Fable 5 is back. What came back isn’t identical to what went offline.
The nine-day shutdown that began June 12 under a US government export control directive ended on June 19. API access is restored. The model that returned has stricter safety filters applied under the terms of the restoration, and those filters have a visible effect on developer workflows in specific domains.
Queries in cybersecurity, chemistry, and biology are now routing to Opus 4.8 at what developer community reports describe as a significantly higher rate than the sub-5% fallback threshold Anthropic advertised at launch. The exact rate isn’t confirmed from a primary source, this is developer-reported behavior, not official Anthropic disclosure. The direction is clear, even if the magnitude isn’t: teams that built production workflows in those domains should expect different response characteristics than they saw before June 12.
The Opus 4.8 fallback isn’t a failure mode in the system design sense. It’s a deliberate routing decision with capability consequences. Opus 4.8 is a capable model, but it isn’t Fable 5. If your workflow was specifically optimized for Fable 5’s reasoning depth in chemistry or biological applications, the fallback rate matters operationally.
Now for what’s useful about the restored version. According to Epoch AI’s Epoch Capabilities Index, as previously reported in this pipeline, Fable 5 scores 161 on the ECI, which placed it at the top of the global ranking as of the last evaluation cycle. On FrontierMath Tier 1-3, the reported figure is 87%, with Tier 4 at 88%. Epoch AI’s benchmarks framework provides independent evaluation methodology; the specific score figures are attributed to Epoch AI’s evaluation as reported in prior pipeline coverage, not directly confirmed from page content available to this pipeline. Treat them as Epoch-sourced but treat the specific numbers as figures requiring direct verification from Epoch’s published output.
SWE-Bench Pro sits at 80.3% per the same reporting. That figure carries an important caveat: SWE-Bench Pro was a contested benchmark for Fable 5 as of June 10. The score was disputed by independent evaluators at that point. Whether the restored version’s score reflects the pre-shutdown evaluation or a new evaluation of the restored model isn’t clear. Don’t treat 80.3% as a clean, uncontested number.
One deadline worth noting. The free trial window for Fable 5 reportedly closes June 22. A discrepancy exists in prior pipeline coverage, a June 20 refund window was referenced in a previous brief. These may be two distinct deadlines (refund window vs. free trial window) or a discrepancy in the reporting. If you’re evaluating Fable 5 under a trial arrangement, check your account status directly with Anthropic rather than relying on either date as confirmed.
The model that came back is the one you’ll be working with from here. The benchmark scores give you a performance baseline. The fallback behavior tells you the constraint. Run your domain-specific tests before committing production workloads.