Gemini 3 Leak?

After they ran 31 leading models, pass@1 scores fell about 6 percent with only five added instructions. Once three or more constraints were introduced, not one of the models broke 50 percent. Comparing results with greater than 800,000 human rankings confirmed what no benchmark had quantified until now: accuracy plus instruction-following tracks real developer preference much better than any single metric in circulation.

Google can be touting a brand new flex: 1.3 quadrillion tokens processed per thirty days. The full is up 320 trillion since June, however the spike traces back to heavier models, not user growth. Gemini 2.5 Flash alone consumes 17 times more tokens per query and might cost 150 times more to run. Even a trivial prompt can spin into multiple reasoning passes, and multimodal inference inflates the count further.

The number signals infrastructure strain, not adoption. It also clashes with Google’s sustainability pitch of 0.24 watt-hours per Gemini request, a figure that only matches lightweight text use and ignores video workloads, agent chaining, and long-context reasoning. Impressive on a slide, less so on a utility bill. Read more.

OpenAI is getting squeezed from multiple fronts. In Recent York, the corporate is staring down a possible multibillion-dollar copyright fight after authors and publishers uncovered internal emails about scrubbing a dataset full of pirated books. They’re now asking the court to force disclosure of OpenAI’s legal communications, arguing the corporate knew what it was doing and could have destroyed evidence. If a judge agrees, damages could explode past a billion dollars, especially after Anthropic already paid $1.5 billion to make an identical lawsuit go away. Insurers are reportedly balking at underwriting either company.

At the identical time, OpenAI is alienating the very policy advocates it claims to collaborate with. Encode and The Midas Project, each tiny nonprofits that backed California’s latest AI transparency law, SB 53, say OpenAI sent sheriffs to serve subpoenas demanding their private emails with lawmakers, journalists, students, and former OpenAI staff. The corporate insists it’s all a part of its lawsuit against Elon Musk and aimed toward sniffing out undisclosed backing. Each groups say they’ve never taken Musk’s money and examine the move as legal intimidation timed to ongoing reviews of OpenAI’s $500 billion reorganization.

To blunt growing distrust, the corporate can be trying to indicate progress on AI bias. In a brand new internal audit, OpenAI stress-tested GPT-4o, OpenAI o3, and the newer GPT-5 models with 500 prompts across hot-button political topics, from immigration to reproductive rights. One other model judged the responses using rules against emotive escalation, one-sided framing, personal opinions, and dismissive language.