Cursor launches Projects, Gemini 3.7 Flash clears review as GA, GLM-5.3-Flash pricing reverts to full rate

Cursor's new Projects feature lets you hand off work that spans months, such as a full feature or a migration, to a coordinator agent that delegates tasks across thousands of subagents and keeps working without repeated prompting. It is a step beyond single-session coding agents toward something closer to an ongoing engineering process.

Gemini 3.7 Flash is now confirmed generally available, positioned as Google's workhorse coding and agent model with broad improvements over the previous Flash generation. Introductory pricing runs through December 31, 2026, worth benchmarking against whatever you use today before that window closes.

GLM-5.3-Flash's launch discount ended on September 9; prices are back to the full list rate of $0.15 / $0.03 / $0.50 per million tokens for input, cached input, and output. Anyone who sized a workload around the discounted rate is now paying double.

Mem0's Pi Agent Plugin v0.3.0 removes the Dream consolidation and pin commands along with their skills, config, types, and exports; only remember, search, forget, tour, scope, and status remain. If your integration still calls the removed commands, it breaks on upgrade rather than degrading gracefully.

GitHub Copilot deprecated MAI-Code-1-Flash across Chat, inline edits, agent mode, and code completions on September 10; the suggested replacement is MAI-Code-1.1-Flash. Enterprise admins may need to enable the new model explicitly through model policies before anyone on the team can switch.

GitHub Copilot Business and Enterprise admins can now centrally control which agent operations (shell commands, file reads and edits, network domains) are blocked, require approval, or run freely, consistently across the Copilot app, CLI, and VS Code. That closes a gap where this kind of control depended on what each developer configured locally.

Claude Managed Agents' permission policies gained an auto mode: the server itself evaluates each agent or MCP tool call and runs, denies, or pauses it for approval, reporting how the call was evaluated in the agent.tool_use and agent.mcp_tool_use events. That moves permission enforcement off the client and into infrastructure you don't have to run yourself.

Claude's Enterprise Admin API, Analytics API, and Compliance API now require the anthropic-version header on every request, matching the rest of the Claude API. Requests without it will start failing, so check any admin tooling or scripts that call these endpoints directly before they break unannounced.

DeepSeek's model naming settled on deepseek-flash (legacy deepseek-v4-flash and deepseek-v4-flash-vision-exp names still work, mapped to the same DeepSeek-V4.1-Flash backend), which now also supports a 1M token context window. Separately, DeepSeek walked back its plan to retire deepseek-v4-pro on September 14: after user pushback, it will keep serving V4 Pro at unchanged pricing instead.

OpenAI's Prompt Cache Diagnostics is now generally available in the Responses API for GPT-5.6 and later: it compares cache reuse against a previous response and points out why a cache miss happened. If your prompt costs depend on cache hits, this turns a guessing game into something you can actually debug.

Together AI's GPU Clusters now offer preemptible compute, the same hardware at a flat 50% of the on-demand rate, with a five-minute drain window before a job gets interrupted. That's a meaningful discount for training or batch jobs that can tolerate a pause, less so for anything that needs to run uninterrupted.

Neon's new Claimable Neon lets an AI agent provision a database mid build and hand the claim link to a human afterward, instead of forcing the agent through a signup and email flow it can't complete on its own. It's a small piece of infrastructure aimed squarely at agent-driven development, not humans clicking through a console.

Claude Code added a maxEffortLevel setting, set globally or per model under modelSettings, that caps how much effort every provider (including Bedrock, Vertex, and Foundry) is allowed to spend, while individual users can still pick something lower. For teams managing cost across a mixed-provider setup, that's a ceiling you control centrally instead of trusting everyone to dial it down themselves.