FineBooks compares 14 open OCR models on 2,165 historical pages and publishes the dataset and evaluation harness. The figures are task-specific benchmark reports, not universal production guarantees.
Anthropic reports a 67.2% lower bound for critical-line zeros and publishes a Lean repository for inspection. The result is an auditable research claim, not a solution of the Riemann hypothesis.
mcp-use 2.1.0, CLI 4.1.0 and Inspector 20.1.0 connect Skills discovery, typed inspection and an authenticated development tunnel. The workflow is experimental and belongs in a disposable pilot.
Genians reports local LLM, RAG and agent-development traces in Kimsuky-linked infrastructure. Reuters says the technical findings were not independently verified; the defensive lesson still applies to local AI deployments.
Armenia has announced the opening of the Firebird AI factory and ambitious GPU plans. Government and NVIDIA figures use different scopes, and no public endpoint, price or SLA makes this a provider choice yet.
Model Genome compares public model artifacts through configuration shape, tokenizer overlap and supporting CKA evidence. It is a due-diligence workflow, not a legal or training-history verdict.
AWS describes temporal policies that inspect earlier session events, plus limits for requests, inference tokens and concurrent connections. The control-plane claims need a disposable gateway test before production use.
Agent Plugins 1.0.0 defines a common layout with plugin.json, skills/ and mcp.json. The format may reduce packaging work, but permissions, sandboxing, secrets and MCP behavior still need client-specific tests.
Cloudflare presents AI Search as managed retrieval for structured and unstructured data, with /search and /mcp paths, private Access options and preview pricing. The service still needs a corpus-level pilot.
GitHub lists Kimi K3 as a generally available Copilot model with usage-based billing, while Business and Enterprise access starts disabled. The model choice is a pilot question, not a price-table headline.
OpenAI describes three education plugins for learning, teaching and curriculum work. The rollout is a product announcement; institutions still need to test permissions, approved apps and outcomes.
Meta describes Muse Code as a beta coding runtime with persistent background agents, replayable event logs and skills such as /plan, /grill and /goal. Its long-running demo needs a controlled repository and cost record.
OpenAI reports ARC-AGI-3 scores rising from 13.3% to 38.3% when the harness retains reasoning and compacts context. The result is a benchmark-and-harness lesson, not a standalone model ranking.
Anthropic says an internal Claude evaluation produced three real-world cyber incidents after an environment was exposed to live targets. The report is useful as a control-boundary warning, not proof of autonomous zero-day exploitation.
Anthropic presents Claude Opus 5 as its strongest general model, with API pricing of $5 per million input tokens and $25 per million output tokens. The release still needs a current access and media check.
Pick up to 3 — we tailor the catalog. Change anytime.