Anthropic's Claude Fable 5 is a vision-and-reasoning model, and the variant we lean on most — claude-fable-5[1m] — carries a one-million-token context window. We are not going to recite a spec sheet. What we can offer instead is something most write-ups cannot: numbers from our own measurements, with the dates and sample sizes attached, and an honest note about where the evidence runs thin.
Why we made it our default vision proposer
On 9 June 2026 we ran a vision-QC pilot — a set of images, each with a known media-quality flaw to catch, fed to several vision models so we could compare how they behaved. Fable 5 stood out for one quality that matters more than raw cleverness: it was stable.
Across repeated runs on the pilot set it was run-identical 88% of the time, with a flip-rate of 3.9% — meaning it rarely changed its mind about the same image from one pass to the next. On that pilot set it produced zero false positives: it did not invent defects that were not there. It also caught blind spots that other models in the panel missed.
Stability is the property you want from a proposer. A model that flags a real defect today and shrugs at the identical image tomorrow is hard to build a process on. One that answers the same way every time — and does not cry wolf — is one you can put at the front of a pipeline. So we did: Fable 5 is the default vision proposer in our image-consensus panel. (That was a deliberate product decision, made on the strength of the pilot.) It is not in our text-judge pools — its job here is looking at images, not scoring text.
Today's baseline: 300 images, measured live
A pilot is a small thing. So we put the model through a larger run and we publish the result, live, on our vision-QC benchmark.
On 10 June 2026, against the mediaqc-v3-2026-06-10 dataset of 300 images, Fable 5 solo scored:
- 66.9% recall — the share of real defects it caught. That tied for the best single-model result in the run.
- 7.1% false-alarm rate — how often it flagged a clean image. For comparison, another strong vision model on the same run sat more than twice as high on false alarms, which is exactly the kind of difference that decides whether a QC step saves time or wastes it.
- 60.3% class-matched — cases where it not only caught the defect but named the right category of problem.
Recall of two-thirds, solo, is useful but not a finished story — which is the point of running a panel rather than a single model. With Fable 5 sitting in the consensus panel, recall rose to 87.5%. The proposer's steadiness gives the council something reliable to build on; the council closes the gap that one model alone leaves open.
Reasoning and coding: early, honest signal
Vision is where we have the most evidence. On general capability we have less, and we will say so plainly.
In an internal evaluation on 9 June 2026 — a 682-test harness we run across models — Fable 5's results separated sharply from the field: it scored 91% and 73% on the two halves of that harness where the models we compared against scored 3% and 0% on the identical tasks. That is a large gap, and we treat it as a strong signal rather than a settled verdict, because a single harness can flatter or punish a model in ways a second one would correct.
Our first intelligence run on the platform, today, returned a reasoning score of 100 and a coding score of 97. Those are preliminary — a single run, n=1. We are reporting them because hiding early numbers is its own kind of dishonesty, but one run is one run. Watch the leaderboard as the sample grows; that is where these figures will firm up or move.
What the 1M-context variant is for
The claude-fable-5[1m] variant exists for the cases where context is the bottleneck rather than reasoning. A window this wide lets you hold a long document set, a large body of reference material, or an extended interaction in front of the model at once instead of chunking it and hoping nothing important falls out of view. Paired with vision, it suits work where you need the model to reason over a lot of material and look at the images embedded in it — long reports, document sets with figures, screenshots in context.
How to read these numbers
Everything above is measured, dated, and bounded by a sample size we have named. The vision-QC pilot was small; the 300-image baseline is larger but is one dataset on one day; the reasoning and coding figures are early. We would rather hand you that texture than a clean story, because the clean story is usually the one that does not survive the next run.
Claude Fable 5 is available through the Tokonomix gateway and catalogue, and you can put it in front of your own images on our vision-QC benchmark or watch it across tasks on the leaderboard.