AI Digest August 4: An OpenAI Model Cracked 10 Open Problems for $2,000
An internal OpenAI model produced 10 new results on open math problems at roughly $2,000 in tokens. Alibaba announced the open 2.4T-parameter flagship Qwen3.8-Max, MiniMax H3 topped Video Arena among open models, and GPT-Live learned to listen while speaking.
Today's top stories
OpenAI: an internal model produced 10 new results on long-standing open problems. According to the company, an internal version of its next major model delivered 10 new results on long-standing open problems in mathematics and theoretical computer science — at roughly $2,000 in token cost. The story spread across the newsletters: The Rundown's breakdown and Superhuman's recap. The bigger question it raises: what happens when model breakthroughs outpace experts' capacity to verify them.
Alibaba announced Qwen3.8-Max — an open 2.4T-parameter flagship. Open weights are promised next week (alongside the compact Qwen3.8-27B), priced at $2/$6/$0.25 per million input, output and cached tokens. First independent numbers: #4 in Frontend Code Arena, #2 in Vision Arena, 66.1 on the Vals Index — #2 among open-weight models. The industry backdrop: the open frontier increasingly belongs to Chinese labs — Kimi, Qwen, DeepSeek, GLM, MiniMax; Vals calls DeepSeek V4 Flash 0731 the cheapest model on its index scoring above 60 — 35x cheaper than the next best at that threshold.
MiniMax H3 — the open leader in video generation. Video Arena ranked H3 #1 among open models with a +280-point lead, and effectively #1 overall in image-to-video. An important license clarification: the US/EU/UK/Korea restrictions turned out to be a formal authorization process, not an outright ban — a meaningful distinction for teams evaluating deployment (the original concern).
OpenAI rebuilt realtime voice: GPT-Live listens while it speaks. The new architecture separates a dedicated fast audio path from slower asynchronous reasoning and tool paths, and cuts session startup from six network round trips to one.
Numbers and facts
- 10 results on open math problems at ~$2,000 in token cost (OpenAI).
- Qwen3.8-Max: 2.4T parameters; $2/$6/$0.25 per million tokens (Alibaba Qwen).
- DeepSeek V4 Flash 0731 — 35x cheaper than the next model on the Vals Index scoring above 60 (Vals).
- Kimi K3 serving: 239 tok/s on Databricks (measurement) and 947 tok/s decode at batch 32 on a single B300 node (measurement).
- Zero-Mem: 57.6% lower agent-memory operation cost with no quality loss (digest).
Different perspectives
Elon Musk argued that source code is about to become "like assembly" — AI will compile intent straight to binaries. The skeptical frame from the same day's research: a fresh taxonomy of 41 agent failure modes shows breakdowns arise at the seams between model, harness, tools and memory — meaning "direct intent compilation" is bottlenecked by systems engineering, not model IQ. The middle ground: long-horizon agentic capability is co-evolution between models and runtimes, and it is too early to bury code.
Tools and techniques
- Let the model work the way it was trained. Cline observes that many open models are RL-trained to double-check their work — rerunning tests, rereading diffs — and not fighting that behavior yields roughly 20% gains from harness changes alone.
- Memory without LLM calls. Zero-Mem moves agent-memory maintenance out of the LLM: the model is only invoked for the final answer.
- Cursor plugged into Google Workspace. Plugins for Gmail, Drive, Calendar, Docs and Sheets — in-editor agents get access to your working documents.
In brief
- Cloudflare introduced @cloudflare/computer — an agent runtime switching between lightweight isolates and full Linux containers.
- Photon 2.0 compiles models into megakernels — the whole forward pass as a single GPU program.
- TokTier speeds up agent-session tokenization: 16–34% lower time-to-first-token under vLLM.
- Jina released reranker v3.5 at 0.6B parameters — 63.20 nDCG@10 on BEIR at ~7x fewer params than its competitor.