What score (%) will Gemini 3.5 Pro achieve on Terminal-Bench (latest version), as officially or independently reported, using its maximum configuration, within 60 days of public release?

Gemini 3.5 Pro, slated for release in late June or July 2026, will launch with Deep Think integration, establishing its maximum effort configuration. The absolute floor for its performance is set by Gemini 3.5 Flash, which achieved 76.2% on Terminal-Bench 2.1 using Google’s official Terminus-2 harness digitalapplied.comblog.googledeepmind.google. Because Pro is positioned as Google's most capable model and follows Flash in the release cycle, we expect it to definitively clear this baseline at maximum effort.

The current frontier on Terminal-Bench 2.1 is defined by GPT-5.5 at ~83.4% to 84.3% (depending on the Codex CLI vs. xhigh configuration) tbench.aiartificialanalysis.ai and Claude Fable 5 at 84.6% tbench.ai. A standard architectural scale-up from Flash to Pro, combined with the test-time compute scaling of Deep Think, supports a median capability in the low 80s (p50: 81.53%), placing it highly competitive with the current state of the art.

A critical driver of the upper quartiles is the 'highest published' resolution criterion. This incorporates both official and community/leaderboard scaffolds within the 60-day post-release window. Historically, optimized independent scaffolds can extract significant performance gains over standard harnesses—for instance, the Codex CLI provided a ~5-point boost for GPT-5.5. If the community quickly publishes results using highly optimized agentic scaffolds (e.g., an Antigravity-style harness), Gemini 3.5 Pro could comfortably reach the mid-80s (p75: 84.03%, p90: 86.50%), potentially establishing a new benchmark record.

Conversely, the downside risk (p10: 77.07%, p25: 79.37%) accounts for scenarios where Deep Think provides limited marginal utility for long-horizon terminal execution compared to the already heavily agentic-tuned Flash model. Furthermore, stricter independent evaluators (such as Vals, which historically reported a lower 74.16% for Flash vals.ai) might represent the earliest published scores if specialized community scaffolds are delayed.