Three days before Gemini 3.5 Pro is supposed to ship, there is no model card, no pricing page, and no gemini-3.5-pro listing in the public API documentation. The July 17 date itself comes from third-party reporting — Google has not officially confirmed it.

This is unusual. Anthropic published Fable 5’s model card and pricing two weeks before launch. OpenAI pushed GPT-5.6 Sol’s pricing page to docs.openai.com before broad access opened. At 72 hours out, Google has issued no parallel announcement — no API docs update, no Vertex AI announcement, no model card.

That silence is the most important piece of builder intelligence right now.

What Google Has Actually Confirmed

Nothing specific. Google’s official communication since the June miss:

  • Google I/O, May 19: Sundar Pichai said “give us until next month.” June ended without a GA release.
  • Vertex AI model garden (as of July 13): no gemini-3.5-pro entry.
  • Google AI Studio (as of July 13): no 3.5 Pro model selector.
  • Google documentation: no pricing page for 3.5 Pro exists.
  • Official blog: no launch announcement.

Two missed deadlines followed by zero documentation update 72 hours before the reported third date.

What Third-Party Reporting Says

Every specification circulating comes from leaks, enterprise preview testers, and analyst inference — not from a model card:

Claim Confidence
July 17 GA date Moderate — consistent across independent sources as of July 13
2M token context window Low — no model card corroboration; from enterprise preview leaks
Deep Think reasoning mode Low — unclear whether GA at launch or gated behind preview/Ultra
Input pricing: $12–15 per million tokens Very low — wide variation across sources
Output pricing: $36–60 per million tokens Very low — range spans $36 (leaks) to $60 (analyst estimate)
Architectural rebuild targeting math and SVG Moderate — consistent with preview tester reports

The pricing range is especially problematic. Early framing positioned 3.5 Pro as a “cost-effective frontier” option, priced below Fable 5 ($10/$50 per million input/output) and competitive with GPT-5.6 Sol ($5/$30). But leaked API numbers — $12–15 input, $36–60 output — would actually make it the most expensive frontier model by input cost. Those two narratives cannot both be true. Until a pricing page exists, treat every number as a placeholder.

Access Pathways Right Now

If you want to evaluate before GA:

  • Vertex AI enterprise preview: Limited seats, not public. If you were approved before July 1, you have access; if not, no path in.
  • LMArena community benchmarking: Gemini 3.5 Pro appears in the blind evaluation pool. This is currently the only way most developers can run head-to-head evals against the model before Thursday.
  • Google Antigravity platform: Internal and select partners only. Not accessible to most builders.

For the majority of developers, LMArena is the pre-launch testing option.

What the Silence Might Mean

Three plausible readings:

WAIC timing. WAIC Shanghai opens July 17 — the world’s largest AI conference, with Xi Jinping delivering a keynote. Google may be coordinating the 3.5 Pro announcement for maximum global visibility rather than dropping API docs mid-week. This would explain the silence: they are holding for a press moment.

Another slip. Two prior deadline misses mean a third is not inconceivable. When Google missed the June window, the silence pattern before the miss was similar. If July 17 moves again, the gap between leaked date and official confirmation would again turn out to be meaningful.

Standard launch ops. Google has historically shipped API documentation on the same day as launch rather than in advance — unlike Anthropic or OpenAI, who pre-publish. The silence may simply reflect their operational pattern, not a problem.

All three are live hypotheses. The silence alone does not tell you which one.

What the Architecture Change Means for Your Evals

The July 7 builder guide covers the rebuild in detail. Short version: Google scrapped the 2.5 Pro base and ran a new pre-training cycle. The model shipping on Thursday is architecturally different from the 3.5 Flash released in May and different from what early enterprise preview testers saw.

Three implications for your launch-day evals:

  1. Gemini 3.5 Flash prompts do not port. Flash and Pro share an API surface but not the same base model or training data. Test your prompts fresh rather than assuming Flash-tuned system prompts transfer.

  2. Enterprise preview results may not hold. The GA build’s architecture differs from the preview build. Approved preview testers should re-run their regression suites on July 17 even if they passed in June.

  3. Benchmark numbers are rebuild-specific. Any benchmark published for 3.5 Pro in the July 8–14 window reflects either the enterprise preview build or an unreleased build, not necessarily the GA model.

Builder Action Plan for July 17

Step 0 — Pricing check first. Do not make any routing or cost decisions until you see the official pricing page. If input pricing lands at $12–15/M, that changes the economics completely versus the $2–5/M range where Gemini Flash and Sonnet 5 sit.

Step 1 — Verify Deep Think availability. Is it GA at launch or gated behind the Ultra subscription ($250/month)? If your architecture depends on extended reasoning, this is a binary decision point — you either have it or you do not, and the price tier matters.

Step 2 — Check access pathway. Does “GA” mean open access today, a waitlist, or regional rollout? Google’s prior GA announcements have sometimes meant open access and sometimes meant a long queue. Verify before committing to a migration timeline.

Step 3 — Confirm the 2M context window under load. Context window claims and real-world stability at the top of the range are often different. If your use case needs 1.5M+ tokens in a single request, test at that length on day one before building assumptions into production code.

Step 4 — Run your Gemini evals, not generic benchmarks. Task-specific regression tests against your actual workflows tell you more than leaderboard scores. Have your eval suite ready to run against gemini-3.5-pro as soon as the API listing appears.

Step 5 — Keep Thursday plans otherwise intact. If July 17 slips again, you want a productive day regardless.

Google is three days out from a launch with no official paper trail. That is either a coordinated press play, an ops pattern, or a warning sign. Builders who are ready to evaluate quickly on launch day — instead of scrambling to remember what they wanted to test — will get the most useful signal regardless of which scenario plays out.

Check the Builders’ Week Ahead for the full July 15–21 calendar of events including WAIC, GitHub’s first brownout, and the Fable 5 plan deadline.