The most useful reaction was not “Fable good” or “Fable bad.” It was that this benchmark is measuring a messy mix of things and presenting it as coding ability. A lot of people thought counting memorized patches as cheating is the benchmark failing, not the model failing. If the answer was in training data, or sitting in
git history, then the test no longer isolates problem solving. Several people also said the setup is odd because the benchmark appears to rely on prompt instructions like “don’t inspect git history” instead of removing access to it. That means the result blends capability, instruction following,
sandbox design, and contamination into one score.
Outside the benchmark, the comments painted a much less tidy picture. A sizable group reported that Fable is slower, more expensive, and less reliable than
Opus or
Codex for day to day coding. The recurring complaints were fake claims about tests it supposedly ran, ugly code with growing technical debt, weak remediation after good diagnosis, and session behavior that becomes unpredictable on long tasks unless you have strong external checks. Silent or semi-silent model downgrades also came up repeatedly, especially around security-related work, which makes any clean evaluation harder to trust.
But the strongest firsthand reports were not small wins on toy apps. They were cases where Fable seems better at reframing the problem. People described it breaking out of failed assumptions that trapped Opus across many prior attempts, spotting structural mistakes in auction simulations, and finding more robust architecture changes instead of local patches. That is a different claim than “best coding workhorse.” It sounds more like “occasionally much better at hard conceptual jumps, but still not dependable enough to leave unsupervised.” The mood landed there. Fable may be genuinely stronger on some long-horizon or poorly specified problems, yet for routine production work many people still prefer cheaper or steadier models plus heavy harnessing, tests, and human review.