Few things motivate a team quite like being written off in public, and Google’s Gemini researchers appear to have taken that idea literally.
Arthur Douillard, a researcher on Google DeepMind’s Gemini RL scaling team, has revealed that his team printed out SemiAnalysis’ now-infamous verdict that DeepMind was “no longer a frontier lab” and kept it on their desks as a daily reminder. The disclosure came a day after Google launched Gemini 4 Neon, the fourth model in a Gemini 4 family that has taken Google from also-ran back to the top of the leaderboards in a matter of weeks.
“I want to thanks @SemiAnalysis_ for the motivation,” Douillard posted on X. “In the Gemini RL scaling team, we printed & put that article on our desks to be reminded daily of what used to be true at the time. But the times they are changing :)”
What SemiAnalysis Said
The passage comes from a SemiAnalysis report titled “Gemini is Cooked but GCP is Cooking,” published on August 7, 2026 by Max Kan, Joey Brookhart, Doug O’Laughlin and Dylan Patel. It was written two days after Google announced a sweeping overhaul of DeepMind’s leadership. DeepMind co-founder Demis Hassabis stepped back from day-to-day operations, Koray Kavukcuoglu took charge of Gemini model development, and longtime Google chief scientist Jeff Dean left to start his own lab, Discovery Loop, taking Gemini co-lead Oriol Vinyals with him. That came on top of a summer of high-profile exits that included Transformer co-author Noam Shazeer leaving for OpenAI and AlphaFold creator John Jumper leaving for Anthropic.
SemiAnalysis read the shakeup as Google effectively conceding the model race. The firm argued that the Gemini team had been losing an internal battle for compute with Google Cloud, which was leasing large amounts of TPU capacity to outside customers, including rival labs. Its conclusion was that Google would make a fine compute landlord but a poor frontier lab. The report was blunt about RL in particular: it pointed to departures from DeepMind’s reinforcement learning teams as one of the main reasons it had given up on Gemini. That makes it fitting that the response came from a team with “RL scaling” in its name.
The critique was not without basis at the time. Google had gone months without a proprietary model above the Flash tier, Gemini 3.1 Pro had lingered in preview, and the company had slid down the independent rankings. It was a sharp fall for a lab whose own leader had said only months earlier that there were three or four leading labs, “of which we are one”.
How Things Changed
The turnaround began on September 30, when Google released Gemini 4 Argon. Artificial Analysis scored it at 53 on its Intelligence Index, level with OpenAI’s GPT-6 Astra and enough to put Google back among the top three labs. Four weeks later came Gemini 4 Carbon, which took the #1 spot on the Artificial Analysis Intelligence Index with a score of 59, ahead of Claude Opus 5.5 and GPT-6 Astra. Carbon also topped Text Arena and Perplexity’s WANDR research benchmark. It even came a close second on SemiAnalysis’ own GDPx benchmark, just behind Claude Opus 5.5 at roughly a third of the cost per task. Gemini 4 Hydrogen, a faster and cheaper option, followed two weeks after that.
Neon, which launched on November 18, is the fourth release. Google pitches it as near-Carbon intelligence at much lower cost and higher speed, and in its launch table Neon beat Claude Opus 5.5 and GPT-6.2 Sol on most benchmarks. Artificial Analysis scored Neon at its xhigh setting at 61 on its Intelligence Index, one point above Carbon and third overall, behind only GPT-6.2 Astra and Claude Fable 5.2. Neon costs about $6.44 per task on that index, roughly 17% less than Carbon, and finishes about 40% faster. On the firm’s Coding Agent Index, Neon scored 59, level with or slightly ahead of Claude Fable 5.2 and Claude Opus 5.5 running in Claude Code, at around half their average cost. API pricing is $3 per million input tokens and $15 per million output tokens, with a 1-million-token context window.
Four models in seven weeks doesn’t look like a lab “meandering.” Google isn’t at the top of every chart this week. Anthropic’s Fable 5.2 and OpenAI’s newly released GPT-6.2 Astra currently sit above Neon on the Intelligence Index, and GPT-6.2 beats Carbon on some benchmarks. But Carbon’s run at #1 has already disproved SemiAnalysis’ claim that Google’s odds of reaching SOTA had “dropped to zero.”
Bulletin-Board Material
Sports teams have long pinned an opponent’s dismissive comments to the locker-room wall, and the Gemini team seems to have done the same with an analyst’s note. Douillard’s post was good-natured, and it’s worth noting that SemiAnalysis’ broader thesis about Google Cloud has largely held up: GCP has kept growing quickly on the back of TPU demand. Even so, the episode is a reminder of how quickly fortunes change at the frontier. Google CEO Sundar Pichai has said that perceptions of who’s leading can swing within four to six weeks because labs ship on different cycles. In the space of a single quarter, Google went from being declared out of the race to holding the top spot on the industry’s most-watched leaderboard.
For the Gemini team, the printout on their desks no longer reads as a forecast. It’s a reminder of where they were three months ago.