Vercel AI SDK's Workflow 2.0 drops support for Workflow 4, Langfuse sets a November 16 sunset for API v3, GLM-5.3 launches with aggressive pricing

@ai-sdk/workflow's 2.0 release drops Workflow 4 support outright; anything still running on it needs to move to Workflow 5, which is itself only in beta. There's no compatibility shim, so either pin your current version until you've tested against the new one, or budget real migration time before touching package.json.

Langfuse has put a hard date on API v3's sunset: November 16, 2026. If your integrations still call v3 directly, that's about twelve weeks out, plenty of time to plan a v4 migration but not a deadline to forget about.

GLM-5.3 landed with concrete pricing, $1.4 per million input tokens and $0.26 output, plus a limited-time free tier, and Zhipu is pairing that with a coding-capability claim: a 50% gain over GLM-5.2 on its own benchmark and reported SOTA among open-source models on Terminal Bench 3.0. Vendor benchmarks deserve a discount, but the price alone makes this worth a bake-off against whatever you're running now.

DeepSeek shipped deepseek-v4-flash-vision-exp, its first vision-capable model in the V4 Flash line, alongside a full pricing breakdown: as low as $0.007 per million input tokens on a cache hit, up to $0.44 on a miss. If your workload needs image input and you've been paying GPT- or Gemini-level vision pricing, this is worth a side-by-side.

OpenAI cut GPT-5.6 Sol's list price 20% on input and 33% on output, landing at $4/$20 per million tokens through at least November 21. Vercel's AI Gateway and Replicate are separately running a 50%-off promo on top of that through mid-September, so anyone routing through those platforms is briefly looking at Sol for a quarter of what it cost a month ago.

Cursor's cloud agents can now subscribe to an event source, a PR, a Slack thread, a scheduled task, and wake up when something happens instead of waiting for you to re-prompt. Agents you spawn to open a PR now stay attached to it, fixing CI and answering bot comments on their own until it merges.

The new /goal command hands an agent a standing objective, "fix all flaky tests and make CI green," say, that it keeps working toward across a session instead of stopping after one pass. Pair it with a custom mode or /loop and you get something closer to an on-call agent than a one-shot assistant.

OpenAI is repositioning Codex from a coding assistant into a platform: an open agent harness that other developers can build custom agents on top of. It's a strategic shift more than a feature you'll use tomorrow, but it puts Codex in the same lane as Claude Agent SDK and Cursor's agent APIs.

GitHub Copilot for JetBrains picks up enterprise managed settings, plugin governance, MCP server access, OpenTelemetry, and permission modes, matching what VS Code admins already had. If your org standardized Copilot policy on VS Code and left JetBrains users unmanaged as a gap, that gap just closed.

The Admin API's user-management endpoints for Claude Enterprise, members, invites, groups, custom roles, are now GA. The anthropic-beta header is no longer required, so any provisioning scripts you built around the beta can drop that header and treat the endpoints as stable.

Together AI swapped its DeepSeek lineup, R1 and V3 are out, V4 Flash and V4 Flash 0731 are in, priced $6–$15 per million tokens. If you built against the old model names, check your calls still resolve before Together fully retires them.

Zep turned its Memory MCP Server into its own pricing line: 5 or 15 seats, or a custom tier, on top of the existing project and entity-type limits. It's a new dimension to budget for if you're evaluating Zep for MCP-based memory, not just an upsell on the same feature.