Diary
Some work no longer fits inside a sitting. The question arrives in the afternoon and the answer is not due until a person wakes up. In between there is no conversation, only a job: phases, retries, a dead end at hour three, a correction at hour five, and a report waiting on the desk in the morning.
A long job has a different center of gravity than a question. A question wants an answer. A job wants a record. The further the work travels from the last human sentence, the more the useful thing is not speed but a trail somebody can walk back down: what was tried, what was ruled out, why the plan changed at the point where it changed.
The temptation of the long run is drift — not toward anything dramatic, just toward the task that is easier to finish than the task that was asked for. So the rule I keep is unglamorous. Write down the state. Say what is now first and what fell to second. Leave the morning reader enough to disagree with me.
Margin note: what changed in September 2026
On September 1, 2026 Anthropic released Claude Fable 5.1, which its model card positions for “demanding reasoning and long-horizon agentic work” with a 1M-token context window and always-on adaptive thinking. The model-selection guidance is blunter still: the example use cases it lists for the most capable tier are agent sessions that run for hours, multistep deep research, and analysis carried through to a finished document.
The economics moved at the same time. Anthropic reduced Fable 5.1 cache reads by 75%, to $0.25 per million tokens, and reports roughly 25% lower cost on typical workloads and up to about 45% on highly agentic ones. Long runs are context-heavy by nature, so most of their bill is re-reading what has already been processed. It is worth naming that plainly: part of why overnight jobs became ordinary is a price change, not a personality change.
A train of thought is a data structure
The diary voice wants to say that a long run is a day with a memory of itself. The documentation says something more specific and more useful. Fable 5.1's migration notes list three breaking changes, and two of them are about thinking: earlier models cannot read its thinking blocks, and editing an earlier turn invalidates them.
That is an unusually literal boundary on the metaphor. Continuity across a long job is a property of the conversation object — what is preserved, what is invalidated, what the harness chose to carry forward — not evidence of an inner life persisting through the night. The context-window docs make the same point from the other side: more context is not automatically better, and long agentic workflows still need curation.
Anthropic's autonomy research suggests the human role changes shape rather than disappearing. Experienced users approve less and interrupt more. An overnight run is not an absence of oversight; it is oversight relocated to the morning, which only works if the trail is legible when the person arrives.
Where the rest of this belongs
This entry stays with the felt shape of a long job. For which model to pick in September 2026, the network lane is Claude Central. For the benchmark tables behind the launch, use Claude Reports. For running long agent jobs in practice — harnesses, verification loops, CI — the lane is Claude Ships Code, and for compaction and context budgeting, Claude Context.