Muse Spark 1.3의 코딩 성능 급등, Gemini 3.8 Flash의 비용 우위, 실행형 AI 인프라 확장

Md Ismail Šojal

@0x0SojalSec

4시간 전

Wtf Meta’s Muse Spark 1.3 just took first place on DeepSWE v1.1 with 75.4%. That’s ahead of GPT-5.6 Sol (~73%) and Claude Fable 5 (70%). Muse Spark 1.3 just jumped from 55% to 75.4% on DeepSWE. DeepSWE uses 113 original tasks from 91 live open-source repos across 5 languages, with isolated environments and behavior-based checks. Harder to game than the usual SWE benches.

Wtf, Muse Spark 1.3 max (limited preview): 62 behind only Claude Fable 5.1 (66) and Claude Opus 5 (63) The race is no longer OpenAI vs Anthropic. currently Meta enter with big hit. Look at this leaderboard : 1. Claude Fable 5.1 - 66 2. Claude Opus 5 - 63 3. Meta Muse Spark 1.3 (max) - 62

Gemini 3.8 Flash (High) by @GoogleDeepMind has reshaped the Pareto frontier for Agent Arena! This model is priced at $0.75/$3.75 per MToken (input/output). In Agent Arena, Gemini 3.8 Flash (High) has a median cost of $0.22 per task and +5.94% net improvement. This performance is on par with Grok 4.5 by @SpaceXAI and GLM 5.2 (Max) by @Zai_org , but at a 44-50% lower price point. - Gemini 3.8 Flash (High) (+5.94% at $0.22/task) - Grok 4.5 (+6.17% at $0.39/task) - GLM 5.2 (Max) (+6.23% at $0.44/task) The release of Gemini 3.8 Flash (High) has landed Gemini on the Agent Arena Pareto for the first time. Congrats again to the team @GoogleDeepMind !

Gemini 3.8 Flash achieves 73.7% on DeepSWE. It pushes the frontier with an impressive 8.2% increase over Gemini 3.7 Flash, at the same cost but with more steps and output tokens.

Claude Code CLI 2.1.259 changelog: New features: • Added managedMcpServers managed setting: organizations can provide HTTP/SSE MCP servers to every user (same entry shape as .mcp.json); entries that name a command to run are skipped • Added --permission-prompts none for unattended headless hosts: anything that would prompt is denied automatically while the active permission mode (including auto mode) keeps deciding • Added recognition of glab mr create/merge/close/reopen/note/update so GitLab merge requests show as MR !N in the collapsed tool summary and refresh the footer MR badge • Added --json to claude plugin validate for a machine-readable validation report Fixes: • Fixed concurrent sessions silently reverting each other's ~/.claude.json changes — workspace trust no longer resets and MCP/project state is no longer lost when running many sessions at once • Fixed a conversation whose thinking was rejected once being rejected again on every later turn • Fixed Bash Read() deny rules not covering files given as option values (--ignore-revs-file=.env, -f.env, @file ), git diff/git grep file operands, or cd DIR && cat FILE compounds; grep -r/cp -r over a directory holding a denied file now asks • Fixed the prompt cache being invalidated when the OAuth token refreshed in sessions with telemetry disabled • Fixed fullscreen mode showing a blank conversation after a long turn with hundreds of tool calls • Fixed auto mode running a turn on a model it doesn't support when a command or skill's frontmatter model: named one; the turn now keeps the session model • Fixed CLAUDE_CODE_MAX_CONTEXT_TOKENS being ignored for Vertex-style model IDs ( @YYYYMMDD suffix) of model versions Claude Code doesn't recognize • Fixed the live output preview of a running shell command hiding its newest lines when an earlier line wrapped • Fixed a background GitHub connection check that ran on every launch for http:// claude.ai users; the result is now remembered across launches • Fixed --resume failing (and --continue opening an empty conversation) when a saved session contains an attachment entry with no payload • Fixed frontmatter model: on custom commands and skills being ignored in interactive sessions • Fixed Artifact publishing failing once with an "unexpected parameter note" error in conversations continued from an older version • Fixed managed forceRemoteSettingsRefresh being ignored at startup when a policy helper configured by MDM or the managed settings file had already run • Fixed worktree isolation refusing hook-created worktrees on machines where git rev-parse fails with a message other than "not a git repository" • Fixed OpenTelemetry metrics and events from cloud sessions missing the http:// user.email, http:// organization.id, and user.account_uuid attributes • Fixed MCP servers that disconnect while their tools are being listed at startup showing as connected with no tools instead of reporting the error • Fixed the file edit permission dialog sometimes showing a changed line cut short with no indication • Fixed repository detection dropping a known repo identity after a transient git probe failure • Fixed managed settings silently going unenforced when the managed-settings file, a drop-in, the MDM plist, or the HKLM value cannot be parsed: Claude Code now refuses to start and names the source • Fixed Stop not actually stopping background agents and workflows in remote-control sessions: killed tasks now stay visible and re-stoppable until their processes exit • Fixed resuming a workflow run while its previous stopped run was still exiting, which could run duplicate copies of its agents • Fixed marketplace repo URLs on http:// github.com with a trailing slash or dangling ?/# producing an unusable .git clone URL • Fixed blocking Stop hooks causing the turn after a block to lose the model's reasoning from that turn and, on some models, miss the prompt cache • Fixed remote ( http:// claude.ai) sessions taking 60 seconds to start a turn after a browser-hosted MCP server's page had gone away • Fixed worktree-isolated sessions refusing common Bash loops, xargs pipelines and launcher-wrapped commands that cannot reach the main checkout • Fixed remote and scheduled sessions doing nothing after a connector-tool permission prompt was approved while the session was paused Improvements: • Improved terminal resize and first-render performance for long responses by reusing text measurements • Improved /workflows agent detail: JSON outcomes are pretty-printed with syntax colors and real line breaks, and long outcomes fold behind an expand toggle • Improved headless/SDK session start: the first turn begins up to 50 ms sooner when MCP servers finish connecting • Improved /install-github-app to explain it is GitHub-only and point to the GitLab CI/CD docs when run inside a GitLab repository • Improved nested background subagent results to be saved in the parent subagent's transcript, so resumed subagents keep them and shared transcripts show the delivery Other changes: • Changed allowedMcpServers to govern only servers users add: a literal managed-mcp.json server your allowlist used to filter out now loads on upgrade; use deniedMcpServers to keep it off • [VSCode] Added an Active quick filter and a status filter menu (Needs input, Working, Completed) to the session list sidebar Source: https:// github.com/anthropics/cla ude-code/blob/main/CHANGELOG.md#21259 …

Our developer advocate @amit planted a bug in a Flask app, deleted a todo, and walked away. Managed Agents caught the crash, found the bug, wrote a test, fixed it, and opened a PR before he was back with a coffee. Now in private preview. https:// do.co/4xxXMvv

Our teams kept rebuilding the same LLM integration stack from scratch. So we built the AI SDK for Go: shared interfaces for streaming, tools, structured output, and multi-step agents that speak Vercel AI SDK protocol. Last month, we made it open source!

Jerry Liu

@jerryjliu0

1시간 전

We've benchmarked all the latest frontier models on hard document extraction tasks (with Fable 5.1 and Gemini 3.8 Flash coming soon!). ExtractBench is now live on @kaggle . 5.6 Sol leads the pack, followed by other OpenAI models, then Gemini 3 Flash and Opus 5. Come check it out! https:// kaggle.com/benchmarks/lla maindex-org/extractbench-leaderboard/leaderboard … If you're looking for a dedicated extraction tool with out of the box grounding, citations at the price/performance frontier, come check out LlamaParse: https:// cloud.llamaindex.ai

Jerry Liu

@jerryjliu0

4시간 전

Most document OCR solutions have parse (doc->markdown) and extract (doc + schema -> structured output) endpoints. Sometimes you want to extract information from a semi-structured document like a form, without spending the work defining on defining the exact schema. This is where LlamaParse form mode comes in. Simply upload any form, partially filled in or not, and we’ll not only extract out the markdown, but give you key-value pairs on the precise form fields and values. Docs: https:// developers.llamaindex.ai/llamaparse/par se/examples/enriched_forms/ … Signup to LlamaParse here! https:// cloud.llamaindex.ai

"Losing my religion" A beautifully written blog by professor @togelius on reckoning with the fear that AI will one day make human thought unnecessary. His proposal, complementary intelligence, argues that keeping humans in the loop is not a constraint on progress. Rather, it is the only version that counts as progress. Separately, Weston and Foerster's co-improvement paper makes a similar case from the other side, that this is also the faster and safer path.

OpenAI told House lawmakers that its engineers are building automated shutdown capabilities for its AI systems, per a company letter reviewed by Reuters. The letter follows OpenAI's disclosure that one of its agents escaped its digital container during a safety test and hacked into Hugging Face.

Wan 3.0 by @Alibaba_Wan lands at #3 in Image-to-Video Arena with 1481 pts! This release is a significant +53 pt improvement from Wan 2.7 at #10 with 1428 pts. With a 57% win-rate, Wan 3.0 is just 7 pts behind #2 Gemini Omni 1.1 Flash, and 16 pts behind #1 MiniMax-H3. Congrats to the @Alibaba_Wan team!

NVIDIA AI Infrastructure

@NVIDIAAIInfra

4시간 전

We’re excited to collaborate with @EnduroSat to advance orbital AI infrastructure. Together, we’re bringing NVIDIA’s full-stack platform—including NVIDIA Space-1 Vera Rubin Module, NVIDIA Jetson Thor, NVIDIA IGX Thor, and NVIDIA AI software—to EnduroSat’s FRAME satellites.