Claude Code 운영 고도화, WAN 3.0 영상 생성, AgentX 추론 최적화와 로컬 모델 경쟁

Claude Code CLI 2.1.243 changelog: New features: • Added a Loops breakdown to /usage: per-loop run count, total tokens, tokens per run, and last run, so runaway or chatty /loop tasks are easy to spot • Added modelPicker setting: curate the /model picker with an ordered, labeled list of models (any id spelling, including Vertex/Bedrock ids), appended to or replacing the built-in lineup • Added promptCacheTtl and subagentPromptCacheTtl settings so API-key and cloud-provider users can keep a 1-hour prompt cache on the main conversation while subagents stay at 5 minutes • Added modelPricing managed setting so an organization's contracted per-model rates and discount multiplier are used for /cost, the status line, and telemetry cost figures instead of list price • Added a keyless sign-in under /login → Anthropic Console: "Sign in with your Console account" (recommended) alongside creating an API key, so organizations that don't allow API keys can sign in • Added a Skipped sources line to /status that lists managed settings sources (for example managed-settings.json) present but not applied because a higher-precedence managed source is active • Added a managed marker in /mcp and /plugins on http:// claude.ai connectors whose authentication is managed by your organization • Added a tip pointing http:// claude.ai users who haven't connected GitHub for Claude Code on the web to /web-setup • Added a /status line showing whether GitHub is connected for Claude Code on the web (Pro/Max), pointing to /web-setup when it isn't • Added the model (and effort level) each subagent ran on to /tasks and the agent detail dialogs Fixes: • Fixed remote MCP servers in non-interactive (-p) and SDK sessions never recovering after a dropped connection; they now reconnect automatically or report as failed • Fixed MCP server sign-in started from the desktop app failing with "Invalid redirect URI" on servers that support client ID metadata documents (for example Linear) • Fixed auto mode staying unavailable at startup when a temporary server-side disable was cached and later flag fetches failed • Fixed auto mode tool calls being denied as "temporarily unavailable" after about a minute of waiting when the API was briefly overloaded and asked the client to retry • Fixed the /model picker silently ignoring an Ultracode selection; picking Ultracode now applies it to the current session • Fixed /resume only listing the 50 most recent sessions; the picker now loads more as you scroll • Fixed cloud sessions resuming after a mid-turn restart with a pending hook or background-task notification re-sent as the prompt instead of the normal continuation message • Fixed cross-session messaging silently turning off inside user namespaces and rootless containers after the 2.1.232 socket-directory hardening • Fixed text that hangs outside its container (for example the sign-in URL in /login) losing its leading columns when another part of the screen repaints • Fixed spellcheck not underlining a misspelled word typed directly after an emoji • Fixed background subagents not waking when their last background Bash task completes • Fixed sessions going silent for 10+ minutes when the Anthropic API never starts a response: the request now times out after ~3 minutes, retries once, then shows API Error: No response from API • Fixed auth, model-availability, and other client-generated error messages rendering like model output instead of as error lines • Fixed workload identity federation in CI: processes in one job share the exchanged token instead of re-exchanging the single-use token; a rejected exchange fails fast with the server's message • Fixed server-managed companyAnnouncements not showing at startup in a session that began with signing in (for example the first launch after /logout) • Fixed hook if conditions like Bash(cat *) firing on unrelated Bash commands when the command contained $() or backtick command substitution followed by more arguments • Fixed plugin dependencies declared with a marketplace field never resolving when both plugins are loaded together via --plugin-dir • Fixed /reload-plugins keeping the LSP tool after the last LSP plugin is disabled; it now also warns before an LSP plugin change that would re-read the conversation • Fixed --agents silently ignoring invalid JSON or invalid agent definitions; it now exits with a clear error, like --mcp-config • Fixed /status showing "Found invalid entries in: ." with no filename when ~/.claude.json has an invalid MCP server entry • Fixed /clear removing the /rename session name from the prompt bar even though the name was kept for the new session • Fixed Ctrl+R history search and up-arrow history breaking when ~/.claude/history.jsonl contains a malformed entry • Fixed Ctrl+[ not leaving vim INSERT mode in terminals that encode modified keys (modifyOtherKeys / kitty protocol) • Fixed the local IDE connection being routed through HTTPS_PROXY (and sometimes failing) when localhost was listed in NO_PROXY but not lowercase no_proxy; both casings are now honored • Fixed sandbox network-violation details being dropped from the Bash tool result when the blocked command still exited 0 (for example curl printing the proxy's 403 page) • Fixed the status line rate_limits fields and /usage still showing a rate-limit window's pre-reset usage percentage after the window reset while the session was idle • Fixed claude --teleport <session> exiting on uncommitted changes instead of offering to stash them and continue, as the session picker already does • Fixed /web-setup repeatedly asking you to log in when an older GitHub CLI (without gh auth token) was already authenticated • Fixed Claude in Chrome losing its connection to Claude Code after an auto-update cleaned up the version it was set up with; the native host now launches via the stable claude launcher • Fixed claude remote-control exiting and stranding attached Remote Control sessions when the server drops its environment mid-session; it now recovers • Fixed Remote Control sessions served by claude remote-control sometimes getting stuck after it was stopped and restarted, for Team and Enterprise members without an admin or owner role Improvements: • Improved startup time: sandbox and MCP bring-up no longer block the first frame, bare launches skip subcommand registration, and workflow discovery, settings, and trust-store work is cheaper • Improved native install and auto-update download size: the binary is now zstd-compressed (about 75 MB instead of 340 MB on Linux x64) • Improved attribution of usage telemetry to your organization for sessions that authenticate with ANTHROPIC_AUTH_TOKEN directly against the Anthropic API, so its data-handling settings apply • Improved native binary size: about 2 MB smaller by storing the bundled skill and prompt text more compactly • Improved memory usage of native builds: code is now loaded on demand instead of keeping the whole bundle resident (roughly 40–70 MB less memory per session) • Improved peak memory usage in long-running sessions (the runtime now garbage-collects sooner as the heap grows) • Improved /login over SSH: the sign-in URL appears immediately, pressing c reports how the URL was copied instead of always claiming success, and a hint explains how to select text in fullscreen • Improved the error when effort xhigh/max is used with thinking turned off: it now names the level, the setting that disabled thinking, and /effort high as the fix • Improved /loop: consecutive wake-ups where Claude has nothing to do now fold into a single line in the terminal instead of printing each one • Improved the notice when resuming a conversation whose Remote Control is held by another terminal: it now says sessions on other machines can't be seen from, or reach, this one Security/safety changes: • [VSCode] Fixed sessions started before feature flags were first fetched (for example right after install) opening in the default permission mode instead of auto mode or your configured default mode • Changed the sandboxed Bash tool prompt to no longer list allowed network hosts, so Claude attempts requests (and you can approve new hosts) instead of assuming unlisted hosts are blocked Other changes: • [VSCode] Fixed Focus view sections you expanded collapsing on their own during subagent tool activity • Updated the /model picker and the bundled claude-api skill to show Sonnet 5's $2/$10 per Mtok pricing as its standard list price rather than a limited-time promo • Changed computer use on macOS so clicking the desktop, Dock, or a Finder window requires granting Finder via the access dialog, like any other app • Changed /model, /fast, and /effort to also run immediately instead of queueing until the turn ends on Bedrock, Vertex, and Foundry and when telemetry is disabled • Changed the cross-session messaging inbox socket to close connections that send no complete line within 30 seconds; scripts posting to it should connect once their data is ready • [VSCode] Improved history trimming in long sessions: older tool-activity rows are dropped first so your messages and Claude's replies stay visible • [VSCode] Improved attribution of the extension's own usage telemetry to your organization when you are signed in with a Claude account, so its data-handling settings apply Source: https:// github.com/anthropics/cla ude-code/blob/main/CHANGELOG.md#21243 …

Claude Code 2.1.243 has been released. 60 CLI changes Highlights: • Loops breakdown in /usage: per-loop runs, total tokens, tokens/run, last run — reveals chatty or runaway loops • modelPicker curates /model into an ordered, labeled org-approved list, making approved models easiest to pick • Added modelPricing setting so /cost, status, and telemetry use org contracted per-model rates and discounts Full details are in thread ↓

WAN 3.0 is really cool! You can now: • Generate native 30-second videos • Use text, images, or Omni Video as input • Give it up to 20 reference assets, including docs and webpages • More consistent outputs across your entire workflow • Make more precise edits (don't need to regenerate as much) I remember when you could only get 5-second videos out of these models. We are now way past that: you can now control a complete production workflow. You can use WAN 3.0 on the Pika API Club. It's one of the least expensive platforms hosting it. https:// dev.pika.art/?utm_source=x& utm_medium=caption&utm_campaign=260823-wan_3_0_launch&utm_content=svpino&utm_term=post&utm_id=7397254a-a663-4c95-bd7e-c7b50a8c39f0 …

WAN 3.0 is now on Runway. Generate state-of-the-art video and audio with multiple image, video and audio reference inputs. Try it now at the link below.

Congratulations to @SemiAnalysis_ on the release of AgentX 1.0 , an open-source multi-turn agentic coding benchmark collected from ~$3M of real traces, running on 1000+ chips and ~2MW of continuously operated compute. We are excited to see @vllm_project ’s competitive performance on frontier open models: 130,093 tok/s/chip for DeepSeek V4 Pro 77,079 tok/s/chip for Minimax M3 12,479 tok/s/chip for Kimi K3. The following thread covers an overview of the work from @vllm_project and @inferact : what we tested, found, and shipped upstream. This work highlights vLLM’s performance on real-world workloads and our committed focus to making vLLM an agentic-first engine. Optimizing AgentX performance meant tackling three major challenges: prefix reuse, efficient long context parallelism, and scaling performance with PD disaggregation. First, long agentic sessions stress prefix caching and KV cache offloading. Modern hybrid models have greatly reduced the required KV cache sizes, and caching every block boundary still saturates the KV cache pool which causes prefix cache thrashing across sessions. The fix was sparse retention: one state per interval-sized segment plus the latest replay boundary (vllm-project/vllm #43447, #45845), then preserving shared-prefix boundaries so the interval can go to 0 for agent sessions (#47782). That gets us >95% hit rate at 14 concurrent requests with contexts to 1M. The bigger structural change in KV cache offloading was making the shared KV pool distributed. With Mooncake Store as a first-class connector, prefill ranks can hit the prefix cache both within a worker and across workers, so we no longer have to trade cache locality against load balance to keep a cluster busy. Session-aware routing (48048) is what lets the router act on it. For a single node deployment, SimpleCPUOffloadConnector has been greatly improved to support all hybrid model architectures and across both CUDA and ROCm platforms. For DeepSeek V4 Pro on ROCm, this implementation gave +81.7% output throughput and 46.6% lower mean e2e latency versus recomputing the prefix. Kimi K3 at 2.8T barely fits on a single node, and squeezing it in leaves almost no headroom for KV cache, which is exactly what a long multi-turn session needs most. Parallelism strategy matters more here than on any other model we tested. TP8/DCP8 tops the K3 agentic frontier across configs, with B300 vLLM peaking at ~12.5k tok/s/chip at ~8 tok/s/user and GB300 NVL72 on Dynamo + vLLM holding the curve out past 200 tok/s/user. K3 also surfaced a routing bug worth pulling: vllm-project/router#194 fixes the router dropping reasoning_content, which hits any reasoning model served behind it. On MiniMax M3, B200 vLLM reaches ~44k tok/s/chip and vLLM leads TRT-LLM on throughput vs p90 TTFT. M3 and Qwen3.5 both shipped day-0. Finally, fully optimized performance requires scaling the deployment with prefill-decode disaggregation and distributed KV cache offloading. Thanks to vLLM’s MultiConnector, this is natively supported with NIXL PD connector + MooncakeStoreConnector. By rate matching to find the ideal prefill-decode ratio, we achieved 4.45x higher throughput at a 60 tok/s interactivity for DeepSeek V4 Pro on GB300 Dynamo compared to B300. Shoutout to @NVIDIAAI , who co-tuned most of these configs with us. Dynamo's router optimizations took AgentX replay time down 23.7% on the vLLM backend, the AIPerf replay harness is what made the traces runnable at all, and NIXL and their kernels sit under each respective performance point. Shoutout to @AIatAMD as well: the team’s AITER sparse-MLA decode selection led to +5.22% AgentX output throughput; the hybrid AITER/native CSA selector led to 1.21–1.76x e2e performance boost, and the team made prebuilt lmcache gfx942/gfx950 wheels and a published Mooncake ROCm wheel. Thank you to @SemiAnalysis_ for building AgentX and for the collaboration throughout. Agentic workloads are what users serve in real production, and benchmarking on real workloads is what improves vLLM and open source inference. Next up: more upstream work and a blog with a full technical deep dive later this week. Stay tuned.

Grok @bot is likely the second (first would be FSD) AI product that is solving daily errands. Something is broken? I took a picture and it found the vendor and warranty, scheduled the service, and checked my calendar. Neat.

Elon Musk

@elonmusk

2일 전

Grok @Bot can do a lot!

Qwen3.8-27B (dense) ranks #9 on Code Arena WebDev only 4 points behind GLM 5.3 Max. - Beats Claude Opus 4.8, DeepSeek V4 Pro, Gemini 3.7 Flash And it runs on a single RTX 3090/4090.

Stop wasting GPU cycles on overthinking. Qwen 3.8 27B was burning 3 minutes deciding a simple tool call. One setting change (reasoning budget = 8k) cut median time in half on real coding tasks. Set reasoning budget to 8k tokens to median time cut in half (190s to 94.7s) on coding tasks.

“Autonomous de novo protein binder design with Claude” Protein binder design is getting powerful, but running a full campaign still takes a lot of heuristics. This paper from Claude Science gives Claude a single general protocol and lets it autonomously research targets, choose epitopes, run open-source design models, optimize candidates, and rank proteins for synthesis. Across 15 protein targets, about 1 in 4 of Claude’s designs actually bound, and its #1 pick worked nearly half the time. On RBX1, Claude’s best design bound about 10x more tightly than the winner of an open protein-design competition. https:// alphaxiv.org/pdf/2608.claud e-de-novov1 …

“ABRA: Scaling Diffusion Image Training” Diffusion image models scale predictably like LLMs, but their compute-optimal recipe is very different. This paper finds they need roughly 200 image tokens per parameter, around 10x Chinchilla, and are far more tolerant to overtraining than undertraining. So for fixed compute, smaller diffusion models trained on more data are the safer bet. https:// alphaxiv.org/pdf/2608.17286 v1 …

Thomson Reuters launched Thomson, its first proprietary large language model, built for about $40 million by adapting Alibaba's open-source Qwen and training it on Westlaw, Practical Law, tax, and news content. The model first powers Tabular Analysis features in CoCounsel Legal, part of a push to cut the company's reliance on third-party AI vendors.