Claude Sonnet 5.5 and Haiku 5.5 break old request shapes, Copilot retires six models, Kimi K2 and Opus 4.8 fast mode get end dates

Sonnet 5.5 is now the default Sonnet on the Anthropic API, with a 1M context window at $2/$10 per million tokens and $0.20 cache reads, and code written for Sonnet 5 can break in five ways. To skip up-front thinking, send thinking: {"type": "between_tools"} instead of "disabled", which applies at high effort or below. tool_choice types any and tool return 400, the older computer_20251124 tool is rejected on the Claude API and Google Cloud, thinking blocks are tied to the model and the conversation, and the advisor tool no longer accepts Opus 4.8, Opus 4.7 or Sonnet 5 as advisors. Work through all five before you switch models.

Haiku 5.5 becomes the default Haiku on the Anthropic API with a 1M context window, priced at $0.10/$0.50 per million tokens for prompts up to 100K tokens and $0.50/$2.50 above that. Manual extended thinking with budget_tokens now returns 400 and adaptive thinking is on by default, so Haiku 4.5 code that sets a thinking budget fails on its first request. The long-prompt tier costs five times the base rate, which matters if you send large contexts to Haiku because it is the cheap option.

On October 19 Copilot deprecated six models across chat, inline edits, ask and agent modes and code completions: Gemini 3.7 Flash, GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini and Grok 4.5. The suggested replacements are Gemini 3.8 Flash, GPT-6.1 Sol and GPT-6 Luna. This follows an October 2 round that removed Gemini 3.5 Flash, Gemini 3.6 Flash, Kimi K2.7 Code and Claude Opus 4.7. Enterprise admins need to enable the replacements in model policy, and any workflow that names an old model needs updating.

Fast mode for Claude Opus 4.8 is deprecated and will be removed on November 30, 2026. After that, requests to claude-opus-4-8 with speed: "fast" return an error, and Anthropic's stated path is Opus 5.5. That leaves about three weeks, and Opus 5.5 has breaking changes of its own (see issue 11), so schedule the migration now rather than at the deadline.

From October 31, 2026, Kimi's K2 series is closed to newly registered users, and all K2 models are scheduled to retire by the end of December 2026. Existing accounts keep working until then. Kimi's docs now point new users to K3.1 Pro, and the K3 series gained a 1,048,576 token context window and Batch API pricing at 60% of the standard rate, so the migration target is cheaper to test than it was.

Supabase CLI v3 is flagged as a breaking change: commands now target the local stack by default and ask for confirmation before changing a hosted project. The scripts to check are the ones that ran against a hosted database without a prompt, because an extra confirmation can stall or abort a non-interactive run. Before you bump the CLI version in CI, read the release notes for how to pass that confirmation non-interactively.

Pinecone Local, the in-memory Docker emulator for development and testing, has reached end of life and its documentation has been removed. Pinecone recommends a separate project in your Pinecone account instead, which means tests that started the container in CI now need network access and credentials. Pinecone also added guidance on mocking its API for tests that should not touch a real project.

LangChain will drop Python 3.10 in langchain-core 1.8.0 and langchain 1.5.0, which makes the current minor versions the last to support it. Nothing breaks today, but environments on 3.10 will stop receiving new releases once those versions ship. If your runtime image is still on 3.10, move it to 3.11 or later, or pin the current minors until you can.

OpenAI-hosted sandbox compute is now billed only during active turns, so idle time between turns no longer costs anything. Long-lived agent sessions that sit waiting on a user or an external event were paying for that wait, and those are the bills that drop. If you built your own teardown logic to avoid idle charges, keeping sessions open may now be the simpler option.

Existing Vercel Pro teams now pay $0.10 per GB-month for Deployment Storage and Functions Storage, and the retention window moved to 30 days on October 23. Projects with many preview deployments or large function bundles accumulate this without anyone noticing, so check the usage page and prune deployments you no longer need.

Mem0 replaced its Hobby, Starter ($19/month), Pro ($249/month) and Enterprise tiers with two plans, Free and Custom. Free gives 10,000 adds and 1,000 retrievals a month for the first 30 days with model calls included, or 50,000 adds and 10,000 retrievals with your own LLM key, and usage past those limits is billed up to a monthly cap you set. Anything beyond that goes through sales, so teams on the old Pro plan should ask what their plan becomes.

Copilot Fusion is generally available. Instead of picking a model per task, you choose Fusion and Copilot routes and orchestrates the models itself. Each model it calls is billed at that model's standard rate, with a 30% discount on model costs for paid plans for a limited time. Because the bill depends on which models Fusion picks, compare a week of usage against your usual model before you make it the default.

Anthropic launched user profiles, which tag each API request with the end customer it serves. For platforms, gateways and resellers, a safety enforcement action can then apply to that one customer instead of the whole organization, and the SDKs and the ant CLI support the new /v1/user_profiles endpoints. If many customers share one organization, this limits how far one abusive tenant can affect the rest.