Anthropic cancels the Claude Sonnet 5 price increase, DeepSeek switches to peak/off-peak pricing today, Neon acquires the Electric team

The footnote on Anthropic's pricing page reading 'introductory pricing of $2/$10 per million input/output tokens through August 31, 2026; $3/$15 standard pricing thereafter' is gone, and this is not page cleanup: the docs pricing page now states that $2/$10 is Claude Sonnet 5's standard price and the increase scheduled for September 1 will not occur. Teams that budgeted for the 1.5x bump or planned migrations to dodge it can drop those plans.

DeepSeek's API moves from a vague "pricing will rise" warning to a concrete peak/off-peak schedule: peak hours (01:00–04:00 and 06:00–10:00 UTC) cost double the off-peak rate, with the new prices taking effect at 16:00 UTC today, August 16. Workloads that can shift outside those four-hour peak windows get cheaper inference for free; batch jobs and anything latency-insensitive are the obvious candidates to reschedule.

Neon is acquiring Electric, the team behind the Electric sync engine and PGlite, folding real-time sync and in-browser Postgres into Neon's stack. Teams using Electric standalone should watch for roadmap changes now that its team answers to a different owner; Neon users get a credible sync story without adding a separate vendor.

Zep pulled the public pricing for its "Emerging Companies" tier, previously $13,000/year for startups that raised $1M–$10M with SOC 2, HIPAA BAA, and guaranteed rate limits included. The page now just invites you to "get Zep Enterprise at an emerging company price," meaning that number is no longer something you can quote without talking to sales first.

Anthropic quietly narrowed the scope of its Claude Haiku 3.5 retirement notice: the model is only dead on the Claude API, not on Amazon Bedrock or Google Cloud, where it keeps running. Anyone who read the original notice as a full retirement and started migrating everywhere can stop the Bedrock/GCP portion of that work; only direct Claude API callers still need to move to Claude Haiku 4.5.

CrewAI bumped its torch dependency to 2.13.0 to close a security vulnerability. The release notes don't name a CVE or describe the exploit, but pulling the update is a routine dependency bump, not a breaking change, so there's no reason to wait for more detail before applying it.

GitHub Copilot's memory feature, retaining context across chat sessions, is now available in JetBrains IDEs, not just VS Code. It's a toggle in the Copilot settings portal, off by default per prior rollout pattern, so JetBrains users who want it need to turn it on rather than assume it's already running.

Together AI added Qwen3.8-2.4T-A95B to its catalog at $2.50/$6.25 per million input/output tokens, with a $0.50 cached-input rate. It's a large model at that parameter count, worth benchmarking against whatever you're currently running on Together if quality per dollar on long-context or complex tasks matters to your workload.

Vercel Connect now exposes line-level visibility into the token lifecycle: filter events, track metrics, and forward data to any log drain. If you've been debugging token usage in Connect by guesswork, this closes that gap without needing a separate observability integration.

DeepSeek-V4-Pro now supports the Responses API, which was previously limited to deepseek-v4-flash. Anyone who picked flash specifically to get Responses API compatibility can now move back to Pro without giving up that interface.

The Compliance API now covers local sessions from Cowork and Claude Code, not just cloud-hosted ones, through three new beta endpoints for listing sessions, fetching metadata, and pulling transcripts. Claude Enterprise organizations doing audit or data-retention work finally get visibility into sessions that were running on employees' own machines.

Claude Code fixed an MCP OAuth sign-in bug that broke authentication for servers using a pre-registered OAuth client, Slack among them, due to a redirect URI mismatch. If MCP OAuth sign-in to Slack or similar servers was failing before, it's worth retrying after upgrading rather than assuming it's a config problem on your end.

The Claude API now returns an anthropic-workspace-id response header carrying the requesting key or token's workspace ID, including the default workspace. Organizations routing logs or costs by workspace can now read that directly off every response instead of maintaining their own key-to-workspace mapping.