Gemini 3.7 Flash goes GA, GitHub Copilot adds a $100 Max plan, Zep moves to monthly billing

Gemini 3.7 Flash (gemini-3.7-flash) is generally available as of August 13: Google calls it its most capable workhorse model for coding and agentic workflows, and it runs at an introductory $0.75 per million input tokens through December 31, doubling to $1.50 from January. Two weeks late here because the Gemini pricing page randomly switches languages and our quarantine rule treats everything from that source as suspect until verified by hand.

The gemini-omni-flash-preview endpoint will be shut down on September 30 in favor of the GA release. Anyone who built against the preview endpoint has a month to move; this is the same pattern as the robotics preview retirements, where the preview name simply stops resolving.

GitHub Copilot's individual lineup now shows four plans: Free, Pro at $10, Pro+ at $39, and Max at $100 per month. The differentiator worth reading closely: delegating tasks to third-party coding agents, Claude by Anthropic and OpenAI Codex among them, sits in Pro+ and Max only, while Pro keeps code review and unlimited completions. If third-party agents matter to you, the effective entry price is $39.

Zep switched its plans from annual to monthly billing: what was $1,250/year ($104/month billed annually) is now $125/month billed monthly. No annual commitment to front anymore, at roughly 20% more per month; teams that were holding off because of the yearly lock-in have a cheaper way to trial it now.

GLM-5.3-Flash launched with a 50% discount running through September 9: $0.075 per million input tokens against a $0.15 list price, $0.25 output against $0.50, plus a limited-time free allowance. The non-Flash GLM-5.3 also landed on Together AI's price list this week, so the family is now reachable outside the first-party API too.

Cursor's Cloud Agents no longer require a connected GitHub or other SCM provider: you can prompt immediately and save the work to a Cursor Origin repo, wiring up your own repository later. Lowers the barrier for trying cloud agents on throwaway work before pointing them at real codebases.

Anthropic's Admin API is now callable from the ant CLI and seven SDK languages (Python, TypeScript, C#, Go, Java, PHP, Ruby) under client.beta.organization, covering organization info, members, and invites. Teams that were scripting raw HTTP against these endpoints can drop that glue code.

The Compliance API's session endpoints are out of beta for Cowork and Claude Code sessions, including session transcript retrieval. This completes the local-session visibility that issue #5 covered as a beta: audit and data-retention pipelines can now build on stable endpoints without beta headers.

The @braintrust/vercel-ai-sdk package is deprecated, with v0.0.7 as the final release. Projects wiring Braintrust into the Vercel AI SDK through this package need a migration plan; there will be no further fixes, so staying pinned only works until something around it moves.

Neon put a price on Object Storage: 5 GB included, then $0.023 per GB-month, though nothing is charged while it stays in beta. The number matters now because it lets you cost out beta workloads before the meter turns on.

Neon added three CLI commands that set it up inside coding agents like Cursor and Claude Code without manual configuration. If your agents provision or query Neon databases, setup drops from config-file editing to a one-liner.

CrewAI promoted conversational flows to stable, one week after shipping them as a new feature. Teams that held off on the feature-flagged version can now adopt it on a stable interface; note last week's caveat still applies, conversation mode is off by default.