AI News5 min read

Claude 5.5 family: Opus, Sonnet and Haiku pricing and coding fit

Direct answer

Anthropic released three models in its Claude 5.5 family: Opus 5.5 on September 22, 2026, Sonnet 5.5 on September 28 and Haiku 5.5 on October 7. Haiku 5.5 lists USD 0.10 input and 0.50 output per million tokens for prompts up to 100k tokens; Sonnet 5.5 lists USD 2 and 10. The Opus announcement publishes a 40% lower running cost than Opus 5 but no per-token rate.

Anthropic released three models under one family name between September 22 and October 7, 2026. Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 5.5 are separate announcements with separate prices, and the cheapest of the three is not a smaller copy of the other two.

We read the three announcement pages on October 28, 2026. This article separates what Anthropic published from what a team still has to measure.

What did Anthropic release, and in what order?

Model Announcement date Anthropic’s positioning
Claude Opus 5.5 September 22, 2026 First model in the Claude 5.5 family; “performs at the level of Claude Fable 5.1 on most work”
Claude Sonnet 5.5 September 28, 2026 Second model in the family; “a clear upgrade over Claude Sonnet 5”
Claude Haiku 5.5 October 7, 2026 “Fastest, most capable small model”; built for high-volume work

The sequencing is checkable, not just narrative. The Sonnet 5.5 announcement listed Haiku 5.5 as a model that would “join the Claude 5.5 family in the coming weeks”. The Haiku 5.5 page is dated October 7, which closes that statement.

What does each model cost per million tokens?

The Haiku 5.5 announcement publishes the only full rate table across the three pages:

Per million tokens Haiku 5.5 up to 100k input Haiku 5.5 above 100k input Haiku 4.5 Sonnet 5.5
Cache reads USD 0.01 USD 0.05 USD 0.10 USD 0.10
Cache writes USD 0.125 USD 0.625 USD 1.25 USD 2.50
Input USD 0.10 USD 0.50 USD 1.00 USD 2.00
Output USD 0.50 USD 2.50 USD 5.00 USD 10.00

Two consequences follow from that table, and both are easy to misread.

Long prompts change the price. The 100k boundary applies to input tokens, and the higher column then applies to the request. Anthropic’s own justification is usage data: prompts up to 100,000 tokens “make up around 90% of requests to our previous Haiku model”. If your workload sits above that line, the small-model label does not describe your bill.

Cache reads are not the input price. The USD 0.01 figure applies only when eligible input is reused from cache. A workflow with a stable system prompt benefits; a workflow that rewrites context on every call does not.

Opus 5.5 needs a separate note. The Opus 5.5 announcement states that it “costs 40% less to run than Opus 5” on typical workloads and does not print per-token rates. We are not going to infer rates from that sentence.

Two related changes ship with the Haiku launch: Anthropic halved Sonnet 5.5’s cache-read price, which it says makes Sonnet 5.5 run around 20% cheaper on most agentic work, and it introduced a monthly API credit for Claude Max and Team subscribers.

What do the published benchmarks actually say?

Benchmark Opus 5.5 Fable 5.1 Opus 5 GPT-6 Astra GPT-5.6 Sol
Terminal-Bench 4.0 66.4% 55.8% 52.3% 57.9% 37.3%
FrontierCode v1.1 54.4% 50.3% 48.0% 53.3% 47.5%
CursorBench 4.0 57.8% 51.8% 46.6% not listed 41.7%
GDPval-AA v2.1 1846 1735 1708 1542 1588
AutomationBench 40.0% 31.4% 26.9% 41.4% 28.8%

The Haiku page publishes a narrower table, with Haiku 5.5 at 1620 on GDPval-AA v2.1 against 735 for Haiku 4.5, 1437 for GPT-6 Luna and 1840 for Sonnet 5.5. On the OSWorld 2.1 offline subset it reports 72.4% for Haiku 5.5 against 15.7% for Haiku 4.5.

Anthropic attaches its own caveat to the Opus numbers, and it is the most useful sentence in the release: “at these levels of capability we’ve found that benchmark margins have become a less reliable guide to real-world differences”, adding that “the gap between Opus 5.5 and Claude Fable 5.1 is narrower than these scores suggest”.

There is a second trap for anyone building a cross-vendor table. The Opus page notes that its Terminal-Bench 4.0 results are reported at xhigh effort for Opus 5.5 and at high effort for GPT-6 Astra, “as reported by OpenAI”. Numbers from different harnesses and effort settings are not a controlled comparison, however confidently they are printed side by side.

Which model should a team actually pilot?

Anthropic’s own division of labour maps to task shape rather than to a score:

  • Haiku 5.5 for high-volume, cost-sensitive work: summaries, compactions, classification, database queries, and subagent roles inside a larger coding task.
  • Sonnet 5.5 for well-scoped everyday work: fixing bugs, and producing documents, slides and spreadsheets. Anthropic reports 70.6% on Terminal-Bench 4.0 against 10.3% for Sonnet 5, a gap large enough to justify re-testing a workflow that was rejected on Sonnet 5.
  • Opus 5.5 for work “requiring careful judgment”. Anthropic’s examples are a 680,000-line code migration completed by a tester in under a day, and a load-time optimisation that succeeded 39 of 40 times where Opus 5 made smaller changes that also altered application behaviour.

Then convert the vendor framing into your own measurement. Score accepted tasks, not tokens: keep the starting commit, the permitted tools and the acceptance check fixed, and record whether a human had to intervene. The cost-per-task worksheet shows the arithmetic; the model evaluation shelf collects the model claims we have re-read.

What access and safeguards did Anthropic attach?

Opus 5.5 is available “on all platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure”, with the API model id claude-opus-5-5. Haiku 5.5 publishes claude-haiku-5-5. We did not find a Sonnet 5.5 model id on the pages we read, so we do not print one.

On safeguards, Anthropic says Opus 5.5 is “comparable to Claude Mythos 5.1 in biology and cybersecurity” and is therefore deployed with safeguards similar to Fable 5.1. Vetted organisations can apply to the Life Sciences Verification Program, and Anthropic says access to the Cyber Verification Program will expand in the coming weeks. A subscription does not by itself establish eligibility for those programs, the same pattern we documented in the Fable and Mythos access comparison.

For teams that route requests between these models, the operational question is unchanged: find out which model id your account actually calls, and record the returned version with each result. The supported models answer covers what MonkeyCode runs, which is a separate list from Anthropic’s.

Frequently asked questions

Does Claude Opus 5.5 have a published per-token price?

The announcement we read on October 28, 2026 does not print per-token rates for Opus 5.5. It states that Opus 5.5 costs 40% less to run than Opus 5 on typical workloads. Treat any per-token figure you see elsewhere as unverified until Anthropic publishes it.

Is Claude Haiku 5.5 cheaper than Haiku 4.5 for long prompts?

Only up to a point. Haiku 5.5 lists USD 0.10 input and 0.50 output per million tokens up to 100k input tokens, and USD 0.50 and 2.50 above that. Haiku 4.5 lists USD 1 and 5. Anthropic says prompts up to 100k tokens are around 90% of requests to the previous Haiku model.

Do the vendor benchmark tables compare models fairly?

Not necessarily. Anthropic says benchmark margins are a less reliable guide at this capability level, and its own table mixes effort settings: Opus 5.5 at xhigh and GPT-6 Astra at high effort as reported by OpenAI. Run the tasks you care about instead.

Cookie settings

We use cookies only for analytics (GA4 + Matomo) to improve the docs. No ads, no tracking across sites.