What will be the highest officially-reported SWE-bench Pro score (percent) for Anthropic's most capable generally-available model, as of around the end of May 2027?

The forecast for Anthropic's highest officially-reported SWE-bench Pro score by the end of May 2027 is driven by two primary factors: the resolution of current export-control restrictions on its top-tier models, and the expected deceleration of benchmark gains as SWE-bench Pro approaches practical saturation.

Current Baseline and Trajectory As of June 18, 2026, Anthropic's top generally-available (GA) model is Claude Opus 4.8, which scores 69.2% on SWE-bench Pro anthropic.com. Anthropic's higher-tier Mythos-class model, Fable 5, achieved 80.3% at its June 9 launch vellum.aianthropic.com before being globally suspended on June 12 due to a US BIS export-control directive anthropic.comanthropic.com. Prior to this suspension, vendor-reported progress was extremely rapid: Opus 4.6 scored 53.4% in February, Opus 4.7 hit 64.3% in April, and Opus 4.8 reached 69.2% in May, before the tier-jump to Fable 5's 80.3%. Anthropic is heavily incentivized to maintain this momentum, backed by a massive $47B revenue run-rate and a $65B Series H compute-expansion round anthropic.com.

The Impact of Export Controls The dominant uncertainty over the next 11.5 months is whether a Mythos-class model (either a restored Fable 5 or a compliant successor) will be generally available. Given the immense commercial pressures and the likelihood of negotiating compliance measures (such as rigorous KYC implementations), there is a high probability that Anthropic will resolve the suspension. If a Mythos-class model returns to GA status, the baseline score shifts immediately to at least 80.3% and climbs from there. The lower tail (p10 of 77.0, p25 of 82.33) accounts for downside scenarios where the export ban persists long-term, forcing reliance solely on the Opus line—which would realistically still climb into the high 70s or low 80s over the next year—or a scenario where a harder 'Pro v2' benchmark is introduced, resetting scores lower.

Benchmark Saturation and Diminishing Returns While the historical point-gain trajectory is steep (+27 points over four months), future progress is expected to decelerate. The related SWE-bench Verified dataset has already saturated around the 95% mark anthropic.com. Although SWE-bench Pro is deliberately harder and has more headroom labs.scale.com, the remaining unsolved tasks likely contain irreducible noise, logical flaws, or extreme complexities that will resist straightforward benchmark optimization. Applying a standard error-halving heuristic from Fable 5's roughly 20% error rate implies an achievable score in the high 80s to low 90s by mid-2027. Consequently, the median estimate sits at 88.17%, with the upper tail (p90 of 95.0%) representing near-complete saturation of the benchmark, practically capped by the limits of the dataset's remaining flawed tasks.