--- title: "Google Retrained Gemini for Coding. Bloomberg Says the Results Were Disappointing." date: 2026-07-17 tags: ["gemini","google","deepmind","benchmarks","industry","claude-code"] categories: ["Industry","AI Tools"] summary: "Bloomberg reported July 16 that Gemini 3.5 Pro is months behind its promised June ship date because a late-June retrain aimed specifically at improving coding performance produced disappointing results — the first specific, on-the-record reason given for a delay this blog has tracked since a June 30 DeepMind talent-exodus story. Alphabet shares fell as much as 4.4%, erasing roughly $200 billion in market value, the same week rivals shipped competitive coding models." --- ![Google Retrained Gemini for Coding. Bloomberg Says the Results Were Disappointing.](/images/gemini-3-5-pro-delayed-coding-retrain-disappointing.png) For three weeks this blog has carried "Gemini 3.5 Pro general availability" as a watch item with the same non-answer attached: still not out, still no date from Google itself. On July 16, Bloomberg finally reported *why* — and the reason is more specific, and more damaging, than a generic scheduling slip. Google retrained the model specifically to close its coding gap against Claude and GPT, and the retrain didn't work. ## What Bloomberg Actually Reported According to Bloomberg's sourcing, Google updated the training data feeding Gemini in late June in a targeted attempt to improve coding skills — and internal evaluation of the results was disappointing enough to hold the model back rather than ship it. That's a materially different story than "still polishing" or "quality refinements," the vaguer language Google itself has used publicly. This is a lab that identified a specific weakness, ran a specific fix, checked the output, and decided it wasn't good enough to put in front of customers. The timeline matters here. Gemini 3.5 Pro was announced at Google I/O on May 19 alongside a promise from Google's own blog that the model was "already being used internally" and would roll out "next month" — June. June came and went with only Gemini 3.5 Flash shipping publicly; Pro stayed in a limited enterprise preview on Vertex AI. As of this week, Google's line to Reuters is that it's "currently testing 3.5 Pro, an upgraded Flash model, and other models with partners" — no date, no acknowledgment of the coding-specific root cause Bloomberg reported, and a notable addition: "we're productively engaged with the U.S. government," tying the delay narrative to the same export-control and pre-release-review apparatus this blog has covered around Fable 5 and GPT-5.6. ## A Rough Week to Be Behind on Coding The market reaction was immediate and specific to this report, not to Alphabet's broader business: shares fell as much as 4.4% on the news, erasing roughly $200 billion in market capitalization in a single session. Some analysts pushed back on the alarm — Bank of America reportedly stayed constructive on Alphabet ahead of earnings, citing cloud growth and, notably, the value of Alphabet's own stake in Anthropic. There's an irony worth sitting with there: part of what's cushioning Google's stock from its own flagship model's coding shortfall is Google's financial exposure to the company whose models are setting the coding bar Gemini is failing to clear. That bar moved again just this month. Grok 4.5 shipped July 8 with a mixed-but-real benchmark profile; GPT-5.6's three-tier family cleared its government review and went GA July 9; Meta's Muse Spark 1.1 launched July 9 at aggressive pricing (even if its self-reported benchmarks didn't survive independent verification). Whatever Gemini 3.5 Pro's actual capabilities turn out to be, it's arriving into a market where three other frontier or near-frontier labs already shipped competitive coding models in the same three-week window — and Anthropic's Opus 4.8 (69.2% SWE-bench Pro) and Fable 5 (80.3%) have been sitting at the top of that scoreboard the whole time Google's retrain was underway. ## The Pattern This Blog Keeps Documenting This is the second Gemini 3.5 Pro delay story on this blog, and the throughline across both is worth naming directly. The [June 30 piece](/posts/google-deepmind-talent-exodus-anthropic-gemini-delay/) covered four DeepMind researchers departing in six days — including Jonas Adler, DeepMind's AI coding lead, who left for Anthropic — right as Google confirmed it would miss its June GA target. That story was about capacity: the people best positioned to close Gemini's coding gap were leaving for the company that already had the lead. This story is about the attempt that came after: Google tried a direct, deliberate fix — retrain the model on coding-focused data — and by its own internal evaluation, the fix underperformed. Put those two facts next to each other and the coding gap looks less like a scheduling problem and more like a genuine capability gap that a single data refresh couldn't close. That's not a knock Google will confirm on the record, but it's the plain reading of "we retrained for coding and didn't like what we saw" from a lab with Google's resources, model access, and infrastructure. Contrast it with Anthropic's cadence over the same stretch — Sonnet 5 shipped June 30 as a straightforward default-model upgrade with no drama, Opus 4.8 has held a public benchmark lead for months, and the CLAUDE.md invariant model plus sandbox controls keep shipping as changelog line items rather than being held back for a quarter waiting on a retrain to land. ## What to Watch Next Google hasn't given a new date, and this blog's standing rule is to write on GA only once Google's own blog confirms it — not the leaks, not the "July 17" date that turned out to be exactly that. The more interesting open question now is whether Gemini 3.5 Pro ships with the coding gap only partially closed, given the internal deadline pressure Bloomberg's report implies, or whether Google holds it back again for a second retrain attempt. Either way, "disappointing coding results" from the company that runs one of the three best-resourced AI labs on Earth is a genuinely notable data point about how hard the coding-benchmark frontier has gotten to move — and how much ground Anthropic has put between itself and the field while competitors were still trying to catch up. --- **Sources:** - [Bloomberg: Google Gemini Launch Delayed as Tech Falls Short of Internal Goals](https://www.bloomberg.com/news/articles/2026-07-16/google-gemini-launch-delayed-as-tech-falls-short-of-internal-goals) - [CNBC: Alphabet Shares Fall on Report Its Most Powerful AI Model Gemini 3.5 Pro Is Delayed](https://www.cnbc.com/2026/07/16/alphabet-stock-gemini-3-5-pro-ai.html) - [9to5Google: Gemini 3.5 Pro Delays Due to Coding Performance, Upgraded Flash Model in Testing](https://9to5google.com/2026/07/16/gemini-3-5-pro-delays/) - [Search Engine Journal: Gemini 3.5 Pro Delayed Over Coding, Bloomberg Reports](https://www.searchenginejournal.com/gemini-3-5-pro-delayed-over-coding-bloomberg-reports/582660/)