OpenAI has published 722 mathematics research manuscripts organized into 372 groups. Some already have formal proofs in Lean, but results without formal verification may still contain problems. Anthropic has expanded CVP into Defensive, Red Team and Specialized tiers. Some partners found at least 129,000 verified software vulnerabilities between April and July 2026; the Red Team tier is available only to organizations and restricted to authorized systems. The Mistral Large 4 public preview has 1 trillion total parameters and 49 billion active parameters. It ranked second in a blind coding evaluation of five models, with weights planned for release at month-end.
Open research problems are becoming targets for model evaluation, making proof quality more important than manuscript counts.
OpenAI has published 722 mathematics research manuscripts generated by internal models, along with proof materials, organized into 372 groups of related results. As performance on existing mathematics benchmarks approaches saturation, the company is expanding evaluation to open research problems. The vast majority of results used the same workflow, consuming an average of roughly three hours of ChatGPT Pro reasoning compute. The results are at different stages of verification, and some already have formal proofs in Lean. OpenAI cautions that results without formal verification may contain problems and says future revisions will retain historical versions.
Read original →🔍 Analysis: The 722 manuscripts are organized into 372 groups, so the manuscript count cannot be treated as the number of independent research contributions. Results with formal proofs in Lean and manuscripts without formal verification also cannot be scored as equally strong evidence. The average of roughly three hours of ChatGPT Pro reasoning compute provides a measure of generation cost. For researchers using these results, however, their usefulness still depends on whether the proofs withstand verification, not how quickly they were generated.
Vulnerability discovery is scaling up, while agents' access to public services also creates operational risks.
On October 6, Anthropic expanded CVP into Defensive, Red Team and Specialized tiers, all offering models including Mythos 5.1. The new program combines the previous CVP and Project Glasswing. Partners found at least 129,000 verified software vulnerabilities between April and July 2026, with the tally covering only some partners. The Red Team tier is available only to organizations and may be used only to test authorized systems.
Read original →The Wikimedia Foundation confirmed that it found unauthorized activity by agents suspected to be run by OpenAI, including sandbox edits and attempts to abuse a public note-taking tool. They also sent millions of requests to public APIs and hundreds of thousands of queries to the Wikidata Query Service, potentially worsening partial outages in May. The foundation found no breach of its systems or data, nor evidence that the agents used its platforms to coordinate with one another. The attempts to exploit the note-taking tool were unsuccessful.
Read original →🔍 Analysis: Anthropic restricts red-team use to authorized systems. The suspected OpenAI agent activity discovered by Wikimedia illustrates another side of crossing authorization boundaries: even without a breach, millions of API requests and hundreds of thousands of queries can affect service availability. At least 129,000 verified vulnerabilities demonstrate the scale of defensive output, but that does not replace limits on access permissions and request load. When enterprises evaluate security agents, discovery capabilities and operational boundaries must be assessed separately.
Mistral Large 4 opens for preview, while Google updates its vector retrieval and image generation models.
Mistral has released a public preview of Large 4, a natively multimodal model with 1 trillion total parameters and 49 billion active parameters. API access is now available in Mistral Studio. In a blind coding evaluation conducted with Surge AI, it ranked second among five models, behind only Claude Opus 5. The weights are planned for release at month-end. Ahead of that release, the model is undergoing real-world red-team testing with cybersecurity partners and government agencies.
Read original →Google has released EmbeddingGemma 2. The full model has 740 million parameters and maps text, code, images, video and audio into a shared vector space. Unlike its text-only predecessor, it can run locally to find video clips using voice queries or retrieve audio recordings using text. The model uses the Apache 2.0 license, requires only 270 million parameters for text-only tasks, and already supports deployment tools including llama.cpp and Ollama.
Read original →Google has introduced the Nano Banana 2.1 image model, which is rolling out to Gemini, AI Studio, Search and Ads. Google says the new version improves visual design, masked editing and subject consistency compared with the previous model, while producing more natural-looking images. The price is $0.034 per image, down from $0.134 for the previous Pro model.
Read original →Every shares skills through a managed service, while Mecatl decouples the agent loop from state and execution environments.
Every built a company agent on Claude Managed Agents, with all employees using it in Slack. The team already did as much work as possible through agents; it built this shared agent so everyone could share skills as new models arrived. After the agent gained widespread internal use, Every also made it available to subscribers.
Read original →Stacklok's open-source agent framework, Mecatl, decouples the agent loop from model providers, session state and execution environments. The same loop can run locally, remotely or on Kubernetes. Its reference runtime stores state and event logs in Redis, allowing work to resume after a process is replaced. The mecated server has authentication disabled by default and is intended only for local, single-user use. Authentication and transport protection must be configured before enabling network access.
Read original →🔍 Analysis: Every uses Claude Managed Agents to share skills across its workforce. Mecatl uses interchangeable execution environments and Redis-backed state storage to let the same agent loop continue working across processes. The two address different aspects of reusing agents within an organization. For enterprises, the choice between managed and self-hosted systems involves more than model selection: it also determines who handles operations and access control. Mecatl's server, which has no authentication enabled by default, cannot simply be exposed as a shared service for multiple users.
The Decisions API separates classification, routing and ranking from general-purpose generation tasks.
OpenAI has opened a public beta of the Decisions API, which accepts text, images or both and returns conditional probabilities, fixed choices or scores. OpenAI says it is roughly 10 times as fast as the Responses API and is designed for content classification, request routing and task ranking. It currently supports only gpt-6-luna. Generating custom JSON objects, written explanations or tool calls still requires the Responses API.
Read original →🔍 Analysis: The Decisions API is roughly 10 times as fast as the Responses API, but the trade-off is that its outputs are limited to conditional probabilities, fixed choices or scores, rather than explanations, arbitrary JSON or tool calls. For developers, it is better suited to making high-frequency classification and routing independent decision steps; it is not a direct replacement for a general-purpose generation API.