The week in AI

Most of the week's news came from Google I/O on 19 May. Google announced Gemini 3.5 Flash, a video model called Gemini Omni Flash, a standalone Antigravity 2.0 agent app, and a background personal agent called Gemini Spark. The next day OpenAI said an unreleased reasoning model had disproved an 80-year-old belief in discrete geometry, and outside mathematicians checked the result.

On 22 May Anthropic published its first report on Project Glasswing. Partners using Mythos Preview found more than 10,000 high or critical vulnerabilities, and the open-source projects receiving them can't patch them fast enough. Andrej Karpathy joined Anthropic's pretraining team on 19 May.

Google reports 76.2% on Terminal-Bench 2.1, 83.6% on MCP Atlas, 84.2% on CharXiv Reasoning and an Elo of 1656 on GDPval-AA. All of these are company-reported. Google also reports that the model is about four times faster in output tokens per second than other frontier models. Gemini 3.5 Flash became the default model in the Gemini app and in AI Mode in Search. Google said a 3.5 Pro would follow the next month.

The model ships with new agent products. Antigravity 2.0 is a desktop app with no code editor at all. It's an agent manager that groups conversations by project and runs subagents in parallel and tasks asynchronously. It adds JSON hooks, scheduled agents, and slash commands such as /goal, which runs a task to completion, and /schedule.

The same Antigravity agent is now available as a Managed Agent in the Gemini API, in public preview. A Managed Agent is a stateful agent that Google hosts in an isolated Linux sandbox, where it plans, writes and runs code, manages files and browses the web. Google also announced Gemini Spark, a personal agent built on 3.5 Flash. It runs in the background of the Gemini app to handle email, scheduling and web errands. Trusted testers get Spark first, followed by a beta for Ultra subscribers in the US.

Earlier Google video models such as the Veo line were separate generators. Omni puts generation inside Gemini, so the model's reasoning carries across edits. Google says this keeps characters and physics consistent when a user asks for changes in a follow-up message. Clips have synced audio, are marked with SynthID, and are capped at 10 seconds.

Omni Flash rolled out to Plus, Pro and Ultra subscribers in the Gemini app and in Flow, and it's free in YouTube Shorts. Google calls it the first of a planned Omni family. Flow Music, Google's music workspace, also gained Omni-directed music videos and edits to specific sections of a song, such as changing the genre, instruments or lyric language. Separately, Lyria 3 Pro, which Google launched on 25 March and which writes songs of up to three minutes with verse and chorus structure.

The unit-distance problem asks how many pairs among n points in the plane can be exactly one unit apart. For about 80 years mathematicians believed constructions based on the square grid were essentially the best possible. The OpenAI model produced an infinite family of point sets that beats the grid by a polynomial factor, and it built them using algebraic number theory.

OpenAI says this was a general model with no special training for mathematics. A group of outside mathematicians checked the construction and wrote a companion paper. OpenAI calls it the first prominent open problem solved autonomously by AI. The model has not been released, and OpenAI hasn't said what it is or when it will ship.

Glasswing gives security partners access to Mythos Preview, Anthropic's unreleased model, to find flaws in software. The figures below are company-reported. Six independent firms or Anthropic assessed 1,752 of the findings and found that 90.6% were true positives. Of those 1,752, 62.4% were confirmed as high or critical.

Mozilla gave a third-party figure. Using Mythos Preview, it found 271 vulnerabilities in Firefox 150. That is more than ten times the count it found in Firefox 148 using Claude Opus 4.6.

The problem now is fixing the flaws. Of 530 high or critical bugs disclosed to open-source projects, only 75 had been patched, and a fix took about two weeks on average. Maintainers asked Anthropic to slow its disclosures. The report doesn't say how Anthropic will pace them from here.

• Qwen3.7-Max is Alibaba's new flagship model for long-running agents across hundreds or thousands of steps. It's available only through Alibaba Cloud and Alibaba's apps, and costs roughly $10 per million tokens according to VentureBeat.
• Command A+ from Cohere is an open-weights mixture-of-experts (MoE) model under Apache 2.0, with 218 billion total parameters and 25 billion active per token. Quantized to 4 bits, it runs on one B200 or two H100 GPUs. Cohere reports 85% on tau2-bench Telecom, up from 37% for Command A Reasoning.
• Stable Audio 3.0 from Stability AI is a family of four music and sound models trained on fully licensed data, with tracks up to 6 minutes 20 seconds long. The Small and Medium models have open weights under a community licence.
• Gemini for Science bundles Google Labs prototypes built on Co-Scientist and AlphaEvolve for generating hypotheses and for computational discovery.

• Andrej Karpathy, an OpenAI co-founder who founded Eureka Labs, joined Anthropic's pretraining team on 19 May to build a group that uses Claude to speed up pretraining research.