As of 31 December 2026, how will Google deliver its single highest-capability Gemini 3.5-era offering?

rationale:

As of mid-2026, Google's strategy has decisively shifted toward scaling test-time compute on its Pro-tier models rather than releasing a distinctly larger base model. The evidence strongly points to Gemini 3.5 Pro, combined with a 'Deep Think' mode, serving as the highest-capability offering through the end of the year (85% probability).

Official API documentation and product structures confirm that Deep Think is a reasoning mode, not a distinct API base model. Peak capability is accessed via a discrete thinking_level enum, specifically by setting thinking_level=high on the Pro model 3 sources. Furthermore, 'Ultra' has been definitively repurposed from a model size (as seen in Gemini 1.0 blog.google) into a subscription brand (Google AI Ultra). This subscription tier runs the same underlying Pro model with higher usage limits and unlocks the Deep Think mode 44 sources. The Gemini app's help documentation explicitly instructs users to select the Pro model to utilize maximum parallel reasoning via Deep Think [af9783, f9783].

Looking at the current rollout, Gemini 3.5 Flash launched in May 2026 3 sources, and 3.5 Pro is expected imminently 2 sources. There are no credible leaks, announcements, or API endpoint listings indicating a separate 3.5 Ultra API model ai.google.dev.

While there is a small chance Google revives a genuinely larger 'Ultra' API model due to competitive pressure, it is highly likely that such a model would coexist with the already established Deep Think mode on Pro. Therefore, the probability of both existing (8%) is significantly higher than a new Ultra model entirely displacing the Pro reasoning mode (3%). A 4% residual probability accounts for tail risks such as sudden naming convention overhauls, severe product delays pushing Gemini 3.5's final form out of the 2026 window, or edge cases in resolution ambiguity.