Gemini 3.5 Pro misses a third deadline: Google scrapped the base model, and a stopgap Flash may ship first

Where things actually stand

Google announced Gemini 3.5 Pro at I/O on May 19, 2026, with Sundar Pichai telling the audience it would arrive “next month.” Since then it has slipped three times: June came and went, an early-July window passed, and the July 17 date that dominated third-party reporting arrived with no launch.

As of July 20, 2026, the verifiable status is short and unambiguous:

• No official Google announcement of general availability.
• No model card, no published pricing, no confirmed spec sheet.
• No gemini-3.5-pro entry in Google’s public API model list — only Gemini 3.5 Flash is generally available.
• Access remains limited to enterprise preview on Vertex AI plus testing surfaces like Antigravity and LM Arena.

Everything about why comes from reporting, not from Google.

What’s reported (and worth reading carefully)

Per Bloomberg’s reporting and 9to5Google, the delay traces to the model falling short on coding performance and complex, long-horizon reasoning. The rebuilt version reportedly failed key reliability standards, including hallucinating too frequently, and fell short of GPT-5.6 in benchmark testing.

The more striking claim: Google scrapped a base model it nearly had ready and restarted pretraining from scratch. Two concrete failure modes were cited for the scrapped version:

1. It could not maintain structural consistency when generating complex, multi-layered SVG scene layouts.
2. It broke down under complex, recursive tool-calling environments.

Meanwhile, model registrations for a new Flash tier — plus an anonymous Flash checkpoint spotted on LM Arena that outperformed current Gemini 3.5 Flash in blind comparisons — point to a stopgap release. Early testers described the gap as incremental rather than generational. One caution the coverage often drops: the name “Gemini 3.6 Flash” is an extrapolation from Google’s versioning conventions by TestingCatalog, not a Google-confirmed product name.

Why this matters

1. The two failure modes are exactly the agentic frontier. This is the most revealing detail in the whole story. Recursive tool-calling and structurally consistent generation aren’t obscure benchmarks — they’re the core competencies for agents that plan, call tools, and produce complex structured output. A model that breaks down there isn’t merely “a bit behind on coding”; it’s not ready for the workloads the entire industry is racing toward. Google isn’t stuck on chat quality. It’s stuck on agentics, in the same quarter GPT-5.6 shipped with subagent “ultra mode”.

2. Scrapping a nearly-ready flagship is both alarming and defensible. Restarting pretraining on a model that’s almost done is enormously expensive and signals real, structural problems. It’s also, arguably, the disciplined choice: shipping a hallucination-prone flagship would have been worse, and would have handed rivals a cleaner narrative. Both readings are fair. What isn’t in dispute is that this was not a minor polish delay.

3. The competitive cost is now concrete. In the space of two weeks, GPT-5.6 went public and Grok 4.5 shipped — both with independent benchmarks — while Google’s flagship slipped a third time. Combine that with the DeepMind researcher departures earlier this summer and the top-tier gap is real. It’s worth keeping proportion, though: Gemini 3.5 Flash is GA and genuinely good, and Google’s Workspace distribution and Apple Siri deal are strategic assets no benchmark captures.

4. A stopgap Flash tells you what Google thinks it needs. Preparing an incremental Flash release while the flagship is rebuilt is the move of a company that needs something on the board. For buyers that’s actually fine — the Flash tier is where most real workloads live on price-performance — but it confirms the Pro tier isn’t close.

The delay timeline, in one place

It’s easy to lose track of how long this has run, so here’s the sequence:

• May 19 — Announced at Google I/O. Sundar Pichai tells the audience it arrives “next month.” Gemini 3.5 Flash ships same-day and is genuinely strong.
• June — The stated window passes with no release and no explanation. Reporting attributes it to “quality refinements after early enterprise testing.”
• Late June — Four senior DeepMind researchers depart in about a week, three to Anthropic, compounding the execution questions.
• Early July — A revised early-July window slips. The model remains in limited Vertex AI enterprise preview.
• July 13–17 — Third-party reporting converges on July 17. Google never confirms it. The date passes.
• July 16–20 — Reporting surfaces the scrapped base model, the restarted pretraining, the hallucination and recursive-tool-calling failures, and registrations pointing to a stopgap Flash release.

Two months from “next month” to “no date at all,” with a full pretraining restart in between. The pattern — repeated unconfirmed dates that slip — is itself the useful signal: Google has consistently declined to commit publicly, which is what a company does when it doesn’t yet know.

What this means for you

• Don’t architect around Gemini 3.5 Pro. After three missed targets and no official date, the safe planning assumption is that it’s not arriving imminently. Anything you’ve deferred waiting for it should move.
• If you want Gemini today: Gemini 3.5 Flash is generally available, beats the old 3 Pro on benchmarks, and is the right price-performance pick for most workloads.
• If you need frontier reasoning now: GPT-5.6 (Sol/Terra) and Claude Opus 4.8 / Fable 5 are shipping and independently benchmarked. See the best AI chatbots guide and Claude vs Gemini.
• If you specifically need the rumoured 2M-token context: it remains unavailable and unconfirmed. Don’t design a document-processing pipeline around a spec Google has never published.
• If you’re a Workspace shop: the integration advantages haven’t changed. This is a flagship-model delay, not a reason to leave the ecosystem.
• If you were waiting to benchmark it against your stack: run that evaluation now against what’s actually shipping. Three missed dates is enough evidence that a comparison held open for Gemini 3.5 Pro is a comparison you may never get to finish — and decisions deferred indefinitely have their own cost.

The honest caveats

• Almost everything about the cause is reported, not confirmed. Google has not announced a delay, explained one, or published a date. The scrapped-base-model account, the failure modes, and the benchmark comparisons all come from press reporting and unnamed sources.
• “Gemini 3.6 Flash” is not an official name. It’s an extrapolation from versioning patterns. The underlying signal — an unreleased Flash checkpoint testing well — is better sourced than the label.
• A rebuilt model could still be excellent. Restarting pretraining to fix structural weaknesses is the kind of decision that produces a better eventual product. A long delay is not a prediction of a bad model.
• Delays can end abruptly. Google could ship without warning; treat “not imminent” as a planning assumption, not a forecast. We’ll update when there’s an official announcement on Google’s own channel.
• This is one product line. Gemini 3.5 Flash, Workspace integration, the Apple deal, and Google’s infrastructure position are unaffected by a Pro-tier slip.

The practical takeaway: three missed deadlines with no official word is enough information to stop waiting. Build with what’s shipping — and re-evaluate Gemini 3.5 Pro when Google, not the rumour mill, actually announces it.

Questions or corrections? Email Pick Right. Want the full list? See all news.