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: The AI race is shifting from peak scores to cost-adjusted capability and the systems that make agents reliable.
Open models are winning a cost-and-adoption test. Together Compute reports four tries with GLM-5.3 on DeepSWE reaching 87.6% for about $16, versus Fable 5 at 69.7% for $21.63. A separate breakdown puts the models near parity—69.0% versus 69.7%—at $3.99 versus $21 per task, with GLM-5.3 using 80k versus 114k tokens. This is one benchmark, not a universal ranking, but Vercel AI Gateway says open-weight models supplied 62% of tokens on Aug. 22, up from 28.4% on June 24; Vercel also says enterprise adoption and model-agnostic tooling remain early. Portability and unit economics are becoming first-order competitive variables.
Harnesses are becoming part of the product. A monitored analysis of Pi’s development notes says Claude models sometimes invent parameters for Pi’s edit tool, suggesting that post-training is increasingly coupled to Claude Code’s harness and schema; it frames Claude+Claude Code, DeepSeek+DSH, and GPT+Codex as emerging whole systems. Pi’s “prune + spill” approach stores full tool results on disk while keeping only a slice in context; across 19 sessions it reports 26–35% lower context use and 72–88% lower uncached prefill, with information recoverable.
Research & Innovation
Why it matters: New work is testing whether agents complete real state-changing work and training tool use earlier instead of assuming post-training will fix it.
Thinkingbox makes reliability an end-state test. Microsoft’s paper introduces an MCP-compatible sandbox and 507 policy-conditioned workflows spanning retail, hospitality, auto insurance, neobank IT, and consulting support. It grades executable backend state and rejects wrong, missing, or extra effects; the strongest model reached 65.36% pass@1 but only 25.25% pass^20, while many failed runs looked clean at the response or tool-call level.
MidTool moves tool use into mid-training. Snowflake’s corpus combines web, PDF, and code data with API, MCP, and document-grounded supervision; it was used to mid-train Qwen3-4B and Qwen3-8B. In the reported results, 4B BFCL rose from 39.51% to 54.18% after RL and τ²-Bench pass@1 from 13.04% to 19.96%, but every model scored 0% on MCP-Universe’s web-search subset. Tool familiarity improves function calling without automatically solving long-horizon research.
Matryoshka nests model sizes in one suite. Cornell researchers stack 500M, 1.5B, and 3B submodels in one end-to-end architecture; the paper reports parity with independently trained baselines, 36% less training compute, and 14–26% faster speculative-decoding throughput.
Products & Launches
Why it matters: AI interfaces are expanding from single-user generation toward controllable, collaborative workflows.
Krea’s Seedance Studio uses Seedance 2.5 and new 3D scene controls to take a project from character design to final cinematic footage in one workflow, according to a current creator demonstration.
ChatGPT may be adding a social layer. Strings in the latest Android app mention “ChatGPT with Friends” for sharing responses, images, and creations, plus private side chats; the observer presents it as a possible next iteration of group chats, not a confirmed release.
Industry Moves
Why it matters: Corporate strategy is increasingly framed around open ecosystems, model ownership, and repeat enterprise usage—not just model releases.
Anthropic’s IPO expectations are escalating, but the report is prospective. A current account says the company’s bankers are telling potential investors it may raise more than $100 billion at a $2 trillion valuation, which would make it the largest IPO ever.
Poolside is being positioned as an open-model US ecosystem play. The Wall Street Journal reports a sweeping agreement intended to build an open AI ecosystem that can compete with Chinese heavyweights and American AI giants; Ollama says it collaborated with Poolside engineers on open models and points to NVIDIA’s Nemotron work.
Runway is packaging video generation as an enterprise operating layer. Its company announcement says the business more than doubled this year and NRR exceeded 300%; its roadmap includes day-one access to third-party models, a media model router, Runway Agent, and customer-hosted model licensing. These are company-reported figures and plans.
Quick Takes
Why it matters: Physical AI and agent tooling are moving from isolated demos toward repeatable systems.
- NVIDIA says its coding harness solved all 183 levels across ARC-AGI-3’s 25 public games.
- The 2026 World Humanoid Robot Games opened with 666 teams and more than 2,000 humanoid robots.
- Jerry Liu’s market framing: SaaS is not dead, but it must be repurposed and remonetized for agent consumption.