Anthropic's Fable 5 answers the old thinking toggle with an HTTP 400

A validation error is the cheapest kind of breaking change. The old thinking block does not quietly degrade to a default on Fable 5; per the dev.to tutorial documenting the updated schema, it fails at the API boundary, as does any `budget_tokens` field ^[3][13]. A schema that accepted those fields and ignored them would have left teams shipping requests whose reasoning spend they could not account for. Failing closed is the right call here.

The grep is the easy half. `budget_tokens` was a number, and the tutorial describes the new `effort` field as the single value that replaces that token-budget dial ^[8]. `effort` has five stops: low, medium, high, xhigh, max ^[2]. So five discrete settings now stand in for an integer ^[9]. If your client derived a thinking budget from prompt length, from a per-tenant quota, or from a latency SLO, none of that arithmetic carries across, because the source names no token count for any effort level and no equivalence table between the two ^[7]. Each call site gets a level chosen by measurement.

The accounting inside `max_tokens` changes shape too. The guidance is to leave enough of `max_tokens` for the final answer at higher effort levels ^[6], which is a heuristic without a number attached: you no longer set the size of the thinking phase, so you cannot subtract it. The worked example pairs `max_tokens: 8192` with `effort: "high"` ^[10]. Treat that as one data point from someone else's prompt. For it to transfer, your answers would have to be about as long as that example's and your effort level no higher, and at `max` neither assumption holds by default.

Two smaller edges. Thinking arrives in streams as content blocks carrying `thinking_delta` deltas, so a parser that switches on block type needs a branch for it ^[11]. And caching system prompts is offered as the lever for reducing redundant thinking cycles ^[12], which is now the main billing control you have, since the direct one was removed.

What the material supports and what it does not is worth being exact about. It supports the enum, the 400, the always-on behaviour, the adaptive option, and a blast radius that also covers Opus 4.7 and 4.8 ^{[1][2][3][4][5]}. It contains no latency figures, no prices, and no per-level token consumption ^[7]. It is also a single publisher's tutorial rather than Anthropic's own reference ^[13]. That is enough to plan a code change and not enough to plan capacity.

The upside of a deleted parameter is that nobody can misconfigure it. The cost is that the migration runs in two passes: strip every `thinking` and `budget_tokens` field, then re-pick an effort level per route while watching for truncated answers. The first pass is mechanical; the second is where the week goes.