Anthropic has confirmed the secret internal testing rumors. In the latest release of Claude Code v2.1.280, Claude Opus5.5 (claude-opus-5-5) has been officially appointed as the default Opus model, putting an end to the recent rumor of "jumping levels." This new flagship model features a context window of up to 1 million tokens, with pricing that is surprisingly aggressive: $4 per million input tokens, $20 per million output tokens, and a cache read price of just $0.20 per million tokens. Looking back at the leaked pricing list from the beta phase, the prices are now exactly the same in the official version—this means that the intelligence of top-tier models is rapidly moving from luxury goods to everyday items.

A more aggressive move was made on the user subscription default tiers. The default model for Pro and Team Standard plans has been switched directly from Sonnet to Opus, aligning it with Max, Team Premium, and Enterprise tiers. In other words, users who previously used cheaper tiers and ran Sonnet by default have been quietly moved to the most expensive Opus model. The direction of the billing curve needs no further explanation. However, Anthropic has left a way out: effort levels saved before the /e effort model distinction will no longer apply to new models like Opus5.5; these start from the default tier until the user specifies otherwise. Opus4.7, 4.8, and Fable5 have also updated their rules, no longer strictly following their original default effort levels.

This update log is quite thick, almost like a small book. Most of the content focuses on those annoying edge experiences. In full-screen mode, the /skills list can now be scrolled with a mouse wheel, and the status toggle for a specific skill in /plugin can now be clicked with the mouse. An added CLAUDE_CODE_MAX_MCP_DESCRIPTION_LENGTH environment variable gives developers control over the 2048 character limit for each MCP server's tool description and instructions. Security barriers have also been improved in key areas: previously, when automatic mode encountered a security check rejection, it would keep retrying endlessly. Now, it simply rejects once and shows a message saying "retrying is futile." When the response is delayed, it no longer keeps rejecting blindly but instead backs off, retries, and stops after ten attempts with a prompt.

All the pitfalls in the writing chain have been addressed. If the model improperly includes path, file_text, file_content, or extra descriptions in the Write call instead of the standard file_path and content, validation no longer crashes immediately. Writing symbolic link paths now clearly tells you the real destination, and acceptEdits, allow rules, and automatic mode no longer approve writes outside the tree. The frustrating API error "role 'system' must precede an 'assistant' message" no longer causes every conversation to fail. When the advisor tag is not supported by the gateway, the request will remove the tag and resend. Format errors in MCP tools no longer cause sessions to fail both during each round and when using /compact. The cache for sub-agents has also been restored—when restoring a forked sub-agent, it rebuilds the list of tools it originally used rather than the one it re-sent, eliminating the bug that destroyed the prompt cache.

For users of non-Latin scripts like Chinese, a seemingly minor yet thoughtful fix deserves special mention: the invisible character cleanup feature used to delete zero-width non-joiners, which are crucial for languages like Persian and Arabic to attach suffixes (such as the plural of "PDF") to Latin words or numbers. This time, they have finally been spared. A long-standing issue with the Windows terminal displaying incorrectly after pressing Enter has now been fixed by redrawing the screen; awkward issues such as the microphone still recording secretly after Ctrl+C, Esc not canceling, and spacebar triggering voice transcription in vim NORMAL mode have all been resolved.

Both the VS Code and cloud platforms have added new dialogues. The editor now includes newly added panels such as /status, /sandbox, /chrome, /export, and /plan, and the /skills dialogue now shows the source, token estimate value, and switch status of each skill. The chat box now marks pasted content exceeding 800 characters or two line breaks, helping the model distinguish what you typed from what you pasted. A startup deadlock caused by the Python extension activation is now resolved after 60 seconds, allowing it to run independently without the Python environment. On the cloud side, the GitHub Enterprise Server cloud session now includes an eight-hour automatic token renewal, removing the previous obstacle of "gh calls failing after eight hours." The Claude Tag in Slack now includes a native Working indicator, Stop button, and thread title, and a brief notification will appear when guests join or leave the channel, explaining the changes in response methods.

From setting Opus5.5 as the default to quietly upgrading Sonnet users to the Opus tier, and then fixing a long list of bite-sized interaction issues, the signal released by v2.1.280 is clear: Anthropic doesn't just want the model to be smarter, but also to polish the entire call chain so developers won't want to leave. With top-level intelligence paired with a unit price of $4 and a million-token context, this battle for developers' wallets has already reached others' living rooms.