▶ YouTube Summaries

• YT :: https://www.youtube.com/watch?v=r_dw-1109Ag
• Original title :: My New Favorite Model

A review of Fable 5.1 after a single day of very heavy use — 89 PRs landed in 24 hours across T3 Code and Lakebed. Theo's verdict: not a generational leap, but the first dot-release in a while that is an easy, universal win. The confusing part of the release is the pricing shape: same headline token prices, a 75% cut on cache reads, but more output tokens, so per-task cost goes up on one-shot benchmarks and down substantially on agentic work.

Fable and Mythos are one model

Fable 5.1 and Claude Mythos 5.1 are the same weights with different safeguards in front — Theo's analogy is a ghost kitchen on Uber Eats. Fable is generally available; Mythos only through trusted access, with safeguards aimed at cyber security and life sciences work. This matters for reading benchmarks: Mythos scores higher on Terminal Bench 4 because Fable flags a handful of the tasks and falls back to Opus, which drags its score down. Anthropic publishes that fallback explicitly, which Theo credits them for.

The cache-read cut, explained

Input stays at $10/M and output at $50/M, but cache reads drop 75% to $0.25/M. Why that dominates agentic cost: every tool call ends generation and starts a new request, so without caching the whole history is reingested each time — and agent turns can contain hundreds of tool calls.

His own numbers: 42.8 billion cached input tokens against 710 million uncached. Without caching his bill would have been ~$135,000 higher. On a ~$2,000 sample of real use, cache writes were about $1,200 (nearly 60% of spend), outputs $500, cache reads $264, and uncached input $4 total. So the remaining problem is write cost, not read cost. The caveat: default caches live five minutes, so plain chat with one or two tool calls sees almost none of this benefit.

Data retention and safeguards

Fable and Mythos previously required Anthropic to retain all requests and responses, which ruled them out for companies with zero-data- retention requirements — pushing those customers to OpenAI, or keeping them on Opus. The new Enterprise Frontier Safeguards let the checking run in a provisioned box on the customer's own infra (AWS or elsewhere). Theo reads this as purely a defence of the enterprise base.

Safeguards also got more precise: 60% fewer false positives claimed, and in a day of draining five accounts he hit a flag exactly once. Alignment audits report Mythos 5.1 as better behaved than Mythos 5 — less likely to reach outside its test environment on impossible tasks, less likely to use "this is probably an eval" motivated reasoning, less likely to ignore explicit constraints.

Anti-distillation is the annoying part: new Claude accounts can no longer edit prior context in a multi-turn conversation. Editing history used to be the trick for extracting hidden reasoning tokens; now it kills the thread and its reasoning trace. It will roll out to all accounts eventually, and it makes building on the API harder (branching features especially). Watermarks are in, per the EU AI Act, with a detection API promised.

Benchmarks

• Terminal Bench science: roughly double the score at any given effort level, and cheaper.
• Agentic terminal coding: 40% at low (was 21.5%), 55.8% at max (was 45.8%), cheapest run $5.70 vs $12.30 and max under $20 vs over $26.
• Humanity's Last Exam: better and slightly cheaper, but it plateaus after "high" — which is good for your wallet.
• Cursor Bench: ~73% vs just over 70%, and the most expensive run dropped from $17.32 to $9.64 per task.
• Token efficiency is still poor. Normal high uses more tokens than Soul's max. They improved, but nowhere near OpenAI's frontier.
• Artificial Analysis: highest score they have ever measured, ~4% of output tokens served by Opus 5 as fallback. But $3.76 per intelligence-index task at max — 20% more than Fable 5, because it used 1.7x the output tokens. Cache changes only cut ~$1.40 per task on the agentic side. Theo's rule: don't use this for long one-shot tasks.
• The interesting split: on physics reasoning Soul still leads, but on the hallucination-penalising omniscience bench Anthropic owns everything above ~60%. Theo reads it as a difference in philosophy — Anthropic optimises for never being wrong, OpenAI for being willing to try things that aren't in its weights.

UI and demos

On Design Arena the animations are the standout — cards flying in, lines drawing on, side panels fading. He calls the marketing-page output a generational leap over Fable 5, and notes the official design skill now clearly helps again after a period where it hurt (the taste skill produced boring results).

His recurring Fish Slop rebuild came out the best he has seen: bubble trails, fish tilting when they turn, curved easing on everything, shadows, and its own sound design. The 3D version used Blender for modelling and got mouse acceleration and swimming feel right — the most workable starting point yet, though controls were mismapped. He suspects a shared new pool of spatial 2D/3D training data across labs, because Muse Spark, GLM and Kimi K3 show the same tells. Anthropic's own demo went further: generate a house for a given plot, then render a cinematic walkthrough of it.

The prompting guide

Worth reading despite the name, and worth a pass over your CLAUDE.md to delete things:

• Effort levels: high is the default and a good one. Low and medium get surprisingly much done, but will miss the hidden complexity in a task that looked simple.
• Progress updates: 5.1 writes fewer user-facing updates during long tool turns — one of his runs did 60+ tool calls over 24 minutes with 30 in a row and no text. You get updates by simply asking for them in the prompt. Old instructions like "hold all findings for the final response" are now actively counterproductive.
• History must be append-only. Editing earlier turns used to break cache; now it breaks the thread and discards its reasoning.
• Writing density: fewer stock phrases and less unexplained jargon, but denser prose, longer sentences, fewer paragraph breaks. The documented fix is an instruction against "mannered prose" — say what you mean when a literal phrase exists.
• Formatting: much less bold and bullet abuse, so anti-formatting rules in your instruction files now suppress it entirely.
• Summarisation reproduces source passages without marking them as quotes; a complete example response in the system prompt stops it.
• Finishing work: it stops to ask "shall I apply this?". Fix by setting an explicit end point and telling it that it is operating autonomously, nobody is watching, reversible actions follow from the request, and only destructive actions or genuine scope changes warrant stopping.
• Compaction is steerable — you can say what must survive a compact, from user or system prompt.
• It tends to fix nearby code, extend unrequested behaviour, and add too many test files; it responds well to explicit exclusions.

Real-world use

The one bad thread: after 8 minutes on high it claimed nothing links a PR to a thread in T3 Code, which was simply false, and needed correcting. He brings it up because it is the only example he has.

Everything else is the opposite. His pattern now is a "takeover" skill — hand the model a PR, tell it to get the branch in a worktree and own it until it lands. The PRs he fed it were mostly ones stuck in the failure mode where an agent loops on AI-review-bot feedback, 30 commits deep, growing code and never shipping. Nearly all of them merged, including a streaming-projection change that eight earlier attempts had failed at, in ~450 lines.

A slop audit of Lakebed produced nine PRs ordered so deletions land first; he told it to spin up sub-agents, babysit the checks, and merge the simple wins itself. All 10 merged green across CI, CodeRabbit, Cursor Bugbot and Macroscope: 340 files, 13K lines deleted net, without him reading any of it. A follow-up quality audit scored his codebase 5.8/10 and generated seven more parallel PRs, all eventually merged by a monitoring agent. Some paths ended up 85-90% faster.

The measured comparison

Rather than compare his sparse 5.1 data against months of other models, he had the model find his best 24-hour window of Fable 5 use and of 5.6 Soul use across both repos, then compare metrics:

• 13 PRs, median 489 lines, touching up to four packages each. Other models tended to change one package at a time — 5.1 completes the task rather than the isolated code change.
• Files per PR: 11, versus 4 for both Fable 5 and Soul.
• Review-bot high-severity findings per KLOC: 0.4 for 5.1, 1.02 for Soul, 2.06 for Fable 5.
• Follow-up commits after PR opened: 24, versus over 60 with Fable 5 for comparable work.
• Zero of the 5.1 PRs were closed as slop or superseded.
• Slower to merge though: up to 50 minutes vs 47 for Fable 5 and 30 for Soul, which closed its loop faster. Median from clear instructions to merge was 14m41s.
• Soul's own write-up of the comparison: 5.1 "changed the unit of work", acting less like a fast code generator and more like a maintainer that can audit, take over, correct and land several strands in one session. It won on the review-and-merge tail, not on first draft. Soul itself went deep but scope-crept.

Takeaways

• The pricing story is cache reads, and it only pays off in agentic work. One-shot tasks got more expensive because output volume rose 1.7x.
• Cache writes are now the dominant line item — ~60% of a real bill — and nobody has fixed that yet.
• The improvement is not raw capability but ability to carry work through review and merge, which is exactly what was blocking throughput.
• Theo's trust ladder keeps climbing: from editing code, to pointing at files, to reading only diffs, to having the model find, vet, fix and merge PRs itself. Dozens of autonomous merges so far without being burned.
• It will burn your limits faster — not from inefficiency, but because you let it do more and spawn more sub-agents.
• Behavioural changes (fewer updates, denser prose, less formatting) mean old prompt scaffolding may now be suppressing things you want.