MODELS · 29 OCTOBER 2026
Claude Opus 5.5: when an overall contender is worth the cost
A capable general model can be a sensible default without being the most economical answer to every task. Our Model Spotlight examines Claude Opus 5.5 through that distinction.
What the provider documents
Anthropic lists Opus 5.5 for long-running coding agents and knowledge work, with text and image input and text output. Its documentation dates release to 22 September 2026 and lists a 1-million-token context window. Standard API pricing is US$4 per million input tokens and US$20 per million output tokens, checked 29 October. This is API usage pricing, not a chat subscription or an all-in agent budget. Official model documentation.
Specialist versus overall capability
AIstratica analysis: broad reasoning is useful when a workflow combines reading, planning, coding and revision. A specialist can still be preferable when its narrow skill, deployment cost or latency better matches the job. A short classification request and a multi-file software repair should not inherit the same model choice merely because one model leads a general ranking.
Our existing three-source charts retain their dated snapshots. Epoch's general ECI combines benchmark evidence; Arena's text preferences and agent evaluations are different measurements; OpenCompass uses its own evaluation suite. None is a universal probability of success on your project. We have not replaced text scores with agent percentages or relabelled older tables as today's data.
A cost calculation you can reproduce
At the listed standard token rates, 100,000 input tokens plus 10,000 output tokens costs US$0.60 before additional billable usage, tool charges, retries or applicable taxes. That arithmetic describes one request shape, not the cost of completing a project. Measure total spending per accepted result and include human review time.
Closed service or open weights?
AIstratica analysis: a hosted closed-weight model reduces the work of operating inference infrastructure, while leaving availability, model updates and access terms with a provider. An open-weight candidate can offer more deployment control, subject to its licence and your hardware. It brings responsibility for serving, monitoring, upgrades and security. Neither label establishes privacy or quality by itself. Compare a hosted Opus workflow with an open-weight coding specialist using identical representative tasks and permission boundaries.
Read our open-weight coding specialist guide and the deployment tradeoff explainer.
Migration and agent limits
The provider documents breaking changes from Opus 5, including always-on adaptive thinking and changes to forced tool use. Re-run application checks before switching an existing agent. Migration notes and limitations.
Our recommendation is to shortlist it for mixed, demanding work, then compare against a cheaper model on your real acceptance criteria. Keep financial actions, external messages and destructive changes behind review. A model can perform well in an evaluation and still misunderstand a particular user's intent.
Provider specifications checked 29 October 2026. Decision framework and cost example are AIstratica analysis. This spotlight does not announce a new release or claim independent hands-on benchmark results.