How many discrete NAMED reasoning/effort levels will the Gemini 3.5 Pro API expose at public release?

rationale:

As of June 18, 2026, the status quo strongly favors Gemini 3.5 Pro maintaining a 4-level reasoning schema (minimal, low, medium, high). Across the entire Gemini 3.x family—including the already-released Gemini 3.5 Flash—Google's API, Vertex, and OpenAI-compatible documentation consistently expose exactly these four discrete levels 5ai.google.dev. Furthermore, an official "What's new in Gemini 3.5" matrix explicitly notes that "'extra high' is not a supported thinking level" ai.google.dev. Google's established pattern is to treat "Deep Think" as a high-effort mode activated at thinking_level=high rather than introducing a separate enum tier [venturebeat, 4e4301]. Because the 3.5 generation has already shipped a model (Flash) adhering to this 4-level interface buildfastwithai.com, it is highly probable that 3.5 Pro will follow suit when it reaches general availability in late June or July.

The primary risk to this continuity—accounting for the 24% probability on a 5-level schema—is the frontier nature of Gemini 3.5 Pro. Google faces significant API-surface pressure from Anthropic, which exposes five effort levels including xhigh and max platform.claude.com, and OpenAI, which exposes up to six developers.openai.com. Given that Deep Think will be integrated into 3.5 Pro, it is plausible Google could break from the current schema to introduce a single new named tier (e.g., max, xhigh, or deep_think) to differentiate the Pro API from Flash and match competitor granularity. However, social media rumors of an "Extra High" tier currently lack concrete documentation and contradict the explicit rejection of that tier in existing matrices 2 sources.

A jump to 6 or more levels is highly unlikely (3%), as it would represent an unprecedented leap, requiring Google to introduce multiple new tiers simultaneously with no prior testing or hints in the early 3.5 rollout. The most reasonable expectation is that Google maintains its stable 4-level interface while differentiating 3.5 Pro's Deep Think capabilities under the hood at the high setting.