GPT-6 Astra 확산, Ling 의료·비전 모델, 코딩 에이전트 검증 강화

GPT-6 Astra is now available to all Pro, Enterprise, and Business Premium users in Work/Codex, and is available in the API. We will start rollout to Plus and Business users next. Thank you for the patience.

🌌 GPT-6 Astra from @OpenAIDevs is now generally available in GitHub Copilot. This new model is designed for long-horizon, autonomous coding and agentic tasks. Our internal testing shows it plans and validates as it goes, batches diagnosis with verification, and independently confirms its results before declaring a task done. That translated into stronger performance on extensive coding jobs with fewer steps than prior OpenAI models. Try it out in the GitHub Copilot app, CLI, or @code . 🚀 https:// github.blog/changelog/2026 -09-04-gpt-6-astra-is-generally-available-in-github-copilot/ …

Today, we are releasing Ling-3.0-flash-VL, built on Ling-3.0-flash with visual understanding and visual agent capabilities. It performs well across visual perception, STEM reasoning, document intelligence, multimodal agent tasks, frontend coding, and medical report interpretation.

Today, we’re introducing Ling-3.0-flash-Sante — an MoE model enhanced for health and medicine, built on Ling-3.0-flash. Inspired by the French word “santé,” meaning “health,” Sante is built for real-world healthcare tasks spanning medical reasoning, professional healthcare tasks, deep research, and evidence-based retrieval. Sante shows competitive results across MedXpertQA-Text, DiagnosisArena-MCQ, AFUMED-Drug, HealthBench Professional, and BrowseComp, with leading performance among open-source models and results competitive with flagship models. Here’s to better health — Santé! More below ↓

Claude Code CLI 2.1.261 changelog: New features: • Added an "Organization policy" line to /status and claude doctor that says why your organization's policy could not be loaded, such as a proxy not passing the endpoint through • Added bashOutputMaxChars and taskOutputMaxChars settings to raise how much command and background-task output Claude receives inline before it is saved to a file, up to 128K characters • Added --append-subagent-system-prompt-file to read the subagent system prompt from a file, for prompts too large to pass on the command line • Added /skill-doctor to show which loaded skills go unused and what they cost in context, so you can prune them Fixes: • Fixed typed or pasted characters occasionally landing out of order or being dropped during fast input or key repeat • Fixed /add-dir <subdirectory> printing a false "couldn't be resolved" error when the working directory is on a /net automount • Fixed the Bedrock setup wizard hanging when AWS or an AWS credential helper never responds (it now times out with a clear error), and its model checks failing behind a TLS-inspecting proxy • Fixed cloud sessions discarding a plugin synced from http:// claude.ai when managed settings force-enable it in enabledPlugins, then falling back to a marketplace clone that could fail • Fixed being unable to delete the character immediately before an inline [Image #N] chip in the prompt input • Fixed resuming a session losing hook output and other context around parallel tool calls, which changed the resumed request • Fixed Remote Control showing a stale permission mode when a phone, browser, or http:// claude.ai app attaches to a terminal session or after the mode changes in the terminal • Fixed Remote Control sessions showing as still working (stuck spinner and Stop button) after stopping a turn from a connected phone or browser, or after a local slash command like /clear • Fixed SDK and cloud sessions ignoring a Stop or interrupt sent just after the first prompt, before the turn had started; the turn now stops instead of running to completion • Fixed Remote Control uploading a session pulled with /teleport into the connected session, which appeared appended to the original on phone and web • Fixed Remote Control's inbound event stream failing behind TLS-inspecting corporate proxies on native Windows • Fixed Remote Control sessions showing the default effort level on http:// claude.ai when the effort comes from settings • Fixed gcpAuthRefresh opening a browser at startup when the Google credential check was slow, even though the credential was still valid • Fixed http:// claude.ai connectors staying absent for the whole session when the startup connector fetch timed out — the CLI now retries in the background • Fixed sustained high CPU usage when a background agent could not be resumed and its wake-up was retried in a tight loop • Fixed feature flags gated to a newer version occasionally applying to an older Claude Code version running on the same machine • Fixed /usage and the VS Code usage panel dropping a model-specific weekly limit row when the usage endpoint is rate limited or when opened right after startup • Fixed claude -p --resume <file> adopting a malformed session ID recorded in the transcript; it now resumes under a fresh session ID instead • Fixed the terminal progress indicator (iTerm2, Ghostty, ConEmu) showing the session as finished while a background workflow or agent was still running • Fixed a rare layout glitch where a box could render with the wrong height after its container switched between row and column direction • Fixed Claude apps gateway client IP when a trusted proxy appends a port to X-Forwarded-For; with an access list set, an unreadable entry now gets 403 • Fixed Claude apps gateway telling Claude Desktop to export OpenTelemetry as JSON even when the terminal CLI uses protobuf, so protobuf-only collectors rejected Desktop's data • Fixed Desktop and web showing a session as busy while it only watches an artifact for updates • Fixed Claude in Chrome file_upload failing with "paths: expected array, received undefined" in local Cowork sessions run from the Claude Desktop app • Fixed SendMessage to an offline Remote Control session on another machine reading as delivered; the result now says delivery is queued until that machine reconnects • Fixed plugin install hints from CLIs run in background Bash commands: they are now detected, and the raw <claude-code-hint> tag no longer leaks into the conversation • Fixed in-process agent-team teammates re-sending their first-turn tool and skill announcements on the second turn, which changed the request prefix and missed the prompt cache Improvements: • Improved the /model picker and the VS Code model pill to show a model's name instead of its raw Bedrock, Vertex AI, or LLM gateway ID when Claude Code recognizes it • Improved startup on Google Vertex AI when GOOGLE_APPLICATION_CREDENTIALS is set: API client creation no longer re-runs Google Cloud project discovery or spawns extra gcloud processes • Improved streaming performance: already-rendered blocks are no longer re-checked by layout on each update • Improved the dangerous-rm safety prompt to also catch rm -rf on positional parameters and inside double-quoted sh -c scripts • Improved handling when the API sends no response headers: the retry now waits up to API_TIMEOUT_MS (10 minutes by default) instead of another 3 minutes, and the messages say what to change Security/safety changes: • [VSCode] Added a fold button to permission and question prompts so the conversation behind them can be read without dismissing them; the space beside the prompt now scrolls the conversation • [VSCode] Fixed the next queued permission prompt keeping text typed on the previous prompt and accepting an immediate second click Other changes: • Changed a Claude apps gateway 403 on the managed settings load (at startup or after /login) to say Claude Code may not be enabled for the organization, instead of advising a new sign-in • Changed machines whose managed settings pin forceLoginMethod: "gateway" to ignore a leftover API key or http:// claude.ai login and ask for /login; Bedrock, Vertex AI, and Foundry sessions are unaffected • Changed auto mode to treat a link that packs content into a public diagram renderer's URL as an upload to that site: no longer auto-approved unless you asked for it • Changed the prompt's word-editing keys to match Bash: Ctrl+W deletes back to whitespace, Alt+F and Alt+D stop at word end, punctuation separates words; keybindingFlavor no longer has any effect • Changed /context token counting to use a local estimate when the token-counting API is unavailable, instead of extra small-model requests • [VSCode] Added a "Build a custom style" walkthrough to the Output styles menu that writes a custom output style file and lists it right away • [VSCode] Added an Add server form and a Remove action to the MCP servers dialog, so MCP servers can be added and removed without leaving the IDE • [VSCode] Added a hollow ring in the session list for sessions open in a terminal, another VS Code window, or Claude Desktop, so they no longer look closed • [VSCode] Added "Archive session" to the session list's right-click menu and gave Unarchive its own icon • [VSCode] Fixed a session teleported from Claude Code on the web treating a question that was cut off when the cloud session shut down as declined • [VSCode] Fixed the session tab's Rename box opening empty for a tab restored with the window; it now starts with the current name • [VSCode] Fixed collapsed sections in the session list panel briefly showing expanded each time the panel loaded • [VSCode] Fixed Focus view showing a tool call as still running after Claude had moved on, such as while a question waited for your answer • [VSCode] Fixed the session list's active-row highlight going stale when an unfocused Claude tab's session ID is corrected • [VSCode] Fixed Cmd/Ctrl+Shift+T reopen and deep-link opens placing the Claude tab outside the Claude editor group when a Claude tab has focus • [VSCode] Fixed the session tab's "Add to group" putting a session opened from Claude Code on the Web in two groups; it now moves the entry the session list shows • [VSCode] Fixed the model picker showing models an organization has since disabled until the window was reloaded twice • [VSCode] Fixed a tab opened from the session list jumping back to that session, and a tab opened from a Web session restarting its teleport or staying empty, after VS Code reloads the tab's view • [VSCode] Fixed /btw side-question history from earlier sessions being overwritten when a question is asked right after a window reload or while a settings file has errors • [VSCode] Fixed the pending question card not reappearing after the Claude panel reloads when signed in with a http:// Claude.ai or Console account • [VSCode] Fixed claude.ai-only features staying visible in a window's other Claude panels after one panel picked up a third-party provider from a settings file • [VSCode] Fixed the sign-in screen appearing despite the Disable Login Prompt setting when Claude Code reports no login or a request fails for lack of one • [VSCode] Fixed install-plugin links opening the Claude sidebar without the install dialog in a window where only the session list had been shown • [VSCode] Fixed the sidebar usage meter staying empty on a new window until the Account & usage dialog was opened, and a 0% usage limit being left out of the meter • [VSCode] Fixed "Start new session in this group" losing the group after New conversation, and a missing unread dot for a session that finished before the sidebar's unread list loaded • [VSCode] Fixed the editor tab badge showing unread during a running turn or missing on a tab opened from the session list, and "Add Session Tab to Group" doing nothing for an archived session • [VSCode] Fixed "Enable Remote Control for all sessions" so flipping it also applies right away to sessions open in other VS Code windows • [VSCode] Fixed the session list's Open filter for sessions continued from http:// claude.ai whose tab was still recorded under the web session, and labeled the filter menu's sections for screen readers • [VSCode] Changed the model picker to one flat list of every model, with rows kept for older model spellings listed last Source: https:// github.com/anthropics/cla ude-code/blob/main/CHANGELOG.md#21261 …

LangChain OSS

@LangChain_OSS

4시간 전

MCP usage is growing... and growing fast! check out our latest blog on our revamped support for MCP and the new, stateless spec!

Checking that a major mathematical proof is correct can take years. Formalization—converting the mathematical reasoning into a form computer proof assistants like Lean can verify—can help. Last month, Claude completed the first formalized proof of Fermat’s Last Theorem, one of

You can Run locally Gemma 4 12B Coder distilled from execution-verified Python CoT (Composer 2.5 + Fable 5). (Q2_K 4.83 GB) - thinks first, then writes - Fable 5 re-solved the failures. - Trained only on CoT that produced code that *passed tests* - thinking mode on - & the sweet spot Q4 fits 12GB cards/Apple Silicon - up to 256K context - the model learns “think to code that works,” not “think to plausible code.” This one was trained on reasoning traces where the solution was actually run and only kept if tests passed.

We've shared research on fine-tuning forecasters, now you can try it for yourself: a cookbook recipe for training a model to predict event probabilities given reference information. Start with the provided dataset from @ProphetArena , then try your own. https:// github.com/thinking-machi nes-lab/tinker-cookbook/tree/main/tinker_cookbook/recipes/forecasting …

Introducing ExtractBench on Kaggle Benchmarks with @llama_index . When AI agents rely on schema-guided extraction before human review, one truncated schedule or invented value becomes a wrong payment or decision. ExtractBench evaluates models in workflows based on real-world documents across industries such as supply chain, healthcare, and finance, by measuring whether the system returns: ➣ Missing fields as null instead of inventing a value ➣ Source evidence for each value ➣ Every record of each repeated structure GPT-5.6 Sol currently leads at 91%.

Aravind Srinivas

@AravSrinivas

3시간 전

Perplexity will be at RustConf to present how we built the sandboxes powering all of Perplexity Computer.

LangSmith for Startups: @raspberry__ai ✅ The agentic platform for fashion. ✅ Works alongside design teams on their boards and turns plain-English requests into finished renders, tech packs, and campaign imagery in minutes. ✅ Runs the full lifecycle of its LangGraph agent on LangSmith. Join Raspberry AI to transform the fashion industry: https:// jobs.ashbyhq.com/raspberry?utm_ source=Langsmith …