AI ComparisonsNov 12, 20269 min read

Claude Opus 5.5 vs Opus 4.8: Cheaper Tokens, New Safeguards

Claude Opus 5.5 vs Opus 4.8: 20% cheaper per token and Anthropic's default, but it always reasons first, and its safeguard declined 7 of our 450 test decisions.

On this page
  1. What changed this week?
  2. What changed from Opus 4.8 to Opus 5.5?
  3. How much cheaper is Opus 5.5 than Opus 4.8?
  4. What do the public coding tests say?
  5. What did our decision-task benchmark find?
  6. Should you upgrade to Opus 5.5 or move to Fable 5.2?
  7. Can you use Claude Opus 5.5 with data kept in Canada?
  8. FAQ
  9. Sources

Claude Opus 5.5 is the better default for most teams on Opus 4.8. The new model costs $4 per million tokens read, against $5. Anthropic's 9 November 2026 launch page for Fable 5.2, its new top model, still sends everyday coding to Opus 5.5. Check one thing first: on our own benchmark of routine business decisions, the new model's cybersecurity filter declined 7 of 450 items.

  • Coding agents or knowledge work on the older model: move up, and re-tune effort before you compare bills.
  • Security work: keep the older model in the loop, or set up fallback first.
  • Still short at high effort on your own tests: try Fable 5.2, at twice the price per token.
  • Prompts that must be processed in Canada: neither qualifies; the last section covers the route that does.
Claude Opus 4.8Claude Opus 5.5
Released28 May 202622 September 2026
StatusLegacy; retirement not before 28 May 2027Current Opus; retirement not before 22 September 2027
Price per million tokens (in / out)$5 / $25$4 / $20
Text it can read at once1M tokens1M tokens
Longest single reply128K tokens128K tokens
Thinking before it answersOff unless you askAlways on
Default effort settinghighmedium
Knowledge cutoffJanuary 2026June 2026
Can you download and run it yourself?NoNo

What changed this week?

Anthropic launched Fable 5.2 on 9 November 2026. It costs $8 per million tokens read, down from $10 for Fable 5.1. The launch page calls it the company's most capable generally available model for technical fields. The same page recommends starting with Opus 5.5 for most everyday coding and knowledge work.

Anthropic's models overview sets the same order: start there, and move up when your evals at higher effort still fall short.

So the new model is the default tier, with Fable above it for the hardest work. The older one is two releases back, with Opus 5 in between since July. The Fable 5.2 vs Fable 5.1 post covers the new top tier, and our Fable 5.1 release notes cover the one before it.

What changed from Opus 4.8 to Opus 5.5?

Three changes reach your code: the new model always thinks before it answers, its default effort drops from high to medium, and it rejects forced tool calls. Anthropic's migration guide lists each one.

  • Thinking is always on. On the older model, a request without a thinking field runs without thinking. Now every request thinks, and asking to disable it returns an error. Effort is the only control left.
  • Default effort is medium. The older model defaults to high. A request that leaves effort out now runs at medium, so set it on purpose.
  • Forced tool use returns an error. Setting tool_choice to any or tool fails. Use auto with strict tool use or structured outputs.
  • Replies can start with thinking blocks. Code that reads content[0].text breaks. Select blocks by their type.
  • Refusals carry a category. A flagged request stops with stop_reason: "refusal" and a category such as cyber or bio. Anthropic documents server-side fallback for those requests.
  • Smaller things. Shorter prompts can now be cached (from 512 tokens, down from 1,024). Priority Tier is gone on the new model and stays on the old one.

The announcement also promises clearer writing that puts the key point first. Its speed claim, over 30% faster output, is measured against Opus 5, not 4.8.

How much cheaper is Opus 5.5 than Opus 4.8?

Per token, the new model is 20% cheaper on input and output, and its cache reads cost 60% less. Whether your bill falls by that much depends on thinking.

Price per million tokensOpus 4.8Opus 5.5
Input$5$4
Output$25$20
Cache read$0.50$0.20
Batch (input / output)$2.50 / $12.50$2 / $10
Fast mode (input / output)$10 / $50, deprecated$8 / $40

Source: Anthropic's pricing page, read on 11 November 2026.

The pricing page gives a sample job: an hour of coding that reads 50,000 tokens and writes another 15,000. At those exact counts, the new model costs $0.50. The older one costs $0.625.

The counts will not stay the same. Thinking is billed as output, and the new model thinks on every request. A job that ran on 4.8 without thinking can write more per call after the switch, which can eat the saving. Anthropic's "costs 40% less to run" line compares the new model with Opus 5, not with the older one. Measure your own jobs at the effort level you choose.

What do the public coding tests say?

No public coding test reports both models on the same version, so there is no fair head-to-head number. The 4.8 announcement reported terminal tasks on version 2.1 of its test. The newer announcement uses version 4.0 and compares only with the release before it and with Fable. Different versions are different tests.

The closest matched step runs from Opus 5 to the new model, on Anthropic's own table:

Task (Anthropic's numbers)Opus 5Opus 5.5
Finishing multi-step jobs in a terminal52.3%66.4%
Long, multi-file software tasks48.0%54.4%
Everyday coding help inside an editor46.6%57.8%
Automating business workflows end to end26.9%40.0%

For the step before that, the Opus 5 announcement says it "more than doubles" the older model on one coding test. The figures sit in a chart, not the text.

Every score above is the vendor's own claim, not yet repeated independently. Anthropic also notes that when its safeguards stepped in during testing, the older model completed some cyber tasks instead. Across two releases, the direction is clear and the size is not measured.

What did our decision-task benchmark find?

On decisions, both models are close to perfect, and the safeguard is the difference. These results cover picking one answer from a list, not coding or writing.

DecideBench v2, measured on 9 November 2026, has 450 decisions written the way production software would see them. Examples include ruling on a return, triaging an email, and picking the next tool for an agent.

DecideBench v2, 9 Nov 2026Opus 4.8Opus 5.5
Items answered450 of 450443 of 450
Declined by Anthropic's cyber safeguard07
Right answers447 (99.3%)442 of the 443 it answered
Cost per million decisions$26,542.70not published
Median wait per answer1,336 msnot published

The seven refusals all came from one family of 50 items: routing an analytics agent to a tool. The safeguard declined them before the model could answer, and they came back with stop_reason: refusal. None of those items is a security task. In one, the right answer is simply the SQL query that fetches the data.

A reference model needs an answer on every item, and the new model left seven blank. So the older model is the benchmark's reference. The raw run for the new model is in the DecideBench repository.

An earlier, easier test agrees on accuracy. On 25 September 2026, the new model got 99.2% of 400 decisions right. Decision model vs LLM has that run, including its cost per decision. The older model was not in it, and the two tests differ, so no cost comparison is fair.

If the new model will sit behind an agent that picks tools, handle refusals before you switch. Read the refusal category and route declined calls to a fallback model.

Should you upgrade to Opus 5.5 or move to Fable 5.2?

Upgrade to Opus 5.5 for most work, and test Fable 5.2 only where the new model at higher effort fails your own evals. The bigger model costs twice as much per token.

On Anthropic's launch page, with each model at its highest effort and safeguards on:

Task (Anthropic's numbers)Opus 5.5Fable 5.2
Long, multi-file software tasks (FrontierSWE v2)62.3%69.9%
Hard agentic coding (FrontierCode 2.0)45.5%49.1%
Paid freelance projects done to standard21.3%41.3%
Long knowledge-work projects, rated head to head (Elo)18001785

The bigger model leads on long coding and agent work. The new Opus edges ahead on long knowledge-work projects such as memos, spreadsheets and decks. The Sonnet 5.5 vs Opus 5.5 post covers the cheaper tier below.

The older model still earns a place in three cases:

  • Priority Tier. Commitments on that tier still need 4.8.
  • Security work. Anthropic says most cybersecurity tasks on the new model are re-routed to 4.8 anyway.
  • Code not yet migrated. Anything that forces tool calls or switches thinking off keeps working on 4.8 until it is rewritten.

Retirement is not before 28 May 2027, so there is time to test.

What we did not test: coding, writing, or long agent runs on either model. Our numbers cover decisions only, and the bigger model is not in DecideBench yet.

Can you use Claude Opus 5.5 with data kept in Canada?

No. Both models are closed, so your prompts are processed on infrastructure run by Anthropic or a cloud partner, and no documented route processes either one in Canada.

  • Anthropic's own API offers global processing or processing inside the United States, which costs 10% more.
  • Amazon Bedrock's Canada profile, launched on 27 October 2026, keeps requests between Montreal and Calgary. Claude Haiku 5.5 is its only model.
  • Google's residency page lists no Canada entry for partner models such as Claude.

Our Canadian residency guide has the full provider table.

For inference in Canada at this level, the route is a model you can download and run yourself, on dedicated GPUs in Canada. The closest from this batch is Moonshot's Kimi K3.1 Pro. It has 2.8 trillion parameters, 104 billion of them active per token, under Moonshot's own licence.

On Moonshot's own table it nearly ties the new Opus on terminal jobs, 65.7% to 66.4%. It trails badly on scientific computing in a terminal, 38.1% to 63.3%. The card gives no GPU requirements, so the hardware is set by your workload. Our head-to-head of the two compares them in full, and our post on Kimi's previous model against Opus 4.8 covers the earlier pair.

Running it yourself has trade-offs. GPU capacity is scarcer in Canadian regions than in American regions, you pay for GPUs while they sit idle, and you own the serving.

XY Space deploys open-weight models on dedicated GPUs in the client's own Canadian cloud account, or on the client's own hardware, so prompts and outputs stay in the country. Then we build the workflow on top. How we deploy covers the build.

If you need this model's capability with inference kept in Canada, bring one job to a free 30-minute call.

FAQ

Is Claude Opus 5.5 better than Opus 4.8?

For most coding and knowledge work, yes, going by Anthropic's own tests and its own advice. No public test reports both on the same version, so the size of the gap is not measured. On our decision benchmark of 9 November 2026, both got almost every item right. The new model's cyber safeguard declined seven routine items that the older one answered.

Is Claude Opus 4.8 being retired?

Not yet. Anthropic lists it as legacy: still available, with retirement not before 28 May 2027, and a recommendation to migrate to Opus 5.5. Anthropic gives at least 60 days' notice before retiring a publicly released model. The older model also keeps Priority Tier, which the new one does not support.

Why does Opus 5.5 refuse requests that Opus 4.8 answers?

The new model runs classifiers for cyber and biology content. A flagged request can end with stop_reason: "refusal" and a category. Anthropic says most cybersecurity tasks are re-routed to the older model, and it documents server-side fallback for the API. In our test, the safeguard declined seven requests about routing an analytics agent to a tool.

Is Opus 5.5 actually cheaper to run than Opus 4.8?

Per token, yes: it costs 20% less to read and to write. Per job, it depends. The new model always thinks, and thinking bills as output, so a job that ran without thinking on 4.8 can write more after the switch. Lower the effort setting where quality holds, then compare real bills.

Sources

Written by

Cho Yin Yong

Principal AI Solutions Engineer, XY Space

Principal AI Solutions Engineer at XY Space. University of Toronto lecturer for five years, co-author of two patents, winner of two competitive AI awards, and nine years of regulated engineering leadership.

More from Cho Yin Yong

Share this article

Work with us

We build the systems these posts describe, and we'll tell you in the first call whether yours is worth building.

Start a project
Work with us

Book a call.We'll come back with specifics.

Start with the map of your organization, or with the one job that hurts. Measured in hours and money, and everything we build stays yours.

Loading form…