← 返回
精选科技模型发布/更新#LLM#AI Agent#Anthropic#安全#What#Added

v2.1.316

Claude Code:GitHub Releases(RSS)T1.5
09:014 天前GitHub Release阅读原文 ↗导出 MD ↓

评分

70

精选理由

来源可信且信息密度较高,适合进入精选池。

AI 点评

「来自 Claude Code:GitHub Releases(RSS) 的这条信息指向模型能力或模型生态变化,适合用于判断近期 AI 圈的真实动向。」

速览

What's changed Added CLAUDE_CODE_DISABLE_ULTRATHINK, which turns the ultrathink keyword off Added pooled group budgets to the Claude apps gateway: one monthly budget that a group's members share, on top of each member's…

  • What's changed Added CLAUDE_CODE_DISABLE_ULTRATHINK, which turns the ultrathink keyword off Added pooled group budgets to the Claude apps gateway: one monthly budget that a group's members share, on top of each member's…
  • After removal, fast mode requests on Opus 4.8 return an error and do not fall back to standard Opus 4.8.
  • Migrate to Opus 5.5 to keep the speedup Fixed shell commands being left running on macOS and Linux when Claude Code exits by itself, for example when credentials are refused or a cloud session is archived Fixed bash ali…

正文

以下为原文,中文全文翻译生成中

What's changed

  • Added CLAUDE_CODE_DISABLE_ULTRATHINK, which turns the ultrathink keyword off
  • Added pooled group budgets to the Claude apps gateway: one monthly budget that a group's members share, on top of each member's own limit, set with pooled_budget on the spend limits admin API
  • Added CLAUDE_CODE_DISABLE_LOOPED_CROSS_CERT_DROP to keep system cross-certificates that certify each other in a loop, for servers whose private CA needs one
  • Added CLAUDE_CODE_SUBAGENT_EFFORT to set the effort level subagents run at when nothing else names one, instead of inheriting the session's
  • Added managed.policy_mode: layered to the Claude apps gateway: every policy that matches a developer applies, each merged over the ones listed below it, not only the first
  • Added guardrail.input_tagging on Claude apps gateway Bedrock upstreams, which wraps prompts in the input tags that an Amazon Bedrock guardrail's prompt attacks filter needs; all_messages_and_system also tags the system prompt
  • Deprecated fast mode for Opus 4.8 as of October 30, 2026; it will be removed on November 30, 2026. After removal, fast mode requests on Opus 4.8 return an error and do not fall back to standard Opus 4.8. Migrate to Opus 5.5 to keep the speedup
  • Fixed shell commands being left running on macOS and Linux when Claude Code exits by itself, for example when credentials are refused or a cloud session is archived
  • Fixed bash aliases with line breaks in their value (except in Git Bash), and functions whose names start with a dash, being garbled or lost in the Bash tool's shell
  • Fixed bash scripts started from the Bash tool stopping after their first command when SHELLOPTS is exported
  • Fixed PowerShell hooks not blocking on exit code 2 when the command begins with a type cast, param, using or an attribute
  • Fixed some network requests hanging when a compressed response was cut off or corrupted
  • Fixed sonnet, opus and haiku on Vertex AI and Amazon Bedrock showing an error mid-session instead of switching to an older model when the newest is not yet available to the account or region
  • Fixed the Claude apps gateway not checking every model a request names against availableModels, deniedModels and models; a request that names a blocked model now gets a 400 naming the field
  • Fixed claude auth login refusing to run where managed settings require a Claude apps gateway; in a terminal it now shows the gateway sign-in
  • Fixed --restricted sessions uploading files from outside the working directories when a Read allow rule covered them
  • Fixed prompts containing invisible characters needing a second Enter to send; plain prompts now send cleaned on the first Enter, and commands still wait for review
  • Fixed sub-agents that inherited a parent sub-agent's model moving to the main model when a background job or message woke them
  • Fixed a managed autoMode.allow list without "$defaults" getting the built-in exceptions back from a host app's --settings, a first rule added in /permissions, or /auto-mode-setup
  • Fixed cloud sessions and cloud scheduled tasks not being marked as failed when auto mode stopped a turn because the server returned no safety verdict
  • Fixed claude -p text output sometimes not saying why auto mode stopped a turn when a background agent was still running
  • Fixed CLAUDE_CODE_AUTO_MODE_SERVER=1 being ignored in sessions started by a host application (such as Claude Desktop) behind a proxy or LLM gateway with experimental betas turned off
  • Fixed plugin MCP headersHelper scripts receiving a corrupted CLAUDE_PLUGIN_ROOT and server URL when a credential-named environment variable was set to a switch value such as 1
  • Fixed the SDK register_repo_root control request not loading the CLAUDE.md of a directory outside the working directory (with CLAUDE_CODE_ADDITIONAL_DIRECTORIES_CLAUDE_MD=1)
  • Fixed the Artifact tool showing ? for artifact database access rules at the self level, and not quoting those rules back when it refuses a republish that would drop them
  • Fixed /model <id> not warning when the model isn't recognized by this version and auto-compact will assume a 200K context window
  • Fixed the "hold space to speak" hint using up its three showings in one session, and showing an empty key when push-to-talk is unbound
  • Fixed Grep and Glob never returning when the search process is killed at the timeout but does not exit, seen on Windows
  • Fixed a rare case where resuming a conversation did not restore its plan, when /clear or /resume had been run while part of that conversation was still being saved
  • Fixed first-run setup dropping words from the security notes and the terminal-setup explanation in narrow terminals
  • Fixed the Enterprise survey card taking a typed 1 or 0 after a short pause; a typed answer now needs Enter
  • Self-hosted runner: Fixed a message sent while Claude was working being answered after a Stop, once the session's process had been replaced
  • Self-hosted runner: Fixed an MCP server's on/off switch sometimes being lost when a session moved to a new machine
  • Windows: Fixed the Bash tool losing your aliases and functions in Git Bash when the rc file sets errexit and pipefail and only winpty aliases, or none, are defined
  • Improved /code-review started from the Code tab in Claude Desktop or on the web: cloud and claude remote-control sessions now show the findings card, not a plain-text list
  • Improved /diff on large repositories: each refresh usually runs one fewer full-repository git command, and refreshes are spaced out when git is slow
  • Improved the messages shown when /ultrareview, a cloud session or a cloud agent finds no GitHub remote: they now say to create a repository at https://github.com/new, and /ultrareview no longer suggests a gh command that gh rejects
  • Improved SSL certificate errors to name the likely causes, among them a server that leaves out an intermediate certificate, and the fix for each
  • Changed auto mode so that an autoMode.allow list without "$defaults" no longer removes the built-in rule on scheduled-task prompts
  • Changed server-managed settings for Team and Enterprise logins: settings fetched from Anthropic's API now apply without an approval prompt, and a stored copy that doesn't verify is refused
  • [VSCode] Added sign-in to a Claude apps gateway from the chat panel, with no terminal needed
  • [VSCode] Changed ultrathink to run its turn at xhigh effort and show in a rainbow in the chat box, as in the terminal
  • [VSCode] Fixed the missing Effort slider, Ultracode switch and model name when the model setting has the [1m] tag, such as opus[1m]
  • [VSCode] Fixed saved /btw side-question history being deleted too early while a settings file has errors
  • [VSCode] Fixed a conversation coming back in Auto mode after a reload or restart when its plan was approved with "Yes, and manually approve edits"
  • [Cloud sessions] Added a form card to cloud sessions, so you can fill in or decline a form an MCP server asks for instead of the request being canceled without you seeing it
  • [Cloud sessions] Added a "Catching up…" line that shows while a cloud or Remote Control session you switch back to is still bringing its messages up to date
  • [Cloud sessions] Changed network secrets so Enterprise members can add and delete them on their own cloud environments, instead of only viewing them
  • [Cloud sessions] Improved the Changes pane for cloud sessions: the note about what the list leaves out is now a badge beside the branch name, not a row above the diff
  • [Cloud sessions] Fixed file search in a cloud session's Files pane showing folders as ../.. paths after Claude changed directory; results now show their path from the project's root
  • [Claude Tag] Fixed Claude refusing to edit or append to a canvas it made in a Slack channel, from that channel, once the canvas was also shared to another channel or a DM
  • [Claude Tag] Fixed untagged replies going unanswered in a Slack Connect thread when both organizations have installed Claude and Slack delivered the reply to the other organization's app
  • [Claude Tag] Fixed a cloud scheduled task's live status checklist in Slack staying half-done; a run can now edit the top-level messages it posted during that run
  • [Claude Tag] Changed a Slack channel's Respond automatically switch to a "How Claude participates" choice on the channel's Configure page and in Claude Tag admin settings
  • [Code Review] Added a findings-by-severity card to the Code Review analytics page, showing the period's total findings, a daily chart stacked by severity, and each severity's count

相关条目

91

五维评分

信息增量9 ×0.3

详细揭示了两个具体设置如何协同作用实现三倍性能提升,信息增量高。

内容质量8 ×0.2

来自OpenAI官方,数据明确,结构清晰,但缺少对比实验细节。

实用价值9 ×0.2

为AI工程师提供了可直接操作的性能调优方案,实用性强。

时效性10 ×0.15

消息刚发布,具有高度时效性。

信源可信度10 ×0.15

信源为OpenAI官方新闻,权威可靠。

基础分 91 × T1 信源权重 ×1 = 91

精选线 60已入选精选

23:00OpenAI:News(RSS)T1RSS

精选启用两项设置使ARC-AGI-3基准测试分数提升三倍

这项发现太关键了!它表明模型的性能不仅取决于训练,还取决于推理时的配置。对于AI工程师和研究者来说,这是一个必须注意的优化方向。特别是那些在资源受限设备上部署模型的团队,可能会因此找到低成本提升能力的方法。

科技模型#ARC-AGI-3#GPT-5.6#OpenAI
91

五维评分

信息增量9 ×0.3

首次公开GPT-5.6具体效率优化细节,此前仅有传闻。

内容质量8 ×0.2

内容简洁,关键指标明确,但缺乏详细技术文档支撑。

实用价值9 ×0.2

直接指导企业如何降低AI部署成本,具有商业参考价值。

时效性10 ×0.15

刚发布的官方新闻,时效性极强。

信源可信度10 ×0.15

来源为OpenAI官方博客,信源权威。

基础分 91 × T1 信源权重 ×1 = 91

精选线 60已入选精选

08:00OpenAI:News(RSS)T1RSS

精选GPT-5.6:前沿智能与前沿效率的融合

这条消息值得关注,因为它标志着大模型竞争从单纯拼参数转向拼效率。对于AI开发者和企业决策者来说,GPT-5.6意味着可以用更少预算获得更高性能,尤其适合那些被推理成本挡在门外的长上下文或Agent场景。OpenAI这一步,可能重新定义‘前沿’的标准。

AI模型#Agent工作流#GPT-5.6#OpenAI
91

五维评分

信息增量8 ×0.3

首次提出针对CPU优化的1M上下文模型,填补了边缘设备长文本推理的空白,信息增量显著。

内容质量9 ×0.2

技术细节清晰,提供了基准测试对比和架构解释,图文并茂,质量上乘。

实用价值10 ×0.2

直接降低长上下文应用的硬件成本,开源权重可立即部署,实用价值极高。

时效性10 ×0.15

当日发布的博客,且长上下文推理是当前AI热点,时效性满分。

信源可信度9 ×0.15

Hugging Face官方博客,作者为Liquid AI,属于第一方技术发布,可信度高。

基础分 90.5 × T1 信源权重 ×1 = 91

精选线 60已入选精选

23:01Hugging Face:Blog(RSS)T1RSS

精选LFM2.5-Encoder:面向CPU快速长上下文推理的新模型

这篇博文让我眼前一亮,因为它直接解决了大模型在CPU上跑长上下文的痛点。对于没有高端GPU的团队或个人开发者,LFM2.5-Encoder系列提供了可立即使用的开源方案。尤其推荐给做RAG、文档搜索或代码分析的朋友——现在用普通服务器就能处理百万级token的文档了。

AI模型#CPU推理优化#LFM2.5-Encoder#Liquid AI
91

五维评分

信息增量9 ×0.3

首次公开自动化红队系统的具体机制与实验发现,信息增量高。

内容质量8 ×0.2

来自官方,结构清晰,但未提供完整技术细节(如具体模型版本),扣1分。

实用价值9 ×0.2

对AI安全工程师、模型开发者和企业部署者直接指导作用,实用性极强。

时效性10 ×0.15

刚刚发布,是当前AI安全领域最前沿进展之一。

信源可信度10 ×0.15

信源为OpenAI官方新闻,权威且一手。

基础分 91 × T1 信源权重 ×1 = 91

精选线 60已入选精选

18:00OpenAI:News(RSS)T1RSS

精选GPT-Red:OpenAI通过自我对弈实现AI红队自动化,提升模型鲁棒性

GPT-Red 是红队测试从人工到自动化的里程碑,特别适合安全工程师和对齐研究者关注。它直接降低了测试成本,但也引发了对“自动攻击”本身的伦理讨论——值得所有关注AI治理的人深入思考。

AI模型#AI安全对齐#GPT-Red#OpenAI
91

五维评分

信息增量9 ×0.3

这是首个针对AI模型生物安全的漏洞赏金计划,为行业提供了全新的安全评估视角。

内容质量9 ×0.2

官方来源,内容清晰,涵盖范围、奖励、准则等关键细节。

实用价值8 ×0.2

对AI安全研究员和生物安全专家有直接参与价值,但普通用户关联度较低。

时效性10 ×0.15

刚刚发布,完全符合新闻时效要求。

信源可信度10 ×0.15

OpenAI官方RSS和博客,信源权威可靠。

基础分 91 × T1 信源权重 ×1 = 91

精选线 60已入选精选

18:00OpenAI:News(RSS)T1RSS

精选OpenAI推出GPT-5.5生物漏洞赏金计划

这个计划值得关注,因为它首次将AI漏洞赏金制度拓展到生物安全领域,不仅为安全社区提供了新的研究方向,也为AI风险治理提供了实际演练场。对模型安全研究者、生物安全专家以及关心AI双重用途的从业者尤其重要。

AI模型#Bio Bug Bounty#GPT-5.5#OpenAI
91

五维评分

信息增量8 ×0.3

公布了vLLM新后端的具体架构设计与实测数据,对了解推理优化前沿有较高信息增量。

内容质量9 ×0.2

技术细节清晰,有定量性能对比,来源为官方博客,内容质量高。

实用价值10 ×0.2

直接降低LLM部署门槛与成本,对AI工程化落地具有极高实用价值。

时效性10 ×0.15

发布即时,属于最新技术动态,时效性满分。

信源可信度9 ×0.15

来源为Hugging Face官方博客,vLLM为开源权威项目,可信度高。

基础分 90.5 × T1 信源权重 ×1 = 91

精选线 60已入选精选

08:00Hugging Face:Blog(RSS)T1RSS

精选vLLM推出原生速度Transformers建模后端,大幅提升推理性能

这项更新对于AI工程团队非常关键:无需格式转换意味着能更快地迭代模型,而2-5倍的速度提升直接降低GPU成本。如果你正在做LLM在线服务或者批量推理,这几乎是必读的技术更新。

AI模型#Hugging Face#PagedAttention#Transformers
90

五维评分

信息增量9 ×0.3

手语转文本是AI无障碍领域的新突破,提供了此前罕见的具体模型信息。

内容质量8 ×0.2

来源权威,内容清晰描述了模型功能、应用场景和开发理念。

实用价值9 ×0.2

对聋人群体、开发者和无障碍研究者具有直接参考与应用价值。

时效性9 ×0.15

刚刚发布的官方公告,时效性非常高。

信源可信度10 ×0.15

Google DeepMind官方博客,信源可信度极高。

基础分 89.5 × T1 信源权重 ×1 = 90

精选线 60已入选精选

22:01Google DeepMind:Blog(RSS)T1RSS

精选Google DeepMind 发布手语转文本模型 SL2T,助力聋人用户

这则消息的意义不仅在于技术突破,更在于AI开始真正关注边缘群体的需求。对开发者而言,SL2T的架构思路和技术路径值得深入研究;对无障碍倡导者来说,这是推动行业标准的重要案例。

AI模型#Google DeepMind#SL2T#手语识别

引用前请核对原文