Claude Opus 4.8 fast mode ends on 30 November

Anthropic has set 30 November 2026 as the end of Claude Opus 4.8 fast mode. The company announced the deprecation on 30 October. After the cutoff, API requests that specify claude-opus-4-8 together with speed: "fast" will return an error. This is a change to the model's faster inference option, not an announcement that every standard-speed Opus 4.8 request will stop working.

For engineering leads maintaining Claude integrations, the first question is whether a production request explicitly selects both that model and fast speed. An application using the same model without the fast setting is not described by this particular retirement notice. An integration that does select fast mode has a dated compatibility problem, even if the underlying prompt and business workflow have not changed.

Moving off Claude Opus 4.8 fast mode takes more than a model name

Anthropic recommends moving fast mode workloads to Claude Opus 5.5. Its documentation says fast mode is available on Opus 5.5 without the beta header required for Opus 4.8 research preview. However, its migration guidance identifies request and response changes that developers must handle. Opus 5.5 always uses adaptive thinking. Requests that disable thinking, force a named tool, or supply certain sampling parameters can fail with HTTP 400 errors. Response handlers also need to account for thinking blocks rather than assuming that every returned block is plain text.

That distinction matters when an agent calls business systems. A model switch that fixes the fast mode deadline can still break a tool integration if its request builder relies on an unsupported setting. Engineering teams should check model IDs, speed flags, tool choice and response parsing together, then test the actual application path. This is an implementation checklist drawn from the supplier's documentation, not a claim that IC has tested either model.

The fast mode price comparison has limits

Anthropic lists Claude Opus 4.8 fast mode at $10 per million input tokens and $50 per million output tokens, compared with $8 and $40 respectively for Opus 5.5. At an illustrative, unchanged volume of one million input tokens and 200,000 output tokens, those rates produce $20 for Opus 4.8 and $16 for Opus 5.5. The $4 difference is arithmetic, not a forecast of a customer's bill.

The illustration assumes identical token counts and excludes caching adjustments, data residency multipliers, other usage and tax. Different model behaviour or output lengths can change consumption. Anthropic also says fast mode does not work with Batch API or Priority Tier commitments. Teams buying committed capacity should check that constraint before treating a model migration as a like-for-like replacement. Our pricing tracker provides broader AI agent billing context, but current model-specific terms take precedence.

The practical decision is therefore twofold. Identify the requests that will fail after 30 November, then validate the replacement against the application's tool calls and billing assumptions. Anthropic's published deadline is clear; the performance and total cost of a particular migration remain workload dependent.