Tag
10 posts tagged.
Two stop reasons mean your response was cut off, and the continuation helper on Anthropic's own stop reasons page only handles one of them. model_context_window_exceeded falls out of the loop and is returned as complete. Adding it to the condition is worse: the continuation re-sends a full window as input, which is documented as a 400 on every model.
The advisor tool is the only server tool Anthropic's count_tokens endpoint accepts, and the count covers the executor's first sampling call only. The top-level usage object is executor-only too, so both default meters miss the advisor sub-inference. Arithmetic on Anthropic's own published example puts 75.2 percent of output tokens outside the headline figures.
An Anthropic Priority Tier commitment is drawn down alongside your ordinary rate limit, not on top of it. Reading the service tiers page next to the pricing page and the models list turns up a 20x burndown spread, a single cache read rate that only holds because of the exclusion list, and three of four current models excluded.
GET /v1/models returns a capability tree and no lifecycle field at all. Meanwhile two of the five platforms that serve Claude run a different retirement clock, and their model tables still list identifiers the Claude API retired months ago.
Context engineering on the Claude platform is not a setting. It is four mechanisms at four points in the pipeline, and the one everybody enables first pays for context window with prompt cache.
A task budget is not a per-task cost ceiling. It resets whenever you send a user message carrying no tool results, which is exactly what most harnesses do when Claude finishes gracefully.
Context rot is real and Anthropic now names it in its own docs. The useful question in 2026 is narrower: your Claude model ships with one of two self regulation mechanisms, or with neither, and they do not overlap. Sonnet and Haiku get context awareness free and automatic. Opus needs a beta header. Opus 4.6 gets nothing.
Comparisons of Claude Opus, Sonnet and Haiku score capability, speed and price per token. None of them mention that the model string is part of your request contract. Swap it and thinking can turn itself on, your answer can move out of content[0], and a max_tokens you never touched can start truncating. Eight measured differences, September 2026.
The Batch API's default prompt cache expires in 5 minutes while the batch itself runs for up to an hour, so the two cost levers quietly cancel each other out. Six documented constraints, and the single assumption underneath all of them. Field log, September 2026.
