Overview
The week ending February 21 moved AI competition beyond chat interfaces. Google put a stronger reasoning model into preview across consumer and developer channels, OpenAI published mathematical proof attempts and pursued ambient hardware, and regulators confronted a steep rise in synthetic child-abuse imagery. Capability, distribution, and liability advanced together.
Developments
1. Gemini 3.1 Pro doubles Google's verified ARC-AGI-2 score
Google released Gemini 3.1 Pro in preview on February 19 through the Gemini API, Vertex AI, Gemini Enterprise, the Gemini app, and NotebookLM. Its model card lists a one-million-token input window and a verified 77.1% score on ARC-AGI-2, up from 31.1% for Gemini 3 Pro.
The broad launch turns one model revision into a simultaneous test of consumer demand, enterprise distribution, and agentic workloads. Preview status keeps the commercial question open: Google's benchmark gain now has to survive real latency, cost, and reliability constraints across five distinct product surfaces.
Sources: Google announcement · Google DeepMind model card
2. OpenAI publishes five promising proof attempts for expert review
OpenAI published its model's attempts at ten mathematical proof problems on February 20 after circulating them to specialists. The company said expert feedback gave at least five submissions a high chance of being correct, while presenting the work as provisional and inviting further scrutiny.
The publication exposes a more consequential bottleneck than generating a plausible derivation: specialists now carry the burden of checking novelty, hidden assumptions, and correctness. A high success rate after review would strengthen the case for AI-assisted mathematics; rejected proofs would map where fluent symbolic work outruns verification.
Sources: OpenAI proof submissions
3. OpenAI's first planned device is a camera-equipped speaker
OpenAI had more than 200 employees working on a hardware family led by a camera-equipped smart speaker, Reuters reported on February 20 from The Information's account. The reported price was $200 to $300, with shipment no earlier than February 2027; glasses and a lamp were also under consideration.
That schedule places OpenAI in direct competition for the physical space already occupied by phones, speakers, and smart glasses. A continuously sensing device would also make camera access, local processing, retention, and household consent central product questions long before model quality becomes the purchasing constraint.
Sources: Reuters
4. Synthetic child-abuse reports force AI policy into enforcement
Actionable cases of AI-generated child sexual abuse imagery more than doubled in two years, according to reporting published February 20. The Internet Watch Foundation counted 3,440 synthetic abuse videos in 2025, up from 13 in 2024, as Spain and Ireland opened platform investigations.
The increase changes the enforcement problem from isolated image generation to platform-scale detection and removal. Prosecutors and privacy regulators are testing whether recommendation systems and image tools create separate liabilities, while the volume of confirmed material raises the cost of human review for every service carrying user-generated media.
Sources: Reuters · Internet Watch Foundation