DeepSeek doubles peak-hour pricing, GitHub Copilot drops two Gemini models, mem0 fixes delete_all() leaving data behind

DeepSeek will charge 2x the regular price during peak hours (9:00–12:00 and 14:00–18:00 Beijing time) once the policy takes effect; no effective date yet. Batch jobs and anything latency-tolerant are worth shifting off-peak before the bill shows up.

As of July 31, GitHub Copilot dropped Gemini 2.5 Pro and Gemini 3 Flash across Chat, inline edits, and agent mode. Copilot Enterprise admins need to enable Gemini 3.1 Pro (Preview) and Gemini 3.6 Flash through policy settings; anyone still pinned to the old names loses access outright.

The deepseek-chat and deepseek-reasoner model names were retired on July 24; both now map to deepseek-v4-flash's non-thinking and thinking modes. Code still calling the old names is running against a renamed model, not the one it was tested against.

The Anthropic provider in Vercel's AI SDK now reports thinking tokens under reasoning token usage instead of wherever they landed before. Anything reading token usage to track cost, quota, or reasoning-mode spend should recheck its math after upgrading.

OpenAI is replacing Priority Processing with Fast mode; existing Priority Processing traffic migrates automatically. Behavior should carry over, but anything hardcoded to the old name in docs, dashboards, or billing exports needs updating.

mem0's Python delete_all() used to list the vector store once and stop, so accounts with more memories than a single page (most stores default to around 100) had the rest silently left behind. It now paginates in batches of 1000. Anyone who called delete_all() for compliance or privacy reasons before this fix should confirm the data actually left the store.

The Copilot app used to inherit its access rules from the Copilot CLI policy; it now has its own, with the same three options (enabled everywhere, disabled everywhere, or let organizations decide). Admins who disabled the CLI to keep the app off will find the app now needs its own policy set.

Neon now lets you grant access at the project level instead of only across the whole organization. If contractors or partial teams have had org-wide access purely because project scoping did not exist, that workaround is no longer necessary.

Vercel's AI Gateway can now pin requests to the US or EU with a single field, working across every provider behind the gateway, and returns the serving region on each response. Relevant if data residency rules have been blocking use of the gateway at all.

Kimi K3 is now priced on Together AI at $3.00 per million input tokens and $15.00 output, with a $0.30 cached-input rate. A concrete number to put next to whatever you are currently paying for a comparable model.