On this page
Claude Sonnet 5.5 costs half as much per token as Opus 5.5, but not half as much per finished task. At their highest reasoning setting, the independent benchmarker Artificial Analysis measured both at about $10 a task, and Opus scored higher. Even Anthropic's 9 November 2026 launch page for Fable 5.2, its new top model, still says to start most everyday work on Opus.
If you run Claude at its highest effort setting, pick Opus. You pay about the same per task and get the stronger model.
If you run lots of routine work at medium or high effort, pick Sonnet. That is where its half price survives to the invoice.
For classifying or summarising at volume, start with Haiku 5.5.
Sonnet 5.5 and Opus 5.5 at a glance
| Claude Sonnet 5.5 | Claude Opus 5.5 | |
|---|---|---|
| Released | 28 Sept 2026 | 22 Sept 2026 |
| Price per million tokens read / written | $2 / $10 | $4 / $20 |
| Cached input, per million | $0.10 | $0.20 |
| Independent index score, max effort | 55 | 59 |
| Tokens written to run that index | 335M | 218M |
| Cost per index task, max effort | $10.02 | $10.04 |
| Inference inside Canada | No | No |
Prices come from Anthropic's pricing documentation, as of 10 November. Index figures come from Artificial Analysis pages for Sonnet, as of 5 November, and Opus, as of 9 November. Both models read up to a million tokens at a time at the standard rate.
What changed this week
Anthropic launched Claude Fable 5.2 on 9 November 2026. It costs $8 per million tokens read and $40 per million written, a cut from Fable 5.1. Fable now sits above both models in this post.
The launch page did not move Opus aside. Its advice: "For most everyday coding and knowledge work, we recommend starting with Claude Opus 5.5."
Sonnet appears on that page only in a list of models eligible for a cyber security programme.
So the order for heavy work now runs Fable for the hardest technical jobs, then Opus, with Sonnet as the cheaper partner. Our Fable 5.2 vs Fable 5.1 comparison covers the new top model.
Is Sonnet 5.5 good enough instead of Opus 5.5?
For most well-scoped work, yes. Sonnet trails Opus by a few points on most of Anthropic's own tests and beats it on one.
The scores below come from Anthropic's Sonnet announcement, so treat them as the vendor's claim:
| Task (Anthropic's figures) | Sonnet 5.5 | Opus 5.5 |
|---|---|---|
| Finishing multi-step jobs in a command-line terminal | 70.6% | 66.4% |
| Everyday coding help inside an editor | 55.5% | 57.8% |
| Very hard expert questions, allowed to search and run code | 64.5% | 67.7% |
| Real professional work tasks, rated against each other | 1844 | 1846 |
A two-point gap on the professional-work rating is noise. Sonnet leads only on the terminal row, and there Opus ran one setting below max.
Anthropic's own summary is that Opus "remains clearly stronger at complex, open-ended work requiring sustained judgment." The independent picture agrees. Artificial Analysis averages several public tests into one Intelligence Index score. At max effort, Opus scores 59 and Sonnet 55.
The Batch, the DeepLearning.AI newsletter, covered the family on 16 October. It reported that Sonnet at max effort tied Opus one setting lower.
Sonnet gets close to Opus, then. The question is what it spends to get there.
Price per token vs price per finished task
The price that matters is the cost of one attempt divided by how often the attempt succeeds. We set out the arithmetic in the cheapest model is the one that finishes. Per-token price is only the first half of it.
At max effort, Sonnet is the wordiest model on the index: last of 68 for verbosity. It wrote 335 million tokens to finish the tests, where the median model wrote 72.6 million. Opus wrote 218 million. Claude bills its reasoning as written tokens, so those extra words are on the invoice.
Half the price times about 1.5 times the writing comes out level. The index puts Sonnet at $10.02 a task and Opus at $10.04. Opus scores four points higher for two cents more.
The Batch adds the other half of the story. Sonnet's cost per task varied more than eighteenfold across its five effort levels. Effort decides which Sonnet you are paying for.
Anthropic says the same thing in plainer terms. In its words, Sonnet "complements Opus 5.5 best when running at lower effort settings, where it costs less per task."
At higher settings, the announcement adds, Sonnet "can perform comparably at a similar cost."
The rule that falls out is simple. Compare what Sonnet costs per attempt with how often it finishes, both as a share of Opus.
Say Sonnet costs 60% of Opus per attempt on your job. Then it wins as long as it finishes at least 60% as often. At max effort it costs about what Opus costs, so it has to finish just as often. The index says it does not.
| Effort | Defaults | What happens to Sonnet's discount |
|---|---|---|
| Medium | Claude Code and Claude apps | Survives: Anthropic's stated sweet spot |
| High | The Claude API | Smaller; measure it on your own work |
| Max | Never a default | Gone: about $10 per index task for both |
More effort does not always buy a better answer either. On a harder coding test Anthropic reports, Sonnet scored 52.1% one setting below max and 46.2% at max.
Cached input favours Sonnet at every effort level: $0.10 per million against $0.20 for Opus. Anthropic halved that cache price when it launched Haiku 5.5, and said it made Sonnet about 20% cheaper on most agentic work. Caching does nothing for the words a model writes, and the max-effort bill comes from those.
Where Opus 5.5 earns its price
Opus earns its price on long, open-ended work where a wrong turn costs a person's afternoon. It also earns it whenever you would otherwise run Sonnet at max.
Three cases come up most:
- Hard runs a person is waiting on. A failed attempt costs engineer time as well as tokens. Spending more per attempt to fail less often is the right trade there.
- Work that needs judgment over many steps. Long refactors, audits and research tasks are where Anthropic itself says Opus stays ahead.
- Anything you were about to run at max effort. At that setting the two cost the same, so take the higher score.
Opus 5.5 is also cheaper than Opus 5 before it. Anthropic cut prices 20% and says typical workloads cost 40% less at default settings.
Opus 5.5 also has a record on our own tests, for decisions only. In our decision-task run, it got 99.2% of 400 classification decisions right. We ran it on 25 September 2026. That run says nothing about coding, and Sonnet was not in it.
Where Haiku 5.5 fits
Haiku 5.5 fits the high-volume work under both models: summaries, classification, database queries and trimming long histories so the bigger model reads less. At $0.10 per million tokens read and $0.50 written, it costs a twentieth of Sonnet.
Anthropic released it on 7 October as a helper model running beside Opus or Sonnet, not as their replacement. Its own table has Sonnet far ahead on the terminal test, 70.6% against 39.2%. Anthropic says Sonnet and Opus "remain better choices for complex agentic coding tasks."
On cost per task, the gap is wider still. The index lists Haiku at max effort at $0.41 a task, with a score of 36. That is about a twenty-fifth of what Sonnet or Opus cost.
One catch decides where it goes. Haiku is the only one of the three priced by prompt length. A request that reads more than 100,000 tokens pays five times the rate, cached reads included. Keep Haiku on short, repeated jobs and it stays cheap. Feed it a long agent history and the discount shrinks fast.
Put the three together and you get a routing pattern. Haiku takes the bulk, Sonnet at medium or high effort takes routine coding, and Opus takes the hard cases. Our cost breakdown for Claude Code covers where that spend lands in a coding team.
What we did not test
We have not run Sonnet or Haiku on our own tasks. Apart from the decision-task result, every score here comes from Anthropic or the independent index.
The index numbers also move between readings. The Batch reported a Sonnet score of 56 in October, and the index showed 55 on 5 November. For Haiku, The Batch reported 43 at launch and the index showed 36 on 5 November.
The cost-per-task figures come from one index's task mix. Your own ratio depends on what your jobs read and write, and how much of the reading is cached. Log attempts and completions per model on your own work, at the effort level you plan to run, before you standardise on either.
Can you use Claude Sonnet 5.5 or Opus 5.5 with data kept in Canada?
No. Neither Sonnet 5.5 nor Opus 5.5 can process your prompts inside Canada. Haiku 5.5 is the only Claude model that can.
The details, from our Canadian residency guide:
- Amazon Bedrock. Its Canada profile launched on 27 October 2026. It routes only between Montreal (ca-central-1) and Calgary (ca-west-1), and Haiku 5.5 is its only model.
- Other Claude models on Bedrock. Canadian accounts reach them through the
us.orglobal.profiles, which can send requests to American regions. - Anthropic's own API. It has no Canadian option. Its pricing page lists American-only processing as the one regional choice, at 1.1 times the standard price.
For Sonnet-level or Opus-level work processed in Canada, the route is an open-weight model on dedicated GPUs. Open-weight means you can download and run it yourself. The closest match we have covered this week is Moonshot's Kimi K3.1 Pro:
- Size. The model card lists 2.8 trillion parameters and a million tokens of context.
- Coding. Moonshot's own table puts it at 65.7% against Opus's 66.4% on the terminal test, run inside Claude Code. That is Moonshot's claim.
- Where it trails. On very hard expert questions without tools, it scores 51.1% against Opus's 59.8%.
- Licence. Commercial use is allowed. A model-hosting business earning over $20M a year needs a separate agreement with Moonshot. Internal use is exempt from that condition.
- Hardware. The card names serving software but gives no hardware figure, so your workload sets the GPU count.
The trade-offs are real. GPUs are scarcer in Canadian cloud regions than in American ones. You pay for them while they sit idle, and you own the serving. Our Kimi K3.1 Pro vs Claude Opus 5.5 comparison covers the gap in detail.
XY Space deploys open-weight models on dedicated GPUs in the client's own Canadian cloud account, or on their own hardware, so prompts and outputs stay in the country. Then we build the workflow on top, with people approving the steps that matter. See how we deploy.
If you need this model's capability with inference kept in Canada, bring one job to a free 30-minute call.
FAQ
Is Claude Sonnet 5.5 better than Opus 5.5 for coding?
Mostly not. In Anthropic's own tests, Sonnet wins only at multi-step terminal jobs, 70.6% to 66.4%. The Artificial Analysis index scores Opus four points higher at max effort. For routine coding at medium or high effort, Sonnet is close enough and cheaper. For long, open-ended engineering work, Opus is the safer pick.
How much cheaper is Sonnet 5.5 than Opus 5.5?
Per token, exactly half. Opus lists at $4 per million tokens read and $20 written, and Sonnet at half of each. Per finished task, it depends on effort. At max effort the Artificial Analysis index measured both at about $10 a task, because Sonnet writes about 54% more. Anthropic says Sonnet costs less per task at lower effort settings.
What effort level should I use with Sonnet 5.5?
Start at medium or high. Medium is the default in Claude Code and high is the default on the API. Max effort makes Sonnet the wordiest model on the Artificial Analysis index and erases its price advantage over Opus. On a harder coding test Anthropic reports, Sonnet also scored lower at max than one setting below it.
Is Haiku 5.5 good enough to replace Sonnet 5.5?
For summaries, classification, short queries and other high-volume jobs, often yes, at a twentieth of Sonnet's per-token price. For agentic coding, no. Anthropic's own table has Haiku at 39.2% on the terminal test against Sonnet's 70.6%. Prompts over 100,000 tokens also cost five times more on Haiku, so it suits short, repeated work best.
Sources
- Anthropic, Claude Fable 5.2, 9 Nov 2026
- Anthropic, Introducing Claude Sonnet 5.5, 28 Sept 2026. Terminal test: Terminal-Bench 4.0. Editor coding help: CursorBench 4.0. Very hard expert questions: Humanity's Last Exam with tools. Professional work: GDPval-AA v2.1. Harder coding test: FrontierCode 1.1
- Anthropic, Introducing Claude Opus 5.5, 22 Sept 2026
- Anthropic, Introducing Claude Haiku 5.5, 7 Oct 2026
- Anthropic, Pricing documentation, as of 10 Nov 2026 (archived copy)
- Artificial Analysis, Intelligence Index pages for Claude Sonnet 5.5 (max) and Claude Haiku 5.5 (max) as of 5 Nov 2026, and Claude Opus 5.5 (max) as of 9 Nov 2026 (archived copies)
- DeepLearning.AI, The Batch, Claude fills out its model family, 16 Oct 2026
- AWS, Claude Haiku 5.5 Canada cross-region inference on Amazon Bedrock, 27 Oct 2026
- Moonshot AI, Kimi K3.1 Pro model card and licence, Hugging Face, as of 9 Nov 2026 (archived card) and 5 Nov 2026 (archived licence). Very hard expert questions without tools: HLE-Full
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 YongShare 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