DeepSeek V4 goes official — peak-hour pricing doubles daytime rates
Pay double during Beijing business hours; plus Kimi K3 weights open July 27, Claude Code drops auto-review, and Cursor's agent breaks into Slack.
Copy markdown
DeepSeek V4 makes you pay by the clock
V4 is live now, and its official pricing adds peak/off-peak tiers: during Beijing business hours (9am-noon and 2-6pm), rates double — V4 Pro output climbs to roughly $1.74/M and Flash to $0.56/M, versus about $0.87 and $0.28 off-peak, all on a 1M-token context. The legacy deepseek-chat and deepseek-reasoner names deprecate July 24, so remap your calls now and consider batching heavy jobs into off-peak hours.
Kimi K3's open weights land July 27
Moonshot will release full weights for K3 — a 2.8-trillion-parameter MoE that activates just 16 of 896 experts with a 1M-token context, which it bills as the first open model in the 3-trillion class — after pausing new signups on capacity. It's already callable via the OpenAI SDK at api.moonshot.ai, and next week you'll be able to self-host a frontier long-horizon coding model.
Claude Code stops auto-running your review skills
As of v2.1.215, Claude Code no longer auto-invokes /verify and /code-review — you must call them explicitly, so any pipeline leaning on those silent gates now ships unchecked until you wire them back in. v2.1.216 adds a sandbox.filesystem.disabled setting (skip filesystem isolation while keeping network egress control) and fixes a quadratic slowdown that stalled long sessions.
Cursor's agent breaks out of its Slack channel
Cursor's Slack agent can now read from and post to other channels and threads, operate across multi-repo environments, and share its plan before executing — enough to wire a coding agent into a team's real Slack workflow instead of a single thread.