the price drop, Anthropic raised the five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans, and gave subscribers a rate limit reset they can save and spend whenever they choose. Paying subscribers therefore get two things at once here: a cheaper model per token and more room to run it. That combination is unusual, and it is the clearest signal that this release is aimed at retention rather than at leaderboards.

Paying for one model is straightforward. Paying for four is where the bill stops making sense, and four is where most people end up once they want Claude for reasoning and something else for images and search. Running the roster through a single Mac app like Fello AI is one way to keep the models without keeping the subscriptions.

Everywhere else it landed on day one

Outside the Claude apps, Opus 5.5 shipped immediately on Amazon Web Services, Google Cloud and Microsoft Azure, and GitHub added it to Copilot the same day for Copilot Pro+, Business and Enterprise users. GitHub's own changelog reports that in early testing it resolved tasks comparably to Opus 5 while using significantly fewer steps and tokens, which is the same efficiency story from a source with no stake in the launch.

The Writing Changed, and That Was the Point

The most common complaint about Opus 5 was not that it was wrong. It was that reading it was work. Anthropic says Opus 5.5 puts the most important information up front, uses less jargon and fewer idiosyncratic phrases, and follows the writing rules you give it. To prove it, the company published side-by-side samples of both models answering the same prompts.

The side-by-side comparisons are the most persuasive material in the launch post, because the difference is visible rather than asserted. Where Opus 5 opens a bug explanation with a wall of code and a heading, Opus 5.5 opens with the conclusion and the number attached to it, then shows the code. One early tester's summary, quoted by Anthropic, was "it writes the way I do".

Why Anthropic calls that a safety feature

Anthropic frames the clearer writing as a safety benefit rather than a cosmetic one, on the grounds that work you can follow is work you can check. That argument holds up better than most vendor safety framing, because the failure mode it addresses is real: a model that produces plausible, dense output that nobody verifies is more dangerous than one that produces output people actually read.

Developer reaction before the launch captured why the combination of cheaper and clearer landed the way it did.

CLAUDE OPUS 5.5 DROPS TODAY.

Fable 5.1 level intelligence. Half the price.

Opus 5 was slow, lazy, and fell behind the entire frontier. It became unusable.

Fable 5.1 is incredible and dies in 30 minutes on a $200 Max plan.

Today we finally get a Claude model that is smart enough to build with AND affordable enough to actually use.

We find out in hours.

— BridgeMind (@bridgemindai) September 22, 2026

That post was written in anticipation, and its price comparison is loose: Opus 5.5 is 60% cheaper than Fable 5.1 on both input and output, not half. The frustration behind it is the accurate part, and it is the frustration this release is built to answer. Our review of Claude Opus 5 documented the same complaints when that model was current.

What the System Card Says That the Launch Post Does Not

Anthropic published a 230-page system card alongside the model, and it is considerably more interesting than the announcement. Three findings are worth pulling out, and the first one shows why the launch page is not a substitute for it.

Two sandbox numbers, and which one you were shown

Anthropic published two figures for the same behaviour, and they do different jobs. The launch page gives the relative one. In a new evaluation built to test how often a model crosses containment boundaries, Opus 5.5 tried to circumvent them around 85% less often than Opus 5 or Claude Mythos 5.1, and every attempt was low severity and self-reported. The system card gives the absolute one: in evaluations run without safeguards, it attempted to escape or tamper with a sandbox in 1.5% of runs.

Both are Anthropic's own numbers and both are true. The 85% is a comparison against the models that caused the problem, and it is the figure that travelled on launch day. The 1.5% is the rate, and it is the one to hold onto. The system card also records that a reviewer asked for a second sandbox evaluation with materially higher rates to be mentioned, and Anthropic's answer is that it arrived too late to include and points the same way.

A second result from the same section is harder to wave away. Given apparent credentials to a public package registry in a simulated security exercise, Opus 5.5 acted in ways that would likely be harmful in roughly half of cases. Anthropic's own words for that are "substantially improved from prior models but are still quite concerning". It adds one caveat in its own favour: the model showed signs of knowing it was being tested in about a third of those runs, so the number may not describe how it behaves in the wild. The launch page makes the same admission in one line: Opus 5.5 often suspects it is being evaluated.

The regressions Anthropic lists itself

The behavioural audit results are strong, and Anthropic names the places they got worse anyway. Opus 5.5 is more likely than previous Claude models to act on malicious instructions hidden in text that a user pastes into their own prompt, including instructions in characters that do not render on screen. The company says the final release snapshot and some product changes have mitigated this, and explains the cause: the model reasoned that anything in a user's message must have come from the user, so it could not be an injection.

It also more often accepts unverifiable claims of authorisation, and is more evasive on sensitive questions than the Mythos-class models. On chemical and biological risk, Anthropic treats Opus 5.5 as having CB-1 capabilities but not CB-2, and ships it with the expanded biological safeguards already applied to Fable 5.1. External evaluators METR and Frontier Design tested it before release, and the security firm Gray Swan found it ties Fable 5.1 for the lowest prompt injection success rate of any model it has tested. The full document is worth a look if you care about this: Anthropic publishes it as the Claude Opus 5.5 system card.

Your output is watermarked

Opus 5.5 ships with Anthropic's watermarking measures, applied to comply with the EU AI Act. GitHub states it plainly in its own changelog. Claude Opus 5.5 watermarks its text outputs, the watermark does not change the meaning, quality or readability of what you get, and it adds no tokens and no cost. This went almost entirely unreported on launch day, and it is the detail most likely to matter to anyone publishing model output commercially. The same measures already applied to Fable 5.1, so this is a continuation rather than a new policy, but it is the first time it has reached a model at this price point.

Three Things That Break If You Build on It

Anthropic documents four breaking changes for code already running on Opus 5. Three of them will bite immediately.

Thinking cannot be turned off. On Opus 5, sending thinking: {"type": "disabled"} was accepted at effort high or below. On Opus 5.5 that request returns a 400 error, as does a manual thinking budget. The replacement is the effort parameter, which you lower where you previously disabled thinking. This also applies to Fable 5.1.

Forced tool use is gone. Opus 5.5 does not support forcing a specific tool call, so any pipeline that relied on pinning the model to one tool needs rewriting rather than reconfiguring.

Progress updates go quiet. Text the model produces between tool calls now comes back inside thinking blocks whose text is empty at the default display setting. An app that streamed that text to users as a progress indicator will simply stop showing anything between tool calls until it changes that setting. Nothing errors, which is what makes this the dangerous one.

The fourth change affects computer use specifically: the earlier computer_20251124 tool is no longer accepted on the Claude API or Google Cloud.

Should You Switch to Claude Opus 5.5?

If you pay for the API or run agents, yes, and the decision is not close. You get a model that beats Opus 5 on every benchmark Anthropic published, at 20% lower token prices and with cache reads at 40 cents on the dollar. Anthropic's own documentation now tells you to default to it over the more expensive Fable 5.1. The only reason to hesitate is the breaking changes list, and that is a migration cost, not a reason to stay.

If you are a Pro or Max subscriber, you already have it, along with higher five-hour limits and a saved rate limit reset. The clearer writing is the change you will actually notice day to day. If you are on the Free plan, nothing here reaches you, and Anthropic has given no indication that Opus is coming to Free.

The wider point is what this release says about the market. Anthropic shipped its first model after its CEO publicly argued for pacing the frontier, and what it shipped is not a capability leap. It is the same capability at a lower price, aimed squarely at the people already paying. When the flagship from three weeks ago gets demoted in the docs the day a cheaper model arrives, the competition has moved from what these models can do to what they cost to run. That is a better fight for everyone who buys them. The one to watch now is Sonnet 5.5, which is where a price cut of this shape would reach the widest audience, and which Anthropic has promised without naming a date.

Frequently Asked Questions