Wiz's rolling AI-platform security tracker has new vendor-confirmed incidents and a second Fireworks intrusion; NEAR's SGLang exploitation shows why confidential enclaves do not secure vulnerable inference software.
LiteLLM's Private Registry is now available to confidentiality-bound customers for early critical security patches, while public fixes remain promised and open-source access policy is unresolved.
llama.cpp 1.0 removes 35 deprecated functions, adds FP8 KV caching and MoE expert caching, and restricts unauthenticated tools and private-address access. The new versioning promise helps downstream builders, but upgrading is not drop-in.
A specialised API offers cheap, bounded decisions for agent routing and classification, but v1beta integration, calibration and benchmark caveats matter more than the headline latency claim.
The premium Claude tier gets cheaper and faster, but the practical decision is when it beats Opus 5.5 on cost per successful task—and whether an integration can accept its API and data-retention requirements.
The price changes are not uniform: H100/H200 rise about 14%, B200 30%, B300 25% and GB300 about 11%. Builders using dedicated inference or training should re-run workload economics before assuming newer accelerators remain the cheapest route per completed task.
This is a hard capability removal rather than a routine model migration. Products built on OpenAI’s video-generation API now need another provider or a redesigned video path because the official deprecation table offers no successor endpoint.
WebKit's Safari MCP server is available in stable Safari 27.0 as well as Technology Preview. It runs locally, but captured page data passes to the connected agent; remote automation must be enabled explicitly.
Mistral Large 4 now has explicit GA and hosted token pricing alongside downloadable weights; self-hosters must still check the model-service licence exception and actual serving costs.
TanStack AI core has moved beyond its 1.0 GA: 1.4 adds MCP request metadata, 1.3 distinguishes loop limits from normal completions and earlier updates improve cancellations and runtime reliability.
PyTorch 2.15 is now generally available. FP8 and faster MoE decoding improve some Mac AI workloads, while the vendor/contributor speed claims remain configuration-specific rather than universal.
Kimi K3.1 Pro's full open-weight release is now public. Its 2.8T-parameter MoE has 104B active per token and a 1M-token context; local hosting requires serious infrastructure and checking Moonshot's specific licence.
Neon now manages OAuth discovery, token issuance and refresh for apps and MCP clients, but developers still host login/consent pages and verify access in their own resource servers. Shared signing keys across branches make issuer pinning essential.
The Zig headless browser's October 23 follow-up adds security and web-compatibility work that matters when agents browse untrusted pages. It still cannot replace Chrome for visual workflows.
The benchmark runs agents against Prometheus, Loki, Tempo and Grafana, checks actual outcomes across three attempts, and makes the task and grading code inspectable. Synthetic tests and vendor-run leaderboard scores still need local validation.
Computer Use is now GA in both Gemini APIs, but it is an action-planning tool, not a hosted remote desktop: clients execute clicks and keystrokes and must handle safety decisions.
New paid Gemini API projects now default to store=true on generateContent, streaming and batch calls; older projects retain their settings. Developers can opt out per request or project.
The new $0-subscription tier no longer lists starter credits, while three simultaneous requests are now permitted. Builders should check prepaid balance, model prices and queues before relying on the service.
OpenAI's new API credit grants can lower the cost of prototyping apps and agents, but eligibility, expiration, API limits and excluded Codex usage matter.