We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Top Stories
Why it matters: Agents are increasingly competed on cost, latency, and scaffolding—not only model capability.
Inference economics is moving into the product surface. OpenAI cut GPT-5.6 Sol API prices 20% for input and 33% for output, to $4/$20 per million tokens through at least Nov. 21. Arena reports that the cut moved Sol’s frontier position in Code and Work, while Luna reached the overall and all three category frontiers at $0.04–$0.08 per task. Artificial Analysis also measured 3,431 output tokens/s for Gemma 4 31B on a private Groq 3 LPX endpoint, stable from 10K to 100K input tokens; NVIDIA says the rack is in full-scale production.
Harness choice is now a measurable competitive variable. In a same-model Opus 4.6 task—cloning Excalidraw from browser inspection through verification—Factory Droid finished in 8 minutes with 20 tool calls at $1.60, versus Claude Code’s 19 minutes, 40 calls, and $1.87. The tester says the harness was the only difference; it is one controlled comparison, but a clear cost and execution lever.
Physical AI posted a sharp benchmark jump. Team Tianjiao’s robot won the World Humanoid Robot Games long-jump final at 7.97 meters, 0.98 meters below Mike Powell’s record and well above the 1.25-meter robot best reported for 2025. It is a concrete athletic-control result, not evidence of general robotics.
Research & Innovation
Why it matters: The highest-leverage work is moving into the agent loop—serving, evaluation, and recoverable execution state.
AgentX 1.0 makes agentic inference measurable. SemiAnalysis released an open-source multi-turn coding benchmark built from roughly $3 million of real traces across 1,000-plus chips and about 2 MW of compute. vLLM reports sparse KV retention above a 95% cache-hit rate for 14 concurrent requests with contexts up to 1M tokens, while rate-matched prefill/decode reached 4.45× the throughput at 60 tok/s on GB300 Dynamo versus B300.
Video serving is being redesigned, not merely scaled. NVIDIA SANA’s Sol Engine work on MiniMax H3 combines a four-step low-resolution draft with a three-step LTX refinement pass and reduced 10-second 768p generation on one GB200 from 414 seconds to 14.93 seconds—a reported 27.7× speedup.
ACES questions static agent-skill gates. Across 145 real skills, structural scan scores correlated with LLM-judge quality at only Spearman ρ=0.14. The proposed Skill Lift instead compares the same task with and without a skill under identical conditions; its evaluation covered 947 paired cases, 58 production skills, and four harnesses.
Products & Launches
Why it matters: Agent products are becoming cross-provider control planes and enterprise-ready connector layers.
AgentSky launched an “OpenRouter for Agents.” Its API connects Claude Code, Codex, DeepSeek, Kimi, OpenCode, and other cloud agents; Agent Playground runs identical tasks with real tools such as GitHub and Gmail while comparing time, cost, and tokens side by side. The launch reports a $150-versus-$2 same-task gap but leaves the cheaper system for readers to guess.
Claude’s enterprise-managed MCP authentication is generally available. Admins centralize authorization through an identity provider, users connect tools without individual OAuth, and developers can apply the system to third-party connectors in Claude’s directory.
Industry Moves
Why it matters: Organizations and governments are pairing open models with proprietary data and dedicated compute rather than relying entirely on frontier APIs.
Thomson Reuters is pursuing model ownership. DatologyAI says Thomson-1.0-Large used its domain-relevant proprietary data, cost $450,000 in compute, and is competitive with closed frontier models at a fraction of deployment cost. A current report says Thomson Reuters built the model on Alibaba’s Qwen to reduce reliance on Claude.
South Korea is funding a narrowed sovereign-model race. Its government-backed competition reduced Round 2 from four teams to Upstage, SK Telecom, and LG AI Research. Each advancing team is expected to receive roughly 1,000 NVIDIA B200 GPUs for six months—about ₩40 billion per team—with the field planned to narrow to two in early 2027.
Policy & Regulation
Why it matters: Regulatory scrutiny is reaching the financial infrastructure funding AI bets.
- The SEC is examining an AI hedge fund’s financing and leverage. MTSlive reports, citing the New York Times, that the agency subpoenaed banks that lent to and traded for Situational Awareness and ordered them to preserve related records.
Quick Takes
Why it matters: Practical deployments continue to spread across climate response, defense, and everyday AI operations.
- Flood forecasting: Google says Flood Hub and Groundsource forecast riverine floods up to seven days ahead and urban flash floods up to 24 hours, with alerts intended for 2 billion people across 150 countries.
- Defense autonomy: A team building autonomous interceptors in Ukraine reports 45 units sold to a special-forces unit and $500 million in letters of intent, all within eight weeks.
- Usage controls: ChatGPT Work and Codex will restore a five-hour Plus-account limit to smooth compute demand and prevent accidental exhaustion of weekly usage; the $100 and $200 Pro tiers remain exempt for now.
- Team Tianjiao’s humanoid robot won gold in the long-jump final at the 2nd World Humanoid Robot Games with a 7.97-meter jump . A related commentary post says this remains 98 centimeters below the male human record and predicts that robots clearing 9 meters could no longer be surpassed by baseline humans; the latter is speculative rather than an observed result .
- Apodex introduced the Apodex 1.1 model family, claiming frontier-level agentic performance for complex professional work, scientific research, financial analysis, and deep search.
- The release adds an asynchronous Agent Team that decomposes tasks, coordinates agents in parallel, continuously integrates findings, and allows user guidance; Apodex also open-sourced the locally deployable FrontierAgent research workbench. Apodex 1.1 is available through its online workbench, while 1.1 mini is offered as an open-weight model for local execution.
- A separate post says Apodex 1.1 mini is based on Qwen 3.5 35B A3B MoE.
- MiniMax H3 is supported in vLLM-Omni for text-to-video with synchronized sound, frame-conditioned video generation with audio, image-plus-audio lip-synced clips, and automatic relighting/compositing of green-screen footage; vLLM published a deployment recipe.
- The surrounding H3 ecosystem includes 24GB-VRAM local ComfyUI setups, enterprise SGLang and vLLM-Omni deployments, quantization guidance down to 8GB VRAM, serving optimizations, and developer tooling.
- Unlimited-token AI coding workflows are described as enabling unusually high productivity, including work across “8–10” projects at once—characterized as consistent with the baseline expectations of Staff+ engineers even before AI.
- The discussion highlights a growing gap between unlimited-token frontier workflows and the economics of token usage, while framing agentic coding as a rapidly advancing frontier.
- The SEC subpoenaed banks that lent to and traded for Situational Awareness, an AI hedge fund, seeking details on its trades and leverage and ordering them to preserve all related records, according to the New York Times.
- Alex Zhang introduced Speculative Programmatic Tool Calling (sPTC), a general technique that speculates on tool calls during code generation and queues them early in a harness to overlap token generation with REPL execution time.
China’s Galileo robotic dog is described as changing shape like a Transformer to adapt to tough environments and varied tasks. The post’s author characterizes the demonstration as a “brutal flex of industrial capability” and evidence of who is a “live player.”
Ben Springwater argued that even AI adopters have not fully grasped what computer/browser use can do, pointing specifically to “5.6 Sol High in Codex” as an example. @mckbrando replied “correct,” endorsing the assessment.
A low-volume prediction market currently attributes the “0x-alpha” model to Zai_org, while noting that insiders may have incentives to leak information . Commentary cautions that participants may also try to misdirect others before betting, so this is unverified speculation rather than a confirmed model release .
An analysis of roughly 500,000 arXiv AI/ML papers reports a major shift toward Chinese open models in research: about 40% of papers now mention a Chinese open LLM versus 25–30% mentioning an American open model, reversing the 2024 split of roughly 10% versus 30%. The analysis cautions that papers lag model releases, so the figures reflect delayed adoption. Qwen appears in roughly one-third of papers that mention any LLM, while OpenAI’s closed models lead at about 37%; Llama peaked near 30% in April 2025 and has declined since, while DeepSeek saw a clear increase after R1’s January 2025 release. The share of AI papers mentioning any LLM rose from 10.43% in January 2023 to 55.49% in January 2026.
- A Team Tianjiao humanoid robot won gold at the 2nd World Humanoid Robot Games in Beijing with a 7.97-meter long jump—just 0.98 meters short of Mike Powell’s 8.95-meter men’s world record. The best robot mark in 2025 was reportedly only 1.25 meters, signaling a sharp improvement in humanoid athletic performance.
- Jerry Liu argues that software and systems of record should become agent-native: agents should be able to use existing tools such as Slack, rather than build replacements or rely on the tool’s own agent. He expects agent consumption of software to become exponentially larger than human consumption.
- Garry Tan predicts that systems of record will need to become “AI harnesses” or risk replacement by agents.
@teortaxesTex argues that DeepSeek has fallen behind two other startups because it relies more heavily on internal RL, and characterizes the gap between “0731” and “ox-alpha” as “Opus-shaped.” The same account alleges that z.AI may have privileged access to a transfer station and Claude trajectories, while explicitly framing the claim as opinion.
A medical-AI practitioner says their team trains personalized-medicine models on rented GPUs, starts from Meta’s DINOv2 model trained on billions of Instagram images for cancer-detection tasks, and uses ChatGPT and Claude to accelerate medical-AI algorithm development. They argue that a blanket ban on generative AI could stall medical-AI progress.
- OpenAI announced that GPT-5.6 Sol API and credit pricing will fall by more than 20% for the next three months, linking the reduction to continued capability gains and improved efficiency.
- @scaling01 speculates that Anthropic could cut its pricing within a month because its current pricing is viewed as excessive, and questions whether Model 2 improved Anthropic’s inference stack; this is commentary rather than a confirmed announcement.
- Hermes Agent parallelism: Teknium disputes a claimed one-agent-in-parallel limit, reporting that he runs about 25 sessions concurrently, with each using 15–25 subagents, and asserting that Hermes has no such limits.
- Enterprise-agent workflow thesis: @imjaredz argues that mobile could be the “ultimate form factor” for enterprise agents because workers are awake beyond desk hours; the post predicts an asynchronous model in which cloud agents run continuously and constantly ask humans for unblocking and steering.
Figure says it is “accelerating again” and undergoing another “step change in capabilities”; it plans to announce a “critical update” tomorrow that it describes as necessary to solve general robotics.
- Alibaba’s Qwen3.8-27B ranks #9 overall in Code Arena: WebDev with 1,595 points, making it the only model in its size class in the top 10; it trails the much larger Qwen3.8-Max by six places.
Hark says it is building models intended to use computers better than humans, while its Handoff system operates across the internet—the dominant environment for computer use—rather than within a single application.
