TabLOID

Google announced Gemini 4 Carbon on October 28, with initial access for selected cyber defenders and a gradual expansion through its Fairwind partner program. Wider availability is still ahead.

Google reports 88.2% on SEC-bench Pro for vulnerability discovery, 54.8% on ExploitGym and 71.0% on CWE-bench for patching. Those external benchmark runs used the model without guardrails. They are company-reported measurements, not a guarantee about a deployed system.

Its reported zero successful attacks on the Gray Swan indirect prompt-injection evaluation means zero in that tested set. It does not establish immunity to every future attack.

Introductory pricing is $2 per million input tokens and $10 per million output tokens. Google's stated post-introductory rates are $4 and $20. The practical question for most readers is access before price: the public launch has not happened yet.

Google's October 28 announcement, rollout and pricing terms

DeepMind's October 27 preview of Genie 4 adds native sound and real-time language controls to interactive environments. Users can steer a scene as it unfolds, changing action, atmosphere or camera position.

The company describes filmmakers trying scene ideas and game creators testing explorable concepts. These are early uses by trusted testers. Access for Google AI Ultra subscribers and early API partners is described as coming soon.

The fine print belongs in the picture: objects and visual details can drift during longer sessions or complex edits. Physical interactions, spatial accuracy and audio precision remain research problems.

For a designer, that makes this a potentially useful sketchbook with a soundtrack. A convincing scene still needs scrutiny before anyone treats it as a dependable simulation.

DeepMind's research preview and stated limitations

In an October 26 announcement, Axiom says it completed a Lean formalization of the classification of finite simple groups. The theorem organizes the basic building blocks of finite symmetry.

The company reports a 24-day effort producing approximately 43.1 million lines of proof code. Human mathematicians and engineers directed groups of AI agents. This was coordinated work built on a substantial mathematical literature.

Axiom says the translation exposed 34 noteworthy departures from published arguments, which required investigation and repair. Its account also describes discarding work that depended on unjustified assumptions.

The result is Axiom's reported computer-checked formalization. We have not independently rebuilt the full artifact. A formal checker can validate deductions from encoded statements; people still need to examine whether those statements express the intended mathematics.

Axiom's account, methods and proof links

Anthropic's October 27 research update describes work with six biotechnology partners and its acquisition of Mithril Medicine. Founder Cody Tranbarger used Claude to coordinate work with 11 outside labs.

The company says a muscular-dystrophy program produced promising mouse results in under 18 months. Human trials could begin as soon as late 2028 if the candidates continue to hold up. That is a conditional development target, not an available treatment.

Other examples concern the paperwork around science. Anthropic says Claude produced a first draft of one Enveda regulatory document in four hours, compared with about three months manually. A first draft is not a completed, approved filing.

The company explicitly describes the work as early and mostly preclinical. Faster experiments and administration could help researchers; evidence of safe, effective treatment in people still has to be earned.

Anthropic's case studies and technical-report links