Gemini 3.5 Pro Will Reach General Availability Before February 2027

The claim

Google will make Gemini 3.5 Pro generally available — not in partner testing, not in a gated preview, but broadly available to developers through the Gemini API or Vertex AI without an access request — on or before January 31, 2027.

Why this is falsifiable

This prediction resolves on a public, checkable event with a hard date. Either Gemini 3.5 Pro is listed as generally available in Google's own model documentation by January 31, 2027, or it is not.

Resolution criteria, stated in advance so there is no room to reinterpret them later:

• Resolves YES if Google publicly documents Gemini 3.5 Pro as generally available (GA) in the Gemini API or Vertex AI model catalog, accessible without an allowlist or access request, on or before 2027-01-31.
• Resolves NO if the model remains in preview, limited testing, or allowlist access past that date; if it is cancelled; or if it is renamed and shipped as a different version number such that no artifact called Gemini 3.5 Pro reaches GA.
• A rename that Google itself describes as the same model shipping under a new label resolves YES. A rename accompanying a substantially retrained model resolves NO.

The reasoning

On July 16, 2026, Bloomberg reported — citing ten current and former employees, speaking anonymously — that Google is months behind schedule on Gemini 3.5 Pro. The attributed causes are coding capability that failed to improve as hoped following a late-June training data update, and internal contention for compute among Google teams. Sundar Pichai had indicated at the May developer conference that the model would ship in June. Google's only on-record response was that it is "currently testing 3.5 Pro, an upgraded Flash model, and other models with partners."

Everything more specific than that — the widely repeated claims about long-horizon reasoning failures, a scrapped base model, a third missed deadline — appears to originate downstream of the reporting rather than in it, which is the subject of my article on how that story was processed into confident fabrication. This prediction is deliberately built only on the sourced portion.

Three things push toward YES:

The model exists and is in partner testing. This is not a research program; it is a shipping artifact in the final stage before release. The distance from partner testing to GA is measured in months, not years, and Google has cleared that distance many times.

Competitive pressure is acute and rising. Moonshot shipped Kimi K3 at roughly 2.8 trillion parameters on July 16 with open weights promised for July 27. Anthropic and OpenAI both have current flagships in market. A six-month-plus gap at the top of Google's model line is strategically expensive in a way the company has consistently acted to avoid.

Google controls the definition of done. The bar Gemini 3.5 Pro is failing to clear is an internal one. Internal bars are adjustable, and a company months past a publicly signalled date has strong incentives to adjust rather than slip indefinitely. The likely resolution is a release that is good but not what was originally promised.

Two things push toward NO, and they are why this sits at 72 rather than 85:

Compute contention is a structural constraint, not a scheduling one. If engineers are genuinely losing internal capacity fights, that does not resolve because a deadline approaches — it resolves when capacity arrives, which is governed by buildout timelines outside the model team's control. I have argued separately that compute is now rationed by capacity rather than price, and this is what that looks like from inside a lab.

The version-number escape hatch is real. Google has renumbered and rebranded models before. A plausible path is that the work ships under a different name entirely, which resolves this NO on a technicality while the underlying capability does reach market. I have written the resolution criteria to handle the honest version of this, but it remains the most likely route to a NO that does not represent a genuine failure to ship.

Confidence: 72

High enough to be a real commitment, low enough to reflect that the binding constraint may be physical rather than organizational. The compute-contention detail is the one that gives me pause; if the delay were purely about capability tuning, I would be at 85.

What would change my mind

An announced ship date from Google before October 2026 moves this to 90. A second credible report of further slippage, or any indication that the 3.5 line is being skipped in favor of a renumbered successor, moves it below 50.