Capability arrived with a different scarce resource in every story. Mythos produced findings faster than defenders could triage them; Anthropic committed billions to GPU supply; a geometry proof gained credibility through expert review; and Google's hosted agents, generated Search interfaces, and video model transferred execution, validation, and provenance into one platform.
1. Security practitioners put Mythos risk in operational context
Cybersecurity practitioners told Reuters that Claude Mythos Preview is a genuine advance in vulnerability discovery but does not instantly grant unsophisticated attackers novel capabilities. The model lowers prompting effort and scans code faster, while substantial compute, a capable harness, and experienced operators still shape end-to-end results.
The nearer bottleneck is verification and remediation. More candidate flaws can overwhelm triage and patching, so a raw discovery count says little about reduced exposure. Time-to-reproduce, time-to-patch, and deployed-fix coverage connect a model finding to an actual security outcome.
Sources: Reuters analysis of Mythos's practical impact
2. Anthropic commits $1.25 billion a month to SpaceX compute
SpaceX's regulatory filing disclosed that Anthropic agreed to pay $1.25 billion per month through May 2029 for GPU capacity at the Colossus and Colossus II data centers. WIRED reported that a reduced fee applies initially before the annualized commitment reaches roughly $15 billion.
The contract turns compute access into a strategic dependency on infrastructure controlled by a company with its own competing AI lab. For enterprise buyers, the scale is also a reminder that model pricing rests on long-lived capacity obligations, not only marginal token costs or short-term demand forecasts.
Sources: WIRED on the Anthropic-SpaceX compute contract · Anthropic on its SpaceX capacity agreement
3. OpenAI model produces a checked counterexample in discrete geometry
OpenAI says an internal general-purpose reasoning model autonomously disproved a longstanding conjecture about the maximum number of unit-distance pairs among points in a plane. The construction uses algebraic number theory to establish an infinite family with a fixed polynomial improvement over the previously expected growth rate.
External mathematicians reviewed the proof and published companion remarks, providing stronger evidence than a self-graded model answer. The result is narrow, but its method is consequential: a model connected tools from distant mathematical fields without a problem-specific training system, while human experts supplied verification and interpretation.
Sources: OpenAI's unit-distance result and proof · External mathematicians' companion remarks
4. Gemini API adds managed agents in isolated Linux environments
Google launched Managed Agents in the Gemini API as a preview, allowing one API call to provision an Antigravity agent inside an isolated, ephemeral Linux environment. The agent can browse, execute code, manage files, call tools, and retain environment state across follow-up interactions.
Custom behavior can be packaged in versionable AGENTS.md and SKILL.md files rather than a bespoke orchestration service. That reduces infrastructure work but transfers meaningful controls to Google, making sandbox boundaries, persistence, network access, observability, and pricing part of the platform evaluation rather than implementation details.
Sources: Google's Managed Agents announcement
5. Google Search begins generating interfaces and mini-apps
Google said AI Mode had passed one billion monthly users and was doubling usage each quarter as it announced deeper agentic features for Search. Gemini 3.5 Flash and the Antigravity harness will generate interactive explanations, dashboards, and task-specific mini-apps from search prompts, with staged rollouts beginning this summer.
The shift changes both interface and web economics: an answer can become generated software assembled from multiple sites without foregrounding their links. Source visibility, generated-code reliability, personalization boundaries, and persistence can be measured separately, exposing whether a useful mini-app also obscures the publishers and transformations behind it.
Sources: Ars Technica on Google's agentic Search plans · Google I/O keynote
6. Gemini Omni launches conversational video generation and editing
Google introduced Gemini Omni Flash, which accepts combinations of text, images, audio, and video and initially produces video. Users can refine scenes over multiple conversational turns, alter actions or camera angles, and carry reference characters, motion, or style into a generated clip.
The model began rolling out through the Gemini app, Flow, and YouTube Shorts, with APIs promised later. Google says every generated video carries SynthID. Export tests can check temporal consistency, edit fidelity, identity retention, provenance visibility, and policy enforcement after a clip leaves Google's own interface.
Sources: Google's Gemini Omni announcement