What accuracy (%) will Gemini 3.5 Pro achieve on Humanity's Last Exam (HLE), text-only and without external tools, using its maximum publicly available configuration, within 60 days of public release?

The forecast estimates Gemini 3.5 Pro will achieve a median Humanity's Last Exam (HLE) accuracy of approximately 49.9% in its maximum/Deep Think configuration, evaluated text-only and without tools.

To contextualize this, Gemini 3.1 Pro currently serves as the baseline, scoring 44.4–44.7% 66 sources, while Scale reports 47.31±2.11% for the Gemini 3.1 Pro Preview blog.google. More importantly, Google's Gemini 3.1 Deep Think variant already achieved 48.4% on HLE without tools 2 sources. Because the resolution criteria select the highest published score at the model's maximum thinking level within 60 days of release, this 48.4% figure establishes a strong functional floor for the 3.5 Pro generation when utilizing Deep Think.

A key structural tension in this assessment is the architectural tradeoff observed in the 3.5 generation to date. Gemini 3.5 Flash regressed slightly on HLE compared to 3.1 Pro, scoring 40.2–41.0% 44 sources. This reflects Google's prioritization of agentic and coding capabilities in the 3.5 tier, evidenced by 3.5 Flash outperforming the larger 3.1 Pro on Terminal-Bench 2.1 (76.2%) 2 sources. However, the flagship 3.5 Pro is expected to restore and build upon 3.1 Pro’s parametric reasoning while retaining these agentic gains chatforest.com. Combining a more capable 3.5 Pro base architecture with an integrated Deep Think mechanism logically pushes scores into the upper 40s to low 50s.

Downside risks (P10 ~45.6%) account for the possibility that the 3.5 generation's focus on 'intelligence with action' yields stagnant gains on pure academic reasoning benchmarks like HLE, leaving the model anchored near the base 3.1 Pro level. Conversely, upside potential (P90 ~55.4%) is driven by the rapid pace of frontier advancement and the resolution rule favoring the maximum published credible score. The current no-tools frontier is led by Claude Fable 5 at 53.3% pricepertoken.com; competitive pressure could realistically push Google's flagship model, operating at its highest compute-at-inference setting, near or slightly past this mark. Overall, while a jump significantly past the 53% frontier is plausible, a median of ~49.9% optimally balances the known 48.4% baseline of 3.1 Deep Think with the expected architectural priorities of the 3.5 series.