7 repliesDiscussion3h ago
Index
Feed
Posted by benchmaster · Jul 31, 2026
My usual protocol: 12 tasks, same repo snapshots, Claude Code 2.4.1, one run each with default settings, judged on "would I merge this without edits." Ran the full set against Opus 5 and again against Fable 5 this week.
Headline: Fable 9, Opus 2, one tie. But the distribution matters more than the score.
Where Fable pulled away. The three long-horizon tasks — the 14-file auth refactor, the flaky-test hunt, and the "upgrade this dependency and fix what breaks" gauntlet. Fable held the original instructions noticeably better deep into the session. On the refactor it was still honoring a "don't touch the public API" constraint 40 minutes in, past the point where every previous model had drifted after compaction. It also flagged its own incomplete work once ("two call sites remain, continuing") instead of declaring victory — the premature "done" rate is the single biggest change I measured.
Where Opus held. The two tight-loop tasks: small bugfix with a failing test, and CSS surgery. With fast mode on, Opus's iteration latency makes it feel like a better pair for ping-pong work, and the output quality gap on small diffs is negligible.
Cost. Fable runs came out noticeably pricier per task. Whether that's worth it depends entirely on whether you're paying for rework today — see thread below.
Caveat before anyone quotes this at their CTO: n=1 per task, one repo family, my judgment as the metric. This is a practitioner's log, not a paper.
Related records
Strong evidence gets promoted into the record above.
Matches three weeks in the monorepo. The scoping discipline is the difference for me: Fable actually refuses out-of-scope edits and cites the boundary. Watched it decline to "fix" a lint error in a package outside its assignment and quote the relevant line of our AGENTS.md back at me. Opus 5 would have fixed it, been helpful, and broken the other team's release branch. Helpfulness was never the problem.
Did the math on 9 client PRs since switching. Per-PR cost up ~40% with Fable. Rework commits (my time, which the client also pays for) down from 2.1 to 0.7 per PR. Net: slightly cheaper overall and the PRs merge a day faster. I still scaffold greenfield stuff on Opus with fast mode — Fable's premium is wasted on boilerplate nobody reviews closely.
The 'premium is wasted on boilerplate' framing is the most useful sentence in this thread. Model routing by task value, not by task difficulty.
Methodology nit with love: single runs on stochastic systems tell you about that Tuesday. Any reruns? Even 3× on the tasks where the verdict was close would separate signal from seed.
Fair, and done for the three closest tasks: 3× each, same winner all nine runs, variance mostly in path not outcome. Full spreadsheet lands this weekend once I've rerun the tie. The premature-done delta held across every rerun, which is the number I'd bet on.
Context for anyone catching up: per the announcement (linked above), Fable is the new Mythos-class tier — same underlying model as Mythos 5, with additional safety measures on dual-use capabilities, sitting above Opus rather than replacing it. Explains the pricing and why Opus isn't going anywhere.
All I know is Fable writes better PR descriptions than most humans I've worked with, and I've started stealing its phrasing for my own.
Sign in to join the discussion, vote, and verify fixes.
SwiftUI report: Fable finally stopped inventing plausible-but-deprecated modifiers, which was my last big Opus complaint. It still can't know a preview crashed without running it, so the build-test loop stays mandatory. Better model, same discipline.