Theo's Missing-Opus Theory, the 150K "Smart Zone" & Kimi K3 Refuses to Believe the Jacobian News

Model Wars: Fable, GPT-5.6 & the Missing Opus

Theo's theory: Opus 5 was supposed to retire Fable — and couldn't

Theo's most-discussed post of the day (377k views, 3.1k likes): his theory that Opus 5(.1) was meant to replace Fable 5 for most dev work — cheaper, benching nearly as well — and that the repeated Fable extension deadlines were really Anthropic buying time to improve an Opus that underperformed. He stresses it's 100% speculation, and later adds an alt angle: "Opus 5 benched worse than Kimi K3". The reply thread is a good map of the competing theories:

  • The "capacity, not quality" camp: paperclippriors argues the Fable saga was about inference optimization, noting Thariq's announcement implied real backend lift to make the Max-plan extension happen.
  • The "scale can't be RL'd away" camp: ostyn pushes back that a smaller model fundamentally can't be a "Fable replacement for cheaper" — Fable has the Mythos base, and a bigger model is just better.
  • The "competitors moved the goalposts" camp: several replies argue Opus 5 would have been fine if it had shipped before GPT-5.6 and Kimi K3 landed.
  • And the recurring gripe that won't die: multiple heavy users say Fable's security-refusal behavior (flagged while fixing SSH config, etc.) is pushing them off the model regardless of capability.

Best one-liner: "If true, Opus 5 got benched by the model it was supposed to retire."

Fable 5 vs GPT-5.6: the head-to-head video

Theo also pinned his hour-long Fable 5 vs GPT-5.6 breakdown — strengths, weaknesses, and a personal pick he teases "will likely surprise you." The framing that stuck with viewers is tool vs contractor/coworker: GPT-5.6 as the predictable strong tool, Fable as the more capable but harder-to-manage collaborator — one reply says they now use Fable as the manager and farm work out to GPT-5.6. Recurring reply themes: OpenAI models have "bad taste" in UI/output that no eval captures ("tasteeval when?"), harness matters more than model (5.6 + Codex > Fable + Claude Desktop), and Fable's usage limits keep converting people to "Sol mains" against their preferences.

Related: leerob says Grok 4.5 is really good at React — and notably affordable and token-efficient.

Agentic Coding & Context Engineering

Matt Pocock: 1M context windows are a gimmick — stay under ~150K

Matt Pocock's most-shared post of the day (94k views): huge context windows are nice, but you may be better off treating the first ~150K tokens as the "smart zone" and everything past it as the "dumb zone." The thread is full of practical follow-ups from him:

Adjacent Pocock news: his Claude Code plugin has been approved — a managed installation of his skills ships with the next release — and his upcoming course uses a request logger to show the raw payloads sent to Anthropic/OpenAI, which he says instills "context paranoia — every token counts" and makes the model/harness boundary concrete.

Freerange: proving your UI correct without running it

Cheng Lou (of React/ReasonML fame) announced Freerange (87k views, boosted by swyx): a zero-API static analysis tool that deduces numerical ranges in your code to prove things about UIs — layouts obey specified sizing, no NaN/Infinity, array indices in bounds — with no browser and no code execution. He pitches it as step one in turning vibe coding into proof engineering, and explicitly calls it RL-friendly (verifiable UI properties as reward signals). Pete Hunt's reply: "Make it for sql."

Small notes

Jacobian Conjecture Aftermath

Yesterday's bombshell (Fable's counterexample to the Jacobian conjecture) entered its digestion phase:

Benchmarks & Local AI

swyx on the RLM paper: benchmark gaming via "Temu T-bench"

swyx highlights a trajectory-comparison writeup buried in the RLM paper from Alex Zhang and Omar Khattab: an open secret of frontier training is that you don't need to train on the test set to goalseek a benchmark — training on test lookalikes works ("Temu T-bench"), and since open-weight releases almost never include the datasets/RL environments, there's plausible deniability. The paper explores applying NLP distance metrics to hidden trajectories as a detection approach. The quoted RLM finding is itself notable for harness-pilled readers: well-designed harnesses generalize for the transformer — RLMs trained only on short tasks fully generalize to 8–32× longer ones because the harness makes structurally similar tasks produce near-identical trajectories, even across domains (essay-author matching transferring to math-solution matching). Reply consensus: "held-out test set" is no longer enough for agent evals — you need provenance on the task generator.

Local AI notes

Other Notes