Google Shipped 4 AI Models in 106 Days. The One That Mattered Still Isn't Out.

The promise made on a stage in May

On May 19, 2026, at Google I/O, Google announced Gemini 3.5 Flash and, alongside it, Gemini 3.5 Pro, the model developers actually cared about, with a release window of June. Only Flash shipped that day. Pro did not ship in June. It did not ship in mid-July, when Bloomberg reported that Google had scrapped and was rebuilding the base model after it fell short of internal coding and reliability goals. It did not ship in early August. It was reportedly expected again on September 2, the same day Gemini 3.8 Flash actually shipped, and it did not ship then either. As of this writing, Gemini 3.5 Pro has no model ID, no public price, and no confirmed release date.

Four targets, four misses, and the gap between the original promise and today is now past 100 days. That is not a rounding error on a roadmap. It is the entire lifespan, so far, of every Flash model Google has shipped this year. Compare that to the company Google is actually racing: Anthropic and OpenAI both shipped their current flagship models inside the same window without a single public delay announcement. Google is the only frontier lab in 2026 whose flagship model missed its own announced date more than once, let alone four times in a row.

Why the cheap tier never noticed the flagship was late

The split is not a coincidence of scheduling. A Flash refresh, on Google's own description of Gemini 3.6 Flash, can ship as “an algorithmic refinement” of the model before it rather than a new base model: cheaper post-training passes, a narrower quality bar, and a much smaller blast radius if something regresses, since Flash traffic is disproportionately low-stakes, high-volume work. A flagship Pro model is the opposite bet on every axis: a new base model, a much higher reliability bar, and a launch that Google cannot quietly roll back if it under-delivers, because every competitor and every analyst benchmarks it on day one. Bloomberg's reporting that Google scrapped and rebuilt Gemini 3.5 Pro's base model after it fell short on coding and hallucination goals is exactly the failure mode a lighter-weight Flash release is built to avoid: ship small, ship often, and let the occasional zero-point release pass unnoticed inside a cadence nobody is scrutinizing release-by-release. The flagship does not get that luxury, and this year, it shows in the calendar: four low-stakes ships landed on schedule, and the one high-stakes ship missed every date it was given.

The treadmill that never stopped

While the flagship sat in the lab, the cheap tier moved on its own clock. Gemini 3.5 Flash shipped May 19. Gemini 3.6 Flash shipped July 21, an algorithmic refinement rather than a new base model, according to Google's own release notes. Gemini 3.7 Flash shipped August 13. Gemini 3.8 Flash shipped September 2. Four releases, 106 days, a new Flash model roughly every 26 days on average, the fastest cadence of any frontier lab tracked by Artificial Analysis this year.

The cadence was not just fast, it was accelerating: 63 days between the first two releases, then 23, then 20. Whatever Google changed about how it ships Flash, it worked. The obvious next question is what all that shipping speed actually bought.

What the speed actually bought

On Artificial Analysis's Intelligence Index, the four Flash releases scored 40, 40, 45, and 47, in order. Read the first two again: 3.6 Flash, an entire release cycle after 3.5 Flash, scored identically to it. Zero gain. The three releases after 3.5 Flash, spanning two full months of shipping, added 7 points to the index, combined.

# Cadence vs. capability: what 106 days of Flash releases bought
releases = ["3.5 Flash", "3.6 Flash", "3.7 Flash", "3.8 Flash"]
scores = [40, 40, 45, 47]
release_days = [0, 63, 86, 106]  # cumulative days since 3.5 Flash

total_gain = scores[-1] - scores[0]
total_days = release_days[-1] - release_days[0]
print(f"Total Intelligence Index gain across 4 releases: +{total_gain} points")
print(f"Total elapsed time: {total_days} days")
print(f"Average gain per release: {total_gain / (len(scores) - 1):.1f} points")
# Total Intelligence Index gain across 4 releases: +7 points
# Total elapsed time: 106 days
# Average gain per release: 2.3 points
# One of those three releases (3.6 Flash) added exactly 0.

The same week, the org chart broke

On August 5, 2026, the story stopped being about models. Google restructured its entire AI leadership in a single announcement. CEO Demis Hassabis, who co-founded DeepMind in 2010 and shared the 2024 Nobel Prize in chemistry for AlphaFold, stepped down from running Google DeepMind day-to-day. He becomes chairman of the lab and chief scientist of Alphabet, with more time earmarked for Isomorphic Labs, Alphabet's AI drug-discovery unit, and for what he has publicly called an approaching AGI moment. Koray Kavukcuoglu, DeepMind's CTO and Alphabet's chief AI architect, now runs Gemini day-to-day as SVP, reporting directly to Sundar Pichai. Fortune reported Alphabet shares fell around 5% the same day; ITPro independently put the same-day drop at roughly 4%.

The people who used to build this

The reshuffle landed alongside a departure that mattered more than a title change. Jeff Dean, a 27-year Google veteran and, most recently, Google DeepMind's chief scientist, left the company the same day to co-found Discovery Loop, a public-benefit corporation aimed at automating machine learning, science, and engineering research. He brought three more senior Google researchers with him: Sanjay Ghemawat, Oriol Vinyals, and Quoc Le. Alphabet is, officially, a founding investor in and cloud partner for the new company, an unusual arrangement that reads less like a clean break and more like Google preferring a stake in whatever Dean builds next over competing against it later.

Dean was not the only senior departure this year. Noam Shazeer, a Gemini co-lead, left for OpenAI. John Jumper, a Nobel laureate for his work on AlphaFold, left for Anthropic.

Three senior exits from the team building Google's most important product, inside one summer, landing in the exact window its flagship release kept slipping.

The bill, while the flagship sits idle

None of this comes cheap to sit through. Google is guiding to as much as $205B in capital spending for 2026. Cloud revenue is up 82% year over year, to $24.8B in a single quarter. The infrastructure spend and the cloud growth are both real and both large. Neither one shipped the model.

Meanwhile, everyone else kept climbing

While Google reorganized its masthead, its rivals shipped leaps instead of laps. On September 1, Anthropic's Claude Fable 5.1 posted the highest Intelligence Index score Artificial Analysis has ever recorded on that benchmark: 66 at max effort, three points past its own predecessor, Claude Opus 5 (63). Two days later, OpenAI's GPT-6 Astra gained roughly 85 Elo points on Artificial Analysis's AA-Briefcase evaluation of agentic knowledge work over the model it replaced. Both of those gains landed in the same two weeks Google spent reorganizing instead of shipping.

The tell in Hassabis's own farewell note

The strangest detail sits inside the internal note Hassabis wrote announcing his own exit from day-to-day control. He did not just talk about Gemini 3.5 Pro. He name-checked Gemini 4, a model with no public release window and no benchmark anyone outside Google has seen. Google is already talking, on the record, about the model after the one it still cannot ship. That is either confidence that 3.5 Pro is close enough not to worry about, or a tell that the roadmap has quietly moved on without waiting for it.

What this means if you are building on a single lab's roadmap

None of this makes Gemini a bad model family. Flash is real, it is fast to iterate, and plenty of production workloads do not need frontier-Pro-tier reasoning. What it does prove is that a lab's release cadence and its capability trajectory are two different things, and neither one is guaranteed to hold steady, especially the same summer its research leadership gets reorganized and three of its most senior people leave. A routing decision hard-coded around “Google will ship the next Gemini Pro on schedule” would have been wrong four separate times this year already.

The same logic applies to every other lab, not just Google. Anthropic, OpenAI, and Meta have all had their own quiet-quarter stretches, and the lesson from this particular summer is not “avoid Google,” it is “avoid betting a production system on any single lab's roadmap holding steady.” A leadership reshuffle, a scrapped base model, or a departing research team are the kind of event that shows up in a benchmark chart weeks or months after it actually happens inside the building. The teams that felt this delay first were not the ones reading the news in August. They were the ones who had already hard-coded a June ship date into a Q3 infrastructure plan.

MegaBrain routes every call across 500+ models, including the full Gemini lineup alongside Claude, GPT, and everything else, at zero markup, so a delay, a reshuffle, or a quiet quality plateau at any single lab shows up as a number you can route around immediately, not a roadmap slide you have to keep re-reading for news that never comes.

Sign up at getmegabrain.com to route across the full Gemini lineup and 500+ other models with real, current cost-and-availability visibility, instead of betting a production workload on one lab's next launch date.