Summary

A release walkthrough of Anthropic's Claude Fable 5.1: the specification and pricing as announced, five single-prompt capability tests run by the presenter, and seven usage tips. The structure is deliberate — he states the vendor's claims, labels them as the vendor's, then tries to test them, and is explicit that every test was one prompt and one attempt because that is 'the only fair way to see what it can actually do without a safety net', while acknowledging nobody actually works that way. The video carries a substantial mid-roll advertisement for the presenter's own paid platform and closes with another, so the enthusiasm is commercially interested; the testing itself is nonetheless specific about what broke. All test results are unreplicated single runs by one operator.

• Pricing as reported: the per-token price did not move, at $10 per million input tokens and $50 per million output. What changed is cached reading — re-reading material the model has already seen — which he says fell 75% to $0.25 per million.
• Anthropic's own cost claim, which he explicitly labels as theirs and not his: roughly a quarter off a typical task, and close to half for heavy multi-step work. He adds that independent tests 'don't always land in the same place'.
• A caveat he supplies that the announcement does not: the cheaper re-reading only applies within the same conversation. Dropping a large document in and then opening a fresh chat for each follow-up pays full price every time, and he warns the bill will then look nothing like the announcement.
• Specification as announced: about one million tokens of context, up to 128,000 tokens of output in one response, text and image input, and reasoning before answering. Anthropic lists six areas of improvement — long coding sessions, documents and spreadsheets and slides, multi-step research, reading charts and diagrams, using the large context reliably, and driving interfaces unaided.
• He treats the million-token window as 'a ceiling rather than a promise', arguing that material sitting far from the question is more likely to be skimmed, and advising that what matters be placed next to the question.
• The point he calls the most useful line in the whole announcement: Anthropic tells users not to reach for this model by default, pointing them to Opus 5 for most normal tasks and reserving Fable 5.1 for genuinely complex multi-step work. He notes it is a strange thing for a company to say about its most expensive model, and his closing complaint is that deciding which side of that line a task falls on is left entirely to the user.
• Reported reduction in unnecessary refusals: around 60% fewer false blocks on cyber-security questions and around 85% fewer on harmless biology and medicine questions. He is careful to say the actual limits did not move — real penetration testing, exploit writing and serious biological research are still refused — and characterises it as 'less jumpy, not more permissive'. These are the vendor's figures.
• A counter-intuitive tip he supports with a third-party figure: set the effort level low rather than high. He reports CodeRabbit ran the same review both ways and low effort caught more problems than high — 61% against 57% — and finished almost three minutes faster, and says his own runs matched. This is a single reported comparison, not a benchmark, and the source is named but not linked.
• He reports the model finished his hardest test about 40% faster than Fable 5 and cost meaningfully less, while the older model did better on simple work. His own measurement, one task, no replication.
• Two behavioural quirks he reports and gives prompt-level fixes for. It behaves as though supervised and will stop mid-task to ask permission for steps the original request already covered — he tells it plainly that nobody is watching and to proceed with anything reversible. And it does more than asked: he reports requesting one feature and finding it had rewritten his test files unprompted, and now names what must stay untouched and asks it to report other problems rather than fix them.
• He is careful not to oversell the first fix, noting it does not cure drift on very long sessions and that the model 'still writes code that runs fine while quietly doing the wrong thing' — so the output must be read regardless.
• A cost tip with a clear mechanism: left alone it will regenerate an entire file when the change is four lines, and every one of those output tokens is billed. Telling it to edit only the part that needs editing is invisible on one small script and 'turns into real money' on a project of a few hundred files.
• Advice to delete instructions written for older models from a project's CLAUDE.md, on the grounds that the new model reads them and spends effort fixing problems it no longer has. His suggested test is to remove one instruction, re-run the same task, and treat it as dead weight if nothing changes. He also refers to a terminal command that audits for stale instructions and suggests edits without touching files; the command name is garbled in the transcript and is not reliable as quoted.
• On writing style, he reports Fable 5.1 reaches for mannered metaphor less often than Opus 5 or the original Fable 5, but that sentences run longer with fewer paragraph breaks, making long answers harder to read. His fix is the single instruction to drop all mannered prose.
• Of the five tests, the two he says changed his view are the least flashy: correctly reading a messy hand-drawn app flow photographed on paper — resolving crossing arrows, then flagging genuine gaps in the logic such as no order confirmation after payment and no route back from checkout — and building an animated aurora simulator in plain JavaScript with every outside graphics library banned, where the three sliders each affected only what they should. He reports the aurora stuttering at maximum star density as the one place it slipped.
• The failures in the other tests are reported plainly. A first-person shooter built from one prompt ran with working waves, an overheat mechanic and a kill counter, but alien bodies 'read as spheres with tentacles glued on', enemies barely collided so they stacked, and the frame rate sagged as waves filled the arena. In a travel-planning test the hour-by-hour timing was optimistic about crossing a large city, though he credits it for marking flight prices as estimates rather than inventing bookings, and for solving the geography before writing the plan.

Why it matters

The durable item here is not the model but the vendor's own instruction not to use it by default. A maker telling buyers its most expensive model is the wrong choice for most work is a genuine signal about where capability and cost have separated, and the presenter is right that the burden of classifying a task is pushed onto the user with no guidance. The low-effort-beats-high-effort finding points the same way: more compute is being sold as better and at least one reported comparison found the opposite, which reinforces the Handbook's standing principle that benchmark leadership and configuration defaults should be judged task by task rather than accepted as ordered. The cache-pricing caveat is the practical trap — a saving announced as a headline rate that evaporates if you open a new conversation. Against that, every capability figure here is either the vendor's own or a single unreplicated run by a presenter selling a competing subscription, and should be carried as reported rather than established. The prices, the context window and the refusal rates will all date quickly; the pattern of a maker hedging its own flagship will not.