Cohere released Embed 5 as Cloudflare expanded its agent tools

TOP 3

Embed 5: Cohere’s September 30 release says Pro and Fast share an embedding space, so an index built with Pro can be queried with either model. Cohere lists API prices of $0.12 and $0.08 per million tokens, respectively; its retrieval comparisons are vendor evaluations.
AREX-2: BAAI published model weights and a card for an agent model trained for feedback-driven, multi-round improvement. The model card describes its training and evaluations; those results are the authors’ claims.

Routing and sandboxes: Cloudflare’s Auto Router routes compatible AI Gateway requests sent to cloudflare/auto. Cloudflare reports up to 30% cost savings in its own early OpenCode use, not a general guarantee. Containers now allow code to choose a sandbox image and instance type at runtime; filesystem snapshots are in public beta.
Production handoffs: Cloudflare Workers Issues is in open beta. It groups failures and can send logs, traces and version context to a configured coding agent; it is not a claim that the agent fixes the bug autonomously.
Agent payments: Cloudflare’s Monetization Gateway is in closed beta, allowing enrolled site and API owners to charge agents for access, including to MCP tools and datasets.

SDK work, not a protocol release: A TypeScript MCP client commit adds one SSE connection retry after onUnauthorized() resolves; this is a repository change, not a new SDK version.
Pi integration: Earendil Engineering says Pi now supports MCP in its core and explains its interpreter-based approach to composing tool calls. The explanation was posted September 29; the separate community discussion below surfaced September 30.

OpenClaw: The 2026.9.7 release notes describe backups before state and agent-database migrations, rollback changes, and a plugin for OpenAI’s Agents API.
Local inference: A merged llama.cpp pull request adds GLM-5.3-Flash text and vision support. Its author notes that MTP is separate and multiple sequences require --kv-unified.
Agent evaluations: The UserProxyBench preprint reports that varying only the simulated user changed mean reward by 15.2 points across 375 tasks. That is the authors’ study result, not a general benchmark correction factor.

MCP debate: A September 30 Hacker News thread on Pi’s MCP support, covered above, argued over whether MCP tools, resources and prompts compose well across clients. These are developer views, not changes to the MCP specification.