Anthropic completed a sequence of Claude 5.5 launches across September and October: Opus arrived first, followed by Sonnet and then Haiku. The useful question for a business is which part of its work benefits from each level of model.
This article uses the release pages as checked on 19 October 2026. Provider performance claims are identified as such; the suggested workflows below are examples to test, not customer results we have measured.
Opus 5.5: the complex-work option
Anthropic introduced Opus 5.5 on 22 September. It positions the model for work requiring sustained judgement and reports improvements over Opus 5 in complex software tasks and its alignment evaluations.
A useful trial would be an application change that crosses several parts of a codebase: a revised permission model, a migration or an investigation into inconsistent behaviour. Give it the relevant constraints and ask for a plan before edits begin. Review whether the plan preserves the product’s existing rules and identifies the areas that need testing.
A larger model is not a substitute for that review. In particular, a convincing explanation of a code change is different from evidence that the change works with your data and integrations.
Sonnet 5.5: everyday development and documents
Sonnet 5.5 launched on 28 September. Anthropic describes it as a faster, lower-cost complement to Opus for well-defined tasks, including bug fixes and document work. It reports output generation more than 30% faster than Sonnet 5. That is Anthropic’s comparison, rather than a speed guarantee for your application.
For an agency or internal software team, test a contained bug report with clear acceptance criteria. Check the fix, the regression coverage and the amount of unrelated code it changes. For document work, compare the output against the supplied figures and source material, not just its formatting.
Measure the whole interaction. If staff need three rounds to correct a result, include those rounds in the time and cost comparison.
Haiku 5.5: short tasks at volume
Anthropic released Haiku 5.5 on 7 October, describing it as a small model for quick, repeated work such as summaries and classification. The announcement uses different pricing tiers for prompts up to and above 100 000 tokens, so a single headline price does not describe every request.
A trial could sort inbound enquiries into agreed service categories or extract reference numbers from sample emails. Define the valid categories and fields first. Send ambiguous requests to a person instead of forcing every message into a category.
Small tasks can still carry sensitive information. Keep the account permissions and information supplied to the model appropriate to the task, regardless of which model handles it.
Build one comparison that staff can review
Choose a set of routine cases and a smaller set of difficult cases from the same workflow. Remove private information or use representative sample records. Have someone familiar with the job mark the expected result before seeing the model’s answer.
Record accuracy, response time, corrections and usage. Include missing attachments, conflicting instructions and requests outside the service you offer. For South African customer work, check dates, rand amounts and the terms your team uses locally.
Use those findings to decide where each model belongs. Keeping a stronger model available for exceptions can be useful, but the routing rule and approval boundary should be explicit. For the wider release picture, see our October frontier AI roundup.