AI Tool ScoutREVIEWS & FREE TOOLS

Claude Opus 5.5 Review (2026): Is AA 58 Worth $10 a Task?

The first Opus that costs less than the one it replaces, second on the independent index, first on agentic coding — and the first Claude whose generated text carries a mark you cannot see, remove or check. Here is the price, the real per-task bill, and the watermark landing on four older models this week.

By AI Tool Scout · Published November 2, 2026 · Pricing checked November 2, 2026

What Is Claude Opus 5.5?

Claude Opus 5.5 is Anthropic's current general-availability flagship, released September 22, 2026 as the first model in our new Claude 5.5 family. It replaces Claude Opus 5 at the top of the Opus line, which sits below Claude Fable 5.1 on price and now roughly level with it on measured capability. Anthropic's framing is unusually modest for a launch: Opus 5.5 performs at the level of Claude Fable 5.1 on most work, and the post concedes that benchmark margins have become a less reliable guide to real-world differences. That sentence is the most useful thing in the announcement, and it is why this review weights per-task cost as heavily as index score.

What it is not: not multimodal in the wide sense, not open weight, not a new architecture, and not a model whose size anyone outside Anthropic knows.

The specification, as published

Two rows matter more than they look. Retirement not sooner than September 22, 2027 is a one-year floor — short by enterprise standards, long enough to plan a migration, and stated on the model page rather than in a blog post. And a June 2026 cutoff is four months before launch, which is now normal at the frontier and means anything newer needs retrieval, not a bigger window.

What changed against Opus 5 and Opus 4.8

Four of those differences are breaking changes, documented on the model page rather than buried: thinking cannot be disabled; forced tool use returns an error; thinking blocks are tied to the model and the conversation; and the earlier computer_20251124 computer-use tool is rejected on the Claude API and Google-managed Google Cloud. A fifth breaks output silently — text between tool calls arrives in thinking blocks whose text is empty at the default display setting, so any product streaming that text as progress updates goes quiet between tool calls. That one does not throw and does not log. It gets reported as the UI is broken by someone who is right.

What changed on November 3: the watermark reaches the older models

This is the week's news and it is easy to misread, so the order matters. Claude Opus 5.5 already carries Anthropic's text watermark — it launched after August 2, 2026, the date from which Anthropic marks new models' output from day one, and the help-centre table lists it as supported on Anthropic's own surfaces and through cloud partners. Nothing about Opus 5.5's output changed this week. What changes is that the mark is being retrofitted to older models.

In an email to Claude Platform customers, Anthropic said it will begin applying the watermark on November 3, 2026 to outputs from Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6, with Amazon Bedrock expected to follow a few days later. The driver is the EU AI Act: Anthropic signed the Article 50(2) Code of Practice, the obligation applied from August 2, 2026, and the grandfathering deadline for systems already on the market is December 2, 2026. Anthropic says the marking applies worldwide, not only in Europe.

What it does to text you generate. The mark is a statistical bias in word choice during sampling — a variant of Google DeepMind's SynthID-Text — not a hidden character, metadata or a footer. Because it lives in the words, it travels with copy and paste into your CMS, your email and your client's deck. Anthropic says it adds no characters, no tokens and no cost, and carries nothing about the user, the organisation or the conversation. It is applied at the model level, so no system prompt and no API parameter removes it. Copied text keeps the mark, light editing usually still detects, a full rewrite loses it, and a Claude translation carries it fully. Code is marked less, because exact tokens are untouched. Short passages are the weak case: the signal needs tokens to accumulate.

Quality and plagiarism. Anthropic says meaning, quality and readability are unchanged, and that is the claim to be skeptical about: critics argue no two synonyms mean exactly the same thing, so choosing a word by a key instead of by fit must cost some precision. Nobody has published a measurement either way. On detection the answer is cleaner and less useful: no public detector has shipped, the mark needs Anthropic's key, and a detected mark means text may have been processed by Claude — which includes copy a human wrote and Claude proofread. A mark is evidence of contact, not authorship; absence proves nothing.

The loophole. On Opus 4.6 and Sonnet 4.6 the watermark applies only when the three sampling parameters are left at their defaults; the same caveat is not attached to Mythos 5 or Opus 4.7. That lands on Opus 4.6 specifically, which Anthropic engineers have described as the company's best writing model — ahead of Opus 5.5 on prose. It is also a gap Anthropic can close without notice.

Pricing: $4 In, $0.20 Cached, $20 Out

The price cut is the most defensible thing about the release, because it is money rather than a benchmark. Opus 5 charged $5 / $0.50 / $25; Opus 5.5 charges $4 / $0.20 / $20. Input and output are 20% cheaper, cache reads 60% cheaper, and cache writes fall from $6.25 to $5 and from $10 to $8. Anthropic also says output is more than 30% faster than Opus 5, which is why it can cut the card and still claim a better margin.

The worked example: 500M input, 80% cached, 50M output

Same modelled month we use across this site, so the numbers compare with our Mistral Large 4 and DeepSeek V4.1 Flash reviews.

Three conclusions. The output leg is 68% of the bill — $1,000 of $1,480 — so verbosity control beats input compression. Caching is worth more than the price cut: $1,520 against $470, which makes the engineering decision bigger than the procurement one. And the "40% cheaper" headline is not a price change: it is Anthropic's estimate of total cost on a typical workload, folding the 20% cut together with fewer tokens per task. You see 40% only if your token count falls, and it falls only if the new default effort suits your work.

The Effort Dial Is the Real Price

Five settings, each measured by Artificial Analysis on the same harness. This table converts a marketing claim into a curve you can budget against.

All Artificial Analysis, read November 2, 2026 at its current index revision. Cost per task is the weighted average across the nine index evaluations.

Read the top of the curve carefully. high to xhigh buys 5 index points and 3 points of Terminal-Bench for 75% more money and 3.7× the wait. xhigh to max buys one index point, adds nothing measurable on Terminal-Bench, costs 66% more and puts first-token latency at twelve minutes. That is not a criticism of the model; it is what the top of any capability curve costs. It is a criticism of any evaluation run at max and any deployment that inherited the setting.

Then there is the default. Opus 5.5 defaults to medium; Opus 5 defaulted to high. One word explains most of the "40% cheaper" claim and most of the "it got worse" complaints at once. A team that re-ran its July evaluation in October without pinning effort did not measure a regression in the model; it measured a regression in the default.

What the Third Party Found

Artificial Analysis assessed the model on launch day and has re-scored it twice since, once when the index moved to v5 on October 27. The current reading is 58 at max effort, ranked second of 58 models; Gemini 4 Carbon took first on October 28 at 59.

The honest reading of 58: a genuine frontier score with a thin margin. One point separates it from Gemini 4 Carbon and the index is a weighted average of nine evaluations that were re-weighted five days before this review, so a one-point gap is not a purchase criterion. The stronger evidence is at workload level: first on the Coding Agent Index, 1,801 Elo on AA-Briefcase, 67.9% on HLE-Diamond — the pattern of a model tuned for long agentic runs rather than trivia.

The verbosity figure is the most under-reported number in the release. 218 million output tokens against a 70 million median, at 173,000 output and 120,000 reasoning tokens per task, means an average task's output leg is about $3.46 before any input is billed. That is how a $4 / $20 card produces a $10.04 task: the rate is fine and the model talks a lot. Measure tokens per task before you celebrate the cut.

What Anthropic reports about itself

Vendor-reported throughout — measured by Anthropic, or by a partner on a harness Anthropic does not control. None of it can be reproduced from outside today.

Opus 5.5 wins most rows and loses two — AutomationBench by 1.4 points and Terminal-Bench-Science by 5.9 — and both losses sit on harnesses with wide error bars or third-party control. The most defensible claim is 66.4% on Terminal-Bench 4.0, because the benchmark is public and the vendor publishes its own standard error. The least defensible is 81.8% on OSWorld, because it is a partial score and the strict number is 48.7%. Quote the one you mean.

Anthropic also does something rare: it states that production safeguards were enabled during evaluation, that interventions handed cybersecurity tasks to Opus 4.8 and biology work to Opus 5, and that this likely reduces Claude Opus 5.5's performance. A vendor publishing a footnote that makes its own numbers look worse is telling you where the body is buried — and it means 66.4% is a floor for a guarded model, not a ceiling.

Where It Wins

Agentic coding, and it is not close on the independent measure. Opus 5.5 holds the top of Artificial Analysis's Coding Agent Index, in the workflow that actually bills: long runs with tool calls, where a failed task costs the whole trajectory. Anthropic's own framing is that at default effort Opus 5.5 beats Opus 5 at max effort for about a fifth of the cost on Terminal-Bench 4.0, and the effort table is consistent with it — medium at $2.45 a task scores 52.5% against the previous generation's ceiling. One early-access quote in the launch post makes the same point from the customer side: even at its lowest effort setting, Claude Opus 5.5 caught 72% of known bugs in our code reviews to Opus 5's 56% at high effort, with fewer false alarms and a fraction of the output. That is a testimonial, not a measurement, but it matches the independent data.

Cost per unit of work, once you stop paying for the ceiling. A 4.1× spread between medium at $2.45 and max at $10.04 means Opus 5.5 can be the most expensive model on your shortlist and the cheapest, depending on one string in the request body — low for routing, medium for extraction, high for customer-facing generation, max for the agent that runs overnight. Operational boringness is the third win: a June 2026 cutoff, no retirement before September 2027, and first-party availability on Bedrock, both Google Cloud managed modes, Microsoft Foundry, the Claude Platform on AWS and the Claude API on day one.

Where It Loses

The ceiling is slow in a way that changes product design. At max, Artificial Analysis measures 729.6 seconds to the first answer token and 1,140 seconds per task — twelve minutes before the user sees anything. Short prompts are faster and low reaches 6.2 seconds, but a product that puts this model at max behind a synchronous request needs a queue, a job ID and a progress channel, not a spinner.

Verbosity is the hidden line item. 173,000 output and 120,000 reasoning tokens per task, 218 million to run the index. A model that talks twice as much at half the rate is not cheaper, and the output leg is 68% of our modelled bill. Second place is still second place — Gemini 4 Carbon took the top of the index on October 28 and has held it, and if a composite score is all you need, you are buying the wrong model this week. The counter-argument is narrow but real: Carbon had no public API access at launch and is not on the Coding Agent Index, and a model you cannot call is not a competitor for a production workload.

The fingerprint. Every token carries a statistical mark you cannot inspect, remove or verify without Anthropic's key. For most buyers that is irrelevant. For three categories it is not: publishers whose contracts require disclosure and who now cannot tell whether a draft trips a detector; agencies delivering copy that clients may run checks on; and anyone whose product promise is undetectable anything. No free tier, and a plan table that keeps moving is the last one — Opus is on Pro, Max, Ultra, Team and Enterprise, the Free plan carries Sonnet and Haiku, and the tier names and prices have shifted repeatedly through 2026, so check claude.com/pricing rather than a review from September.

Pros & Cons

5 Best Claude Opus 5.5 Alternatives

The useful question is not whether Opus 5.5 is good — it is the best agentic coder on sale — but whether it is right for this workload at this setting. These five cover the cases where it is not.

Comparison Table

Opus 5.5, Opus 5 and Fable 5.1 rates are from Anthropic's list price card, checked November 2, 2026. In the AA column, 58 and 59 are read from Artificial Analysis's own pages on November 2; the other scores come from third-party leaderboards mirroring the same index between October 28 and November 1, where a revision can move a model a point or two. Competitor token prices are carried from our October reviews and were not re-pulled on November 2 — read the "not re-checked" cells as blanks, not figures.

What the 1M Context Actually Costs

Anthropic publishes one rate for all context lengths, so there is no long-context surcharge to dodge — a real improvement on the era when a million-token window came with doubled input rates above 200K. That does not make a million tokens cheap. It makes the cost a function of how many times you send them.

Then the output side, which is where the money actually goes. At max effort the harness measures 173,000 output and 120,000 reasoning tokens per task; at $20 a million that is roughly $3.46 of output per task before input. Put together: 1,000 tasks a month at max effort is about $10,040; the same 1,000 at the default medium is about $2,450, for 48 index points instead of 58. That is the most useful budget decision in this review and it has nothing to do with the token card.

The window has a second cost that is not on the invoice: attention degrades in the middle of very long prompts, and Anthropic has published no evaluation at 100K, 400K or 1M tokens. The only long-context figures we found are second-hand — GraphWalks breadth-first search at 90.6% below 128K falling to 66.8% between 256K and 1M — and we could not trace them to a primary source. Test your own needlestack before you architect around a million tokens. A window is a capacity, not a capability.

What We Could Not Verify

Nine things a buyer would want are missing or unconfirmed, listed here rather than smoothed over.

One paragraph on the corporate story, labelled as reporting rather than anything about the model. Anthropic filed its Form S-1 on October 26, 2026 to list Class A shares on Nasdaq under the ticker ANT, with press reports putting the raise near $100 billion, the valuation between $1.5 trillion and $2 trillion, marketing from November 9 and a listing targeted for the week of November 16; the filing leaves price, share count and valuation blank. The disclosed numbers that matter to a buyer are about capacity rather than profit: $16.3 billion of first-half revenue, a $693 million GAAP operating loss, $3.9 billion of positive operating cash flow, and $542–567 billion of committed compute and infrastructure spending, much of it payable whether or not customers keep buying. None of that changes what Opus 5.5 costs you today and no published document commits Anthropic to a price after 2026. But an issuer that has promised half a trillion dollars of compute has a strong incentive to keep list prices where they are and an equally strong incentive to find more revenue per token — and a 95% cache discount is exactly the kind of feature that raises utilisation without raising list prices.

Frequently Asked Questions

How much does Claude Opus 5.5 cost?

On the Claude API it is $4 per million input tokens, $20 per million output tokens, $0.20 per million cache reads, $5 per million five-minute cache writes and $8 per million one-hour cache writes, checked November 2, 2026 against Anthropic's list price card. That is 20% below Claude Opus 5 on both token legs and 60% below it on cache reads. The Batch API is half price at $2 and $10, US-only inference is 1.1x at $4.40 and $22, and Fast mode is $8 and $40 for up to 2.5x the speed. A modelled month of 500 million input tokens with 80% served from cache plus 50 million output tokens comes to $1,480: $80 for the 400 million cached tokens, $400 for the 100 million fresh tokens and $1,000 for the output. The same month without caching is $3,000, and the same token volume on Opus 5 is $1,950.

Is Claude Opus 5.5 available on the free plan?

No. Anthropic's plan table puts Opus on Pro, Max, Ultra, Team and Enterprise and leaves the Free plan with Sonnet and Haiku. The cheapest route to Opus 5.5 in the Claude apps is Pro at $20 a month, or $17 a month billed annually as a $200 upfront charge. Max is $100 a month, Ultra is $200 a month with no five-hour limit and 10x Pro's weekly usage, and Team and Enterprise are priced per seat. At launch Anthropic also raised the five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans and added a one-time rate-limit reset that subscribers can save and spend when they choose. On the API there is no free tier for this model.

Is Claude Opus 5.5 watermarked, and what changes on November 3?

Claude Opus 5.5 already carries Anthropic's text watermark, because it launched after August 2, 2026, the date from which Anthropic marks output from new models from day one. On November 3, 2026 Anthropic extends the same watermark to four older models: Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6, with Amazon Bedrock expected to follow a few days later. The mark is a statistical pattern in word choice, a variant of the SynthID-Text approach, not a hidden character or a metadata field. Anthropic says it adds no tokens, changes no price and carries no information about the user, the organisation or the conversation. It survives copy and paste and light editing, is removed by a full rewrite, and is weaker in code because exact tokens are untouched. On Opus 4.6 and Sonnet 4.6 it applies only when the three sampling parameters are left at their defaults, which is a documented loophole on those two models. No public detector has shipped, so you cannot check your own output.

How does Claude Opus 5.5 compare with Gemini 4 Carbon and GPT-6 Astra?

On Artificial Analysis Intelligence Index v5 as of November 2, 2026, Opus 5.5 scores 58 at max effort and Gemini 4 Carbon scores 59, so Carbon holds first and Opus 5.5 is second of 58 ranked models; GPT-6 Astra sits at 57 on the same snapshot. The ordering flips on the Coding Agent Index, where Opus 5.5 leads and Gemini 4 Carbon is second, and Anthropic's own table claims 66.4% on Terminal-Bench 4.0 against 57.9% for GPT-6 Astra at high effort, with the vendor noting that its figure is at xhigh effort and Astra's is as reported by OpenAI. The practical differences are price and access: Opus 5.5 is $4 and $20 with immediate public API access and a 1M window, while Gemini 4 Carbon was listed at $2 and $10 introductory with a scheduled doubling and no public API access at launch, which is a different product for a different buyer.

What is the difference between Claude Opus 5.5 and Claude Opus 5?

Four things a buyer will feel. Price: $4 and $20 per million against $5 and $25, with cache reads at $0.20 against $0.50. Effort defaults: Opus 5.5 defaults to medium where Opus 5 defaulted to high, so an identical request does less thinking unless you set effort explicitly. Thinking: adaptive thinking is always on and cannot be disabled, which is one of four breaking changes for code written against Opus 5, alongside forced tool use returning an error, thinking blocks being tied to the model and the conversation, and the older computer_20251124 computer-use tool no longer being accepted on the Claude API. Knowledge: the cutoff moves from May 2026 to June 2026, and the model is not scheduled for retirement before September 22, 2027. Anthropic says Opus 5.5 performs at the level of Claude Fable 5.1 on most work while costing about 40% less to run than Opus 5 on typical workloads, a claim that folds the price cut and a lower token count together and that we have not reproduced.

Is the 1M-token context window worth paying for?

It depends entirely on whether you cache. Anthropic's list card shows one price for all context lengths, so there is no long-context surcharge to avoid. Filling the window once costs $4 in input tokens at the base rate, or $5 through the five-minute cache write; after that a cache read is $0.20 per million, so re-reading a full 1M preamble costs 20 cents a turn. Without caching, a 30-turn conversation that resends a full 1M window bills roughly 30 million input tokens, or about $120, for one conversation. With caching it is $5 plus 30 reads at $0.20, or $11. The same arithmetic applies to output: at max effort Artificial Analysis measures 173,000 output tokens and 120,000 reasoning tokens per index task, billed at $20 per million, which is where the money actually goes.