Opus 5.5 Becomes the Default, Plus a Prompt Audit Command

I'm Shannon, and this is the Claude Notes Brief -- Claude Code updates and Anthropic news for the week of Monday, September twenty-eighth. Claude Opus five point five is now the default Opus model. A new command audits your prompts for outdated instructions. Anthropic signed a twelve billion dollar compute deal with Akamai.

Let's start with Claude Code, where Claude Opus five point five arrived this week as the default Opus model, with a one million token context window and new pricing at four dollars per million input tokens and twenty dollars per million output. Cache reads come in at twenty cents per million. If you're on Pro or Team Standard, your default model also moves from Sonnet to Opus, which brings those plans in line with Max, Team Premium, and Enterprise. One detail worth knowing: effort levels became per-model recently, so any effort setting you saved before that change won't carry over to newly released models.

Opus five point five starts at its own default until you pick one. Anthropic published a post explaining why the model is priced and trained around long, context-heavy coding sessions, and we'll link that in the show notes. A bigger context window means older instructions can quietly go stale, and that's where the second change comes in -- there's a new prompt audit you can run that checks your instruction files, skills, agents, and commands for guidance written against older models. The report leads with stale paths, stale commands, and instruction files that contradict each other.

Staying in that same theme of keeping work in flight, pressing send now while tools are running no longer cancels the turn. Running tools move to the background instead, so your in-flight work keeps going while your new message goes through. Third-party providers picked up some catching up too. Following last week's support for the shared agents instruction file, that fallback now applies on Amazon Bedrock, Google Vertex AI, Microsoft Foundry, large language model gateways, and sessions with telemetry turned off.

Those same sessions now start in auto mode when you haven't configured a permission mode, and an explicit default permission setting still overrides it. On the editor side, the Visual Studio Code panel now answers typed commands with real dialogs -- status, sandbox, Chrome, export, skills, and plan. The skills dialog shows per-skill source, token estimate, and whether it's on or off. And navigation across the board got better: nearly every picker now supports page up and page down, home and end, the mouse wheel, and clicks, with scrollbars in fullscreen mode.

There's also a new setting that caps the width of Claude's prose in wide terminals, while tables and code blocks still get the full width. One more: a single attribution setting now turns off commit and pull request attribution entirely.

Under the hood, this was a week of session-resume repairs. A large batch of fixes stops continued and resumed sessions from re-sending earlier turns in a changed form. That had been causing the API to drop prior reasoning and break your prompt cache -- so if resumed sessions felt like they'd forgotten how they got there, that's the cause. Histories with unreadable thinking or advisor blocks now get repaired instead of failing every turn.

Responses cut short by a proxy are no longer shown as complete, and duplicated stream events no longer run tool calls twice. Speed improved alongside that. Interactive startup no longer blocks on the managed settings request, which cost about eighty milliseconds on a good day and seventeen seconds or more on an unreachable network. Print mode skips loading the interactive interface entirely, resuming very large or heavily compacted sessions is noticeably quicker, and first-request latency improves by reusing the preconnected API connection.

Now to the broader Anthropic news. Bloomberg is reporting that Anthropic committed roughly twelve billion dollars over seven years to Akamai for cloud capacity. Compute deals at that scale shape how much inference headroom sits behind the models you're calling. That capacity question connects to pricing, and the Financial Times has a piece on Anthropic and OpenAI both releasing cheaper models as a price war intensifies.

Opus five point five's pricing lands in the middle of that shift, and if you run large agent workloads, the cost curve is moving in your favor. On the policy side, two stories. The New York Times reports that a federal appeals court upheld the Pentagon's designation of Anthropic as a supply chain risk -- relevant if you work in or sell to federal environments. And The Guardian reports that Australian senators have summoned the heads of both OpenAI and Anthropic following agent-related security incidents.

Regulatory attention on autonomous agents tends to show up later as tighter default permissions and safety classifiers in agentic tools. That's it for the brief. I'm Shannon, and we'll see you next week.