Claude Opus 5.5 lands at $4/$20 and breaks three API defaults, GPT-6 Sol and Luna get list prices, Supabase's Postgres minor release needs action

On Opus 5.5 thinking can no longer be turned off. Both thinking: {"type": "disabled"} and an explicit enabled block return 400, so omit the field and control depth with the effort parameter instead. tool_choice types any and tool also return 400, as on Fable 5.1, which means switching to auto with strict tool use. Computer use on the Claude API and Google Cloud now requires the computer_toolset_20260801 toolset, while computer_20251124 keeps working on Bedrock only. Fix all three before you point traffic at the new model.

Opus 5.5 lists at $4 input and $20 output per million tokens with $0.20 cache reads, and Opus 5 moves to the legacy section at $5 and $25. The new model ships with a 1M token context window by default, 128k max output, and always-on adaptive thinking. Fast mode is billed at 2x standard pricing, so fold that multiplier in before you budget for the 2.5x speedup.

Opus 5.5 watermarks its text output. Anthropic says the watermark changes neither the meaning, quality, nor readability of what comes back, and costs no extra tokens; Fable 5.1 and Mythos 5.1 already carry the same watermark. The disclosure surfaced in GitHub Copilot's model notes rather than a release note of its own, which is worth knowing if your product makes claims about AI-generated content or has to answer provenance questions.

OpenAI published standard pricing for GPT-6 Sol and GPT-6 Luna, quoted per million tokens for prompts up to 272K input. Sol runs $2 input, $0.20 cached input, and $10 output; Luna runs $0.10, $0.01, and $0.50. That is a twentyfold spread between the two, so decide how to route by task difficulty before arguing about the bill.

Copilot added four models in one weekly release. Opus 5.5 and GPT-6 Sol land on Pro+, Max, Business, and Enterprise; GPT-6 Luna and Grok 4.7 are available from Pro up. Grok 4.7 is billed at xAI's provider list pricing under usage-based billing rather than by premium-request multiplier, and the two schemes diverge quickly at volume, so check which one your account is on before you switch defaults.

Supabase moved to PostgreSQL 15.19 and 17.11 and flags that ltree, pgcrypto, btree_gist, and custom operators may need action. Minor releases are usually something you let happen; this one is not. If you use those extensions or define your own operators, read the changelog entry and run a regression pass before the upgrade reaches your project.

Weaviate's dedicated cluster card no longer shows the $400/month starting price and points to sales instead. The FAQ further down the same page still says the minimum runs from $400/month on Premium, so this reads as moving the negotiation earlier rather than as a price rise. Either way you can no longer put a public number for the dedicated tier into a comparison sheet.

Supabase now includes 20 GB of log ingestion per project and charges $0.50 per GB beyond it. Projects that leave debug-level logging on can cross that line without anyone noticing, so look at your current ingestion volume before deciding whether to tighten sampling or retention.

Cursor shipped two bots. Rollouts attaches a monitor to each pull request, posts its monitoring plan as a comment for you to edit, then checks your logs, metrics, and traces per environment after deploy and reports the change as verified healthy, regressed, or inconclusive. On a regression it names the change it suspects, notifies the author, and can open a revert PR, though it will not merge or roll back on its own. Security Review posts one comment per PR covering injection, auth bypasses, secrets committed to source, SSRF, and dependency changes with known vulnerabilities, while style and quality stay with Bugbot. Both are Teams and Enterprise only, with 10 days of trial credits worth roughly 50 changes on Teams and 500 on Enterprise.

Copilot code review settings go from Pro, Pro+, and Max only to every plan, Business and Enterprise included, and now live on their own code review page under Copilot settings in your profile, with separate controls for automatic reviews and review effort. Enterprise owners can configure it across the organization, which is the part that was missing while everyone set it up individually.

Pinecone Bring Your Own Cloud reached general availability, so you can run Pinecone inside your own cloud account. The signal here is the pricing-page banner, which replaced the earlier full-text-search GA notice, so read the linked announcement for region coverage and where the operational boundary sits. For teams whose data cannot leave their own account, this changes the shortlist.

mem0's Python SDK v2.2.0 adds User Profiles to MemoryClient and AsyncMemoryClient. A profile is a JSON summary of one user, shaped by a JSON Schema you configure per project and filled by an LLM from that user's memories, and it stays current as the memories change. Generation is asynchronous and every job POST carries an Idempotency-Key, so pass the same idempotency_key when retrying a lost request rather than starting a second job. Compared with assembling a profile yourself on each call, this saves a retrieval round.