本周 Claude Code、Codex 与 Gemini CLI 更新汇总（2026 年 9 月 27 日当周）

Originally published at https://aicoding-guide.com. The headline this week: Claude Opus 5.5 (claude-opus-5-5) is now Claude Code's default Opus model (v2.1.280). The official pricing table lists it at $4/MTok input and $20/MTok output, below Claude Opus 5's $5/$25. On the permissions side, the same release fixed writes through a symlinked path being approved against the wrong location. On the settings side, "attribution": false became a valid shorthand, and auto mode's classifier moved to the server. Covered here: Claude Code 2.1.278 and 2.1.280 through 2.1.283, Codex CLI 0.156.1 and 0.157.0, and Gemini CLI 0.61.0. Key point What you will learn Claude Opus 5.5's model ID and the prices you can verify in the official table The symlink write fix and what changed in auto mode Codex adding GPT-6 Sol and Luna, and Gemini CLI 0.61.0's security fixes Claude Code 2.1.280: Claude Opus 5.5 becomes the default Opus model The changelog reads: "Added Claude Opus 5.5 (claude-opus-5-5), now the default Opus model — 1M context, $4/$20 per Mtok with $0.20/Mtok cache reads". The official pricing page lists the same numbers. Item Claude Opus 5.5 Claude Opus 5 Model ID claude-opus-5-5 claude-opus-5 Input $4 / MTok $5 / MTok Output $20 / MTok $25 / MTok Cache reads $0.20 / MTok $0.50 / MTok 5-minute cache writes $5 / MTok $6.25 / MTok The pricing page notes that cache hits on Opus 5.5 cost 0.05x the base input price, where most other models use the standard 0.1x multiplier. The longer your sessions run, the more that gap matters. Because it is now the default Opus model, a setup that simply selects opus in /model switches to 5.5 after the update. To pin a version, name the model ID in /model or set ANTHROPIC_DEFAULT_OPUS_MODEL. If your organization pins which models can be used 2.1.283 added the availableModelsMatch managed setting. With "exact", an availableModels entry allows only the model version it names, so new releases stay blocked until you list them. A companion deniedModels setting blocks specific models even when availableModels allows them. 2.1.280: writes through a symlinked path Three permission-related fixes landed together. Writes through a symlinked path were judged by their in-tree spelling rather than where they landed. The prompt now names the real destination, and acceptEdits, allow rules and auto mode no longer approve a write that lands outside the tree. Auto mode retried an action over and over when a safety check declined to review it. The action is now denied once, noting that retrying won't help. Auto mode also denied actions over and over without pause when a safety check gave no answer. Retries now back off, and the turn stops with a message after ten in a row. If your allow rules are written around paths that look in-tree, writes could previously land somewhere you did not intend. For how those rules are written, see Configuring permissions in Claude Code's settings.json. 2.1.281: "attribution": false as a shorthand You can now write "attribution": false in settings.json to hide all commit and PR attribution at once. { "attribution": false } Enter fullscreen mode Exit fullscreen mode The changelog attaches a caveat: older CLI versions skip a settings file that holds it, so keep the object form in files shared across versions. If your team commits .claude/settings.json, stay with the object form below until you can confirm everyone is on 2.1.281 or later. { "attribution": { "commit": "", "pr": "", "sessionUrl": false } } Enter fullscreen mode Exit fullscreen mode Each sub-key is explained in Removing Co-Authored-By from Claude Code commits. 2.1.278: auto mode's classifier moved to the server For Claude API and Enterprise users, and on Bedrock, Vertex, Foundry and gateways, auto mode now defaults to the server-side classifier. Per the changelog, that classifier does not charge for its own overhead, and you get a warning if the session falls back to a billed path. On Bedrock, Vertex, Foundry and gateways you can opt out with CLAUDE_CODE_AUTO_MODE_SERVER=0. /status gained an "Auto mode server" row showing whether this session's classifier runs on the server. For where auto mode sits among the permission modes, see Changing Claude Code's default permission mode. 2.1.282 and 2.1.283: prose width and prompt auditing A maxProseWidth setting caps the width of Claude's prose in wide terminals. Tables and code blocks keep the full width. /doctor prompt-audit (also /checkup prompt-audit) audits your CLAUDE.md files, skills, agents and commands for prompting patterns written for older models. 2.1.282 fixed every request failing with a 400 error in conversations whose history holds web search results the API cannot decrypt — for example, from a turn answered through a third-party gateway. 2.1.280 added CLAUDE_CODE_MAX_MCP_DESCRIPTION_LENGTH, which changes the 2,048-character cap on MCP tool descriptions and server instructions for every MCP server in the session. Codex CLI 0.156.1 and 0.157.0: GPT-6 Sol and Luna 0.156.1 was a hotfix that lets you "choose GPT-6 Sol or GPT-6 Luna from the model picker", with the rate-limit switch prompt now recommending GPT-6 Luna. 0.157.0 follows with "Added GPT-6 Sol and Luna, including Amazon Bedrock support and migration prompts for older models". 0.157.0: changes to how you drive the CLI From the release notes' New Features, the items that touch day-to-day use: Change What it does Fullscreen transcripts Enabled by default; Shift-click extends a text selection Background server Starts automatically for eligible interactive sessions, with recovery choices when server settings are incompatible f shortcut Forks a conversation open in another app, preserving drafts and queued prompts /import Now available in remote sessions and local background-server sessions Terminal rendering Unicode bullets, checkboxes, aligned equations and optimization notation Among the bug fixes, network restrictions are now enforced across redirects and on ongoing HTTP and WebSocket traffic, with cancellation when a policy change revokes access. Configured proxies are also honored for realtime connections and standalone web search, including search redirects. Gemini CLI 0.61.0: almost entirely security work v0.61.0's release notes list four substantive changes, all of them fixes. Prevent indirect prompt injection via build file modifications and untrusted flags (#29250) Harden sandbox filesystem boundaries and isolate runtime state (#29214) Preserve explicit versioned Flash model IDs (#29252) — a version-pinned ID was being resolved to something else. Ensure AgentLoopContext properties survive an object spread (#29335) There are no new features. The release notes carry no detailed description of the fixes, so exactly which attack paths were closed could not be confirmed from the official release notes. If you run Gemini CLI without the sandbox, this is a good moment to turn it on. What to check now Update Claude Code to 2.1.283 or later, which includes the 2.1.280 symlink write fix. If you select opus in /model, confirm that landing on Claude Opus 5.5 is what you want. To pin a version, name the model ID or set ANTHROPIC_DEFAULT_OPUS_MODEL. Before writing "attribution": false into a shared .claude/settings.json, confirm everyone is on 2.1.281 or later. Keep the object form if versions are mixed. If you use auto mode on Bedrock, Vertex, Foundry or a gateway, check the "Auto mode server" row in /status. Set CLAUDE_CODE_AUTO_MODE_SERVER=0 to opt out of the server-side classifier. Update Gemini CLI to 0.61.0, and review your sandbox settings if it is not enabled. Summary Claude Opus 5.5 (claude-opus-5-5) is Claude Code's default Opus model, listed at $4/MTok input and $20/MTok output 2.1.280 makes symlinked writes judged by where they land, and fixes two auto mode retry loops "attribution": false in 2.1.281 is convenient, but older CLI versions skip the whole settings file that holds it Codex 0.157.0 adds GPT-6 Sol and Luna and turns on fullscreen transcripts by default Gemini CLI 0.61.0 is prompt-injection and sandbox hardening, with no new features