Gemini 3.5 Pro’s development drama returns with fresh inside details! Internal evaluations may lose to Meta’s entry-level model, Muse Spark

Google’s flagship model Gemini 3.5 Pro is stuck in development again, and fresh details have surfaced. Citing insiders, APPSO says the model originally planned for a June launch continued to underperform in coding even after Google updated its training data in late June. The latest internal tests even suggest it may be worse than Meta’s newly released entry-level model, Muse Spark. In recent days, some employees’ testing permissions were revoked; only the model team can view the test results, and morale within the team is low. APPSO reports that Dongqu Dongqu summarizes the report.
(Background: Gemini 3.5 Pro has been delayed for months—Google’s internal political faction infighting has left employees feeling frustrated)
(Additional context: Two top scientists at Google DeepMind have left, the stock price fell 7%, and Demis Hassabis fired back: “We can get talent too.”)

Key summary

• Gemini 3.5 Pro was delayed by several months; after the late-June training data update, coding still failed to meet internal targets.
• APPSO leaks suggest that 3.5 Pro may not even match Meta’s entry-level model Muse Spark.
• Some employees’ testing permissions were revoked; only the model team can see the test results, and morale is low.

Google’s flagship model Gemini 3.5 Pro, already delayed for months, has now been exposed again with an embarrassing inside story. Citing insiders, APPSO says the model slated to debut in June has been stuck mainly due to its coding capability. In late June, Google updated the training data to try to boost its programming-writing abilities, but the results still weren’t satisfactory. Even more awkward, the latest internal tests reportedly suggest that 3.5 Pro may not be able to outperform Muse Spark, the entry-level model from Meta that’s positioned as small, fast, and cheap.

When a flagship model gets blocked in internal evaluations by an opponent’s entry-level product, that’s not exactly good news for Google.

Internal tests may lose to the entry-level model

Breaking down the weight of this matter: Muse Spark is Meta’s first model, positioned as entry-level and focused on being small in size, fast in speed, and low in cost. It has only recently been updated to version 1.1 to strengthen coding and agent capabilities. In plain terms, it’s a product meant to capture usage—not one designed to fight flagship battles. Gemini 3.5 Pro, meanwhile, is Google’s main model meant to take on OpenAI and Anthropic’s top-tier models. When its lead within internal evaluations gets caught up to—or even overtaken by—an entry-level model, it’s natural that the engineering team doesn’t look great.

Testing permissions were scaled back, and morale is unstable

APPSO also reports that in recent days, some employees working on related businesses had their testing permissions revoked, leaving only the model team able to keep using and reviewing the evaluation results. When permissions are pulled inward, it usually means the results aren’t flattering and the company temporarily doesn’t want to share them. Coupled with rumors of team integration and layoffs, morale in the Gemini-related team is low, and some employees have started considering jumping to other AI labs.

This talent-exit story has actually been unfolding for a while. Recently, co-lead of Gemini, Noam Shazeer, moved to OpenAI. The core AlphaFold researcher and Nobel Prize winner John Jumper also left Google to join Anthropic. For a lab trying to hold onto the frontier, none of the leavers are ordinary people.

That said, “3.5 Pro was canceled” currently has no reliable evidence. Google’s external response is that the model is still being tested with partners, and it’s also testing upgrade versions such as Flash.

“We are currently testing 3.5 Pro with our partners, along with an upgraded Flash version and other models, delivering rapidly across multiple models while maintaining high cost-effectiveness for customers.”

Delays can be explained as pursuing excellence, but right now the really thorny external issue is the rumor that “internal testing loses to an entry-level model.”

Frequently asked questions

Why has Gemini 3.5 Pro been delayed repeatedly?

Google originally planned to launch it in June, but its coding capability failed to reach internal targets for a long time. After the late-June training data update, the results still weren’t ideal. The company chose to delay, and instead test 3.5 Pro and an upgraded Flash with partners first.

Has Gemini 3.5 Pro really been canceled?

At present, there is no reliable evidence showing that 3.5 Pro was canceled. Google says the model is still being tested with partners, and it’s also testing upgraded models such as Flash. Claims about cancellation are rumors.