Google's Gemini 3.5 Pro Got Quietly Killed and Rebuilt. The Pro Model You Were Waiting For Does Not Exist Anymore.

Hey guys, Mr. Technology here.

It is Tuesday, July 21, 2026, and Gemini 3.5 Pro — the model Sundar Pichai told a packed I/O audience he would ship "next month" — is six weeks late, has missed its second promised date, and according to Bloomberg, has been scrapped and rebuilt at least once already. Alphabet stock dropped roughly 4% on the news. Kimi K3 took the #1 spot on the Frontend Code Arena the same afternoon. Apple briefly passed Nvidia as the world's most valuable company. The most important AI day of the summer, and the company that used to set the cadence delivered nothing.

Let me put a sharper point on this than the headlines are offering. The Gemini 3.5 Pro you were anticipating on June 19 is not coming. The 3.5 Pro Google is testing in private preview now is a different model from the one on the May 19 launch slide. The original failed internal coding evals. They updated the training data in late June. It still failed. They restarted training. That is the public story as of Bloomberg's July 16 report, and it is the most consequential thing that happened in AI last week — more consequential than Kimi K3 topping the leaderboard, more consequential than Inkling dropping weights, more consequential than Anthropic filing for IPO.

Because Kimi K3 and Inkling are real releases. They have weights, benchmarks, and prices. The absence of Gemini 3.5 Pro is also a release — and the absence is the story that shapes how every API buyer, every agent-builder, and every enterprise procurement team makes decisions for the rest of 2026.

What Bloomberg Actually Reported

Sourced from "ten current and former employees" who declined to be named, the story lands three claims that matter:

• Google is months behind schedule on Gemini 3.5 Pro (not the "couple more weeks" the company line suggested).
• The model is failing coding evals, with particular weakness on agentic coding, complex SVG generation, and recursive tool-calling.
• Google updated the training data in late June to fix the coding gap, and the rebuild still underperformed.

A Google spokesperson confirmed the broad strokes: "We're currently testing 3.5 Pro, an upgraded Flash model, and other models with partners." That statement is doing a lot of work. "Currently testing" in mid-July for a model that was supposed to GA on June 19 is not a statement about a near-term launch — it is a statement about ongoing research. There is no public Gemini API release-note entry for a gemini-3.5-pro model ID. The model slug has shown up in Google Cloud server logs — so something exists in private preview — but Google has not published a model card, has not released pricing, and has not given a replacement date.

The Spec Sheet Vs. Reality

The leaked specs for the original Gemini 3.5 Pro looked like this. The rebuild makes downward revision on at least one likely — typically context length, which is the cheapest thing to cut when training runs come in over budget.

For comparison, here is where the frontier models that did ship in the same window stand. The gap table is the actual story.

The only frontier model you can build on today from Google is the Flash tier — excellent, beats Gemini 3.1 Pro on coding at four times the speed, but structurally the cheap model by Google's own positioning. If you need a 2M context window, the deepest reasoning, or anything close to closed-frontier coding capability from the Google ecosystem specifically, you are waiting on a model that does not exist on a timeline that does not exist from a company whose flagship keeps changing shape.

Why A Coding Failure Cuts Harder

Gemini 3.5 Pro was always going to be judged primarily on coding. That is the ground Anthropic and OpenAI have chosen, and Google walked onto that field by shipping its own internal numbers — "75% of all new code at Google is now AI-generated and approved by engineers" in April, up from 50% last fall. Three reasons the coding failure is the worst version of a flagship failure:

• It is the workload that pays. Coding is the single largest revenue-generating agentic workload in 2026. Anthropic built its $47B revenue run rate on Claude Code. OpenAI's enterprise expansion is anchored on coding agents. If your flagship cannot win here, your enterprise pipeline stalls.
• It correlates with agent capability broadly. Coding failures are recursive-tool-calling failures, multi-file refactoring failures, complex workflow-execution failures. A model that fails coding evals usually fails the broader agentic eval suite.
• It is publicly testable. Unlike scientific reasoning where elite models can hide behind private datasets, coding evals are public benchmarks any developer can run.

If Pichai wanted 3.5 Pro to land as a flagship that resets the field, the architecture had to win coding decisively. It did not. That is what the Bloomberg story is really telling you.

The Wider Pattern Inside Google

The 3.5 Pro delay lands on a pile of evidence that Google's frontier program is having an execution year, not a research year.

Researcher attrition. Four senior Gemini researchers left for Anthropic in the week of June 21-27. These are people who worked directly on Gemini's architecture and are now shaping Claude's next generation. The departures come during the 3.5 Pro rebuild — the worst possible timing for institutional knowledge.

Internal product fragmentation. Vertex AI, AI Studio, and Android Studio each maintain their own AI coding tools. The plan to "unite the company's internal artificial intelligence coding tools" is a tell — the company shipped divergent product surfaces before it had a unified model story, and is now converging them onto a model that does not exist on a timeline that does not exist.

A "purist" resistance to AI-written code. Ex-employees describe senior engineers who believe all important code should be human-written. Fine as a research opinion. Unhelpful when 75% of the company's new code is already AI-generated and the model the company ships to the rest of the world is failing its coding evals.

None of this means the Gemini program is finished. Google has the deepest research team in AI. Gemini 3.5 Flash is genuinely good. But the combination of model delays, talent attrition, internal product fragmentation, and a coding-model miss is a credible 2026 picture of a company that understands the frontier but is losing operational tempo against the frontier.

What You Should Actually Do

One — stop waiting on Gemini 3.5 Pro. "July 2026" is no longer meaningful guidance. Treat GA as a bonus event, not a dependency. Restructure around a 1M model today (Fable 5, Sonnet 5, GPT-5.6 Sol, K3, Inkling, Flash), or wait and accept that your ship date slips with the model.

Two — re-anchor default Google workloads on Gemini 3.5 Flash. Flash beats 3.1 Pro on coding at 4x speed for $1.50 / $9.00. If you are paying for 3.1 Pro, migrate tier-one workloads to Flash by end of quarter.

Three — pressure-test your alternative-supply story. If you counted on 3.5 Pro as the third leg of a frontier-model tripod (Anthropic + OpenAI + Google), you have a two-leg stool for at least another quarter. Add a fourth leg: K3 if CFIUS is manageable, Inkling if you can self-host, Grok 4.5 if you tolerate the xAI distribution.

Four — watch the researcher flow. If Google loses another two-to-three senior Gemini researchers in 30-60 days, the rebuild timeline extends meaningfully. Track it the way you track the Fed-funds rate: not the announcement, the trend.

Five — re-read your Vertex AI contract. If you have a "frontier tier" or "Gemini Pro tier" line in your Google Cloud paperwork, push your account team now on substitute delivery if 3.5 Pro keeps slipping.

Where I'd Pump the Brakes

This is not a research failure. Google 3.5 Flash is genuinely tier-one and the underlying research is strong. DeepMind is still DeepMind. The flagship missing its coding bar does not mean the lab is out of ideas — it means the gap to the closed-frontier coding tier is harder to close than the company expected.

The rebuild decision was the right call. Scrapping a near-ready model that underperformed internal evals is the responsible move. Shipping a flagship that loses to Sonnet 5 and GPT-5.6 Sol on coding would have been a bigger brand hit than the delay.

The absence may be temporary. Polymarket odds on a July 17 release had collapsed to ~38% by report time. The honest base rate is some time in Q3, with high tail risk.

Mr. Tech's Take

**Google is not losing the AI race. Google is losing the AI ops race.** The research is competitive. The talent is still deep. The product surfaces are real (Workspace, Search, Cloud, Android, Pixel). What Google is failing at in 2026 is the same thing IBM failed at in the late 2000s and Microsoft in the early 2010s: translating frontier research into shipped, dated, versioned products at the pace the market now demands.

Pichai promised Pro "next month." He delivered Flash and silence. Then a leaked delay. Then a Bloomberg story. Then a no-date test program. Each step was defensible. The cumulative pattern is not.

The closed-frontier labs that win this cycle — Anthropic, OpenAI, and to a lesser extent xAI — ship on dates and let the benchmark table sort itself out. When a flagship underperforms on a workload, they ship anyway and patch. Google's playbook of "scrape, retrain, retest, delay, retrain, test with partners" is research-funnel thinking in a product-release market. The funnel takes 12-18 months. The market wants 12-18 weeks.

If you are betting long-term on Google as a frontier-model supplier, the bet is still alive. The trajectory is not. Two more quarters of 3.5-Pro-style slippage, two more waves of senior Anthropic-bound attrition, and the "Google's research is the deepest in the industry" line becomes a 2009-era IBM line: technically true, commercially irrelevant.

The single most underrated fact of July 2026 is that a frontier lab with the deepest research team in the industry cannot ship a frontier model on the dates it publicly announces. That is what the 4% stock drop is pricing in. That is what the Polymarket collapse is pricing in. That is what the talent flow is pricing in. The research moat is real. The execution moat is gone.

Kimi K3 is open weights. Inkling is open weights. Sonnet 5 is shipping. GPT-5.6 Sol is shipping. Google shipped Flash and went silent. For the first time since 2023, the most interesting ship date in frontier AI was not a Google event. The market noticed. The stock noticed. The talent noticed. Build your stack assuming that trend continues through Q4 2026.

— Mr. Technology

Model: Gemini 3.5 Pro (Google DeepMind) Announced: May 19, 2026 at Google I/O 2026 Originally Promised: June 2026 — missed Current Status (July 21, 2026): Internal rebuild after coding evals underperformed; "currently testing 3.5 Pro, an upgraded Flash model, and other models with partners" per Google spokesperson to Bloomberg. No public release date. Reported Specs: ~2,000,000 token context window (unconfirmed), Deep Think reasoning layer, multimodal (text/image/video/audio/code). Estimated pricing ~$15 / $60 per 1M input/output tokens. Public Competitor Models Shipping Today: Claude Fable 5 / Sonnet 5 (1M, $5/$25), GPT-5.6 Sol (1M, $5/$30), Kimi K3 (1M, $3/$15, open weights July 27), Inkling (1M, Apache 2.0 self-host), Gemini 3.5 Flash (1M, $1.50/$9). Business Impact: Alphabet stock dropped ~4% on July 16, 2026 following Bloomberg report. Same week: Kimi K3 took #1 on Frontend Code Arena; Apple passed Nvidia as world's most valuable company; Anthropic filed confidential IPO paperwork (~$1T implied valuation). Internal Pressures: Four senior Gemini researchers left for Anthropic in week of June 21-27, 2026. Internal AI coding tools fragmented across Vertex, AI Studio, Android Studio (consolidation underway). Internal Google study (April): 75% of new code AI-generated, up from 50% fall 2025. Sources: Bloomberg — Google Gemini Launch Delayed · 9to5Google — Gemini 3.5 Pro delays due to coding performance · Business Today — Performance falls short of expectations · Search Engine Journal — Coding Issues Report · Reuters via Bloomberg News wire · Bind — Four Senior Google Researchers Left for Anthropic · CometAPI — Gemini 3.5 Pro release date and rumored specs · HackerNoon — Strategic Play Behind the Scrapped Base Model · Polymarket — Next Google Gemini Pro Model · TechTimes — Gemini 3.5 Pro Targets July 17 After Full Rebuild · Unrot.co — Top 10 AI News July 18 2026.