AITOP WeeklyIssue 10885 stories
Anthropic's $1.07 Trillion Compute Bill, OpenAI's Pando, Mistral ML4
Singapore·2 Nov – 8 Nov·Ten minutes
This week was about infrastructure at a scale few can underwrite. OpenAI unveiled Pando, a Git-compatible backend built for agent-scale code pushes, cutting median push wait times by roughly 80% while distributing nearly 1PB of code daily. Anthropic's amended IPO filing showed Q3 revenue of $16.96 billion, up 1,127% year-on-year, but compute commitments reaching $1.07 trillion, roughly 32 times revenue, with institutional investors pushing for a $1.5–1.6 trillion valuation. Mistral released open-weight ML4, targeting coding agents and cybersecurity. Marvell's ~$115 billion all-stock acquisition of Lumentum set a semiconductor record. For Singapore, the compute-and-code arms race now decides which AI services actually land here.
01
Five things this week
Enough to retell1
OpenAI 自研 Git 服务器 Pando，Anthropic 则用 Scion 替换 Git 应对 agent 写码潮
OpenAI 与 Anthropic 分别重做或替换 Git，只为扛住 agent 写码速度。
宝玉
2
3
4
5
RAGFlow Harness Hits 91% on FRAMES With BM25 and Grep Only
Accuracy on agentic retrieval benchmarks may already be saturated, so the real design trade-off is token cost versus latency rather than recall. Builders of agentic RAG systems should benchmark indexes on tokens and latency instead of assuming richer structures lift accuracy. The 38% average token reduction with a decision gate keeping latency near 30 seconds offers a concrete configuration to test against dense-retrieval baselines.
DAIR.AI
02
US vs China
The same week, both sides| Dimension | United States | China | What it means for ASEAN |
|---|---|---|---|
| New models | OpenAI ships GPT-6 Astra for embodied engineering; Mistral open-weights ML4 for coding agents | DeepSeek V4.1 Flash 1017 tops three ARC-AGI benchmarks; Tencent open-sources WeVisOne-VL | Rich open-source pipeline means more self-hosting options for ASEAN |
| Prices | Claude API launches Message Threads beta; long-conversation cost structure unchanged | DeepSeek Search/Fetch end free tier, now billed at $0.75 per 1,000 calls | Chinese tooling starts charging; rework your cost models |
| Benchmarks | MIT uses Astra to edit Qwen3.5-0.8B weights; MATH-500 jumps from 7.9% to 48.1% | Tencent Hunyuan Hy4 preview ranks #2 open model on Agent Arena, +5.4% net | Leaderboards now genuinely useful; check them before choosing models |
| Policy | Anthropic S-1 reveals five-tier equity; founders retain 50.1% voting power | DeepSeek v5 reportedly training almost entirely on Huawei chips, accelerating de-Americanization | Compute supply splitting; ASEAN buyers should hedge both sides |
885Stories
147China
2ASEAN
156Sources
AITOP · generated by the editorial system