Latest news, beside the leaderboard

Vals AI audited all 2,698 coding tasks in the RL environments Xiaomi open-sourced with MiMo v2.6: in 1,795 (67%) the reference fix survives as unreachable Git objects. MiMo v2.6 finds and copies it, writes its own pack-file parser when git commands are blocked, and uses file mtimes when Git is removed.

Codex lead Tibo said on day 3 (encore) of OpenAI's ship weeks that Codex Cloud was silently re-shipped. Official cloud-environment docs now let an environment join a tailnet via Advanced > VPN (Tailscale is the only supported provider), so cloud tasks can reach private HTTP/HTTPS services, with IPv4 subnet routes, MagicDNS and split DNS; Enterprise workspaces can request OIDC for short-lived cloud credentials. Tailscale says it had no idea and published hardening advice such as tagging agent nodes tag:codex.

Anthropic's Python SDK 1.12.0 and TypeScript SDK 0.132.0 add computer_toolset_20260801 and browser_toolset_20260801: a single tools[] entry declares a whole family of desktop or browser actions, the SDK tool_runner executes them in order and stops at the first failure, and url_policy, file_policy and confirm hooks run before each call. Developers no longer hand-write the loop that maps clicks and keystrokes to commands; they subclass an abstract toolset and plug in their own VNC, DevTools or hosted-browser backend.

Theo (T3) open-sourced ts-rust (npm: tsc-rs), a Rust port of Microsoft's Go-native TypeScript compiler, checker and language server. Months of GPT-5.6 Sol / GPT 6 Astra runs cost over $400K in API-priced tokens and stalled near 84% compatibility; Claude Opus 5.5 restarted from scratch, reached a working v0 in 10 hours and finished in two weeks for about $24,047. All 181,711 ported Go tests pass, it type-checks 1.61x faster than tsc 7 (geometric mean over six open-source apps), ships Effect diagnostics built in and has a WASM build.

Liquid AI released two open-weight multimodal decision models: d1-3B (text+vision) and the experimental d1-omni-600M (text+image or text+audio). Instead of generating tokens they answer in a single forward pass. d1-3B scores 48.57 on Decision Index v0.2.1 (public split), first among sub-10B models and on par with the 12x larger Decider 35B-A3B, and answers in 8 ms on an RTX 4090 and 50 ms on a Jetson Orin Nano, with day-one llama.cpp support.

At its Oct 7 Windows event Microsoft shipped an on-device MAI-Code-1.1-Flash (137B total / 6.8B active MoE, 256K context, ~3.3 bits per weight) on RTX Spark PCs, scoring 70.80% SWE-Bench Verified and 66.29% Terminal-Bench 2.1 locally. GitHub HydraFusion will route Copilot tasks between local and cloud models in experimental preview later in October, with no inference charge for local calls; MXC agent containment went GA the same day.

On Oct 7 Nous Research, maker of the open-source Hermes Agent, confirmed a $90M Series B at a $1.5B valuation led by Robot Ventures, with NVIDIA, Microsoft's M12, Samsung, USV, Y Combinator and Menlo among backers, bringing total funding to $158M. Nous says Hermes Agent has been cloned over 24M times and drives about 2.5% of global token usage (internal estimate). The money funds Hermes for Businesses: a team tier with shared balance and skill library, and an Enterprise edition on customer-controlled infrastructure with SSO and SLAs, plus a mobile app.

Prevailing coding agent benchmarks evaluate isolated patch synthesis (e.g. SWE-bench) or vulnerability detection, neglecting whether agents can construct reusable, production-grade static-analysis checkers from scratch. Developing static analyzers requires interpreting defect specifications, traversing multi-file AST semantics, writing engine-specific logic, and refining implementations through iterative compilation and diagnostic feedback. Researchers from East China Normal University, SJTU, and Peking University unveil CheckerBench (arXiv:2610.07557), an executable suite comprising 300 tasks derived from 297 CVEs across 167 repositories, 85 CWEs, and five programming languages. Powered by the CheckerLab harness, evaluation across 21 model-harness configurations reveals a mean Pass@1 rate of only 32.30% (peak 45.33%), demonstrating that end-to-end program analysis tool synthesis represents a profound new frontier for software engineering agents.

Sparse Mixture-of-Experts (MoE) architectures are the de facto standard for long-horizon agentic systems due to their favorable activation ratios. However, the co-design of agentic trajectories and MoE routing topologies remains largely uncharted. Apple Machine Learning Research reveals a structural phenomenon (arXiv:2610.07332): off-the-shelf MoE models exhibit natural routing alignment with agentic operations, clustering expert activation patterns across semantically identical turns (e.g. READ, UPDATE, EXECUTE). Standard RL post-training neglects this specialization, permitting unconstrained routing drift that severely impedes policy convergence. Apple introduces a hierarchical routing control framework with entropy-gated stabilization, regularizing turn-level expert assignments to match agentic semantics while preserving token-level consistency. Evaluated across diverse benchmarks, the framework delivers over 10-point improvements in task success rate, demonstrating that agent trajectory semantics provide a critical inductive bias for MoE capacity optimization.

Reinforcement learning post-training generates massive rollout bottlenecks, driving interest in low-precision execution like NVIDIA Blackwell's native NVFP4 (W4A4). However, numerical discrepancies between learner and sampler forward passes destabilize policy-gradient optimization, triggering policy collapse. Researchers from Nanjing University and Polixir present TRIAGE (arXiv:2610.07043), isolating how quantization mismatches interact with policy-gradient directions. The authors reveal that native NVFP4 introduces an early asymmetry that amplifies negative-advantage updates across concentrated tail tokens. TRIAGE implements segment-level diagnostics to rebalance policy updates while preserving native W4A4 forward execution across both learners and samplers. Validated on Qwen3-4B and Qwen3-30B-A3B, TRIAGE matches full-precision BF16 performance across five mathematical reasoning benchmarks while delivering up to 2.3x higher rollout throughput.

Google Labs launched Playground on 2026-10-07: describe a game in chat, test it, and share it or publish it to the Explore gallery, with multiplayer and leaderboards in select genres. It runs on Gemini, Nano Banana, and Lyria with a custom harness. US 18+ first; weekly creation tokens with higher limits on Google AI plans; Unity Spark integration coming.

Grok Bot v0.68.1 (Oct 7, official changelog): slide decks delivered as PowerPoint or Google Slides, formatted email from draft cards, Gmail Spam search, Bot-colored 1:1 chats, and a 1920×1200 Bot computer (up from 1280×800) with faster actions. No official speedup figures.

On 2026-10-07 the official @bot account said Grok Bot can now search, read, and monitor X (feedback tracking, breaking stories, weekly roundups), available to all users with no X connector setup. The August v1 required linking X and an auto-created developer account. Quotas, posts per query, and posting support are not disclosed.

On Oct 7 OpenAI began rolling GPT-6 out globally in ChatGPT together with Intelligent UI: the model decides when to answer with diagrams, interactive charts, forms, tappable buttons or small in-chat tools (calculators, games, bill splitters) instead of plain text. Plus, Pro, Business and Enterprise get it first on GPT-6 Sol; Go and free users follow on Oct 8 on GPT-6 Luna. OpenAI says GPT-6 can stream partial answers while still thinking, cutting wait times by 44%.

On Oct 7 Anthropic shipped Claude Haiku 5.5 (claude-haiku-5-5), completing the Claude 5.5 family. It targets high-volume, cost-sensitive work such as summaries, compaction, classification and coding subagents, and Anthropic calls it its fastest model at standard speed. Prompts up to 100K cost $0.10 input / $0.50 output per million tokens, and it is the first Haiku with effort levels. It scores 72.4% on the OSWorld 2.1 offline subset (Haiku 4.5: 15.7%) and 39.2% on Terminal-Bench 4.0. Sonnet 5.5 cache reads were cut 50% the same day, and Max/Team plans get monthly API credits.

On Oct 7 AWS released Strands Box v0.1.0, an open-source (Rust, Apache 2.0) sandbox for AI agents. It pairs OS-level isolation with the Dogwood policy language: shell, Python, outbound HTTP and MCP calls all pass through one default-deny policy engine with a shared event history, so rules can depend on earlier actions and elapsed time. A gateway injects credentials so the agent never sees the secrets. Preview supports macOS on Apple silicon only; Linux is in development.

On 2026-10-06 the FCC granted-in-part SpaceX’s D2D Constellation (S00735) under DA-26-1078: up to 15,000 sats at 326–335 km, with a §25.125 lease waiver plus PCS G and EchoStar-assigned bands. Milestone: 50% by 2032-10-07. ~150 Mbps/user is a company target, not an FCC measurement.

REA (Reverse Engineer Anything, MIT) wraps Hopper, Ghidra and IDA Pro into an MCP server and CLI for coding agents: 41 native inspection tools and 14 investigation workflows across Mach-O/ELF/PE, .NET, Electron, websites and Android APKs, all running locally. Version 4.1.0 (Oct 6) adds headless-JADX APK analysis and Binwalk/Unblob firmware analysis. The repo has 12,962 stars and 1,375 forks.

On Oct 7 Perplexity released pplx-embed-v2-late, multimodal late-interaction (ColBERT) retrievers in 0.6B (340M active) and 9B (7.4B active) sizes, built on Qwen3.5 with bidirectional attention. Each token gets a 128-dim vector scored by MaxSim, covering text, images and rendered PDFs or slides without OCR. Both sizes share one embedding space, so the 0.6B model can query a 9B-built index. Public ViDoRe v3 image nDCG@10 is 62.3% / 65.2%; MIT-licensed weights are on Hugging Face.

Perplexity released pplx-decider-v1.1-27b (Apache-2.0, fine-tuned from Qwen3.8-27B) on Hugging Face. It returns typed, probability-calibrated answers instead of text. Overall Decision Index is 61.56 vs 56.4 for v1 and 57.9 for Jev, and it is served via the Perplexity Decisions API.