Monday, August 10, 2026 — Inference costs are the story today: DeepSeek signals a big API price hike, Amazon reportedly burns through money running Claude at scale, and memory-chip shortages keep squeezing the whole stack. Meanwhile, GPT-5.6 helped crack a 25-year-old math problem, Meta shipped a serious coding agent, and China flipped on its first fully domestic 100,000-card AI supercluster.
Top 3 Highlights
1. DeepSeek says API prices are going up — and by a lot
DeepSeek announced it will "substantially" raise API pricing soon, telling developers to plan their budgets accordingly. The context: OpenCode data shows DeepSeek V4 Flash alone processed 8 trillion tokens on a single day, with 5 trillion from free trial credits and 3 trillion from paid calls. The era of near-free frontier inference is clearly winding down.
2. GPT-5.6 and Fable crack a math problem that stumped mathematicians for 25 years
A collaboration between GPT-5.6 and the Fable framework solved a problem the paper's author had been working on since his PhD days — 17 years of human effort, closed by AI. It's another sign frontier models are moving from benchmark trivia into genuine research contribution.
3. Meta ships Muse Code — a coding agent built for large codebases
Meta's new terminal-based agent, built on the Muse Spark code model, handles full software-engineering tasks across big repositories: planning, writing code, and verifying results. It spawns multiple agents in parallel for large projects and installs with a single command. Meta is clearly racing to close its gap in the AI application race.
More News
- Amazon's Claude bill is getting painful. Reports put Anthropic usage at around $1.8M per day for Amazon — a number that's unsustainable even for a cloud giant, and a sharp reminder of how expensive frontier models have become. QbitAI
- ByteDance bans distillation of open-source models. At an internal meeting, Zhang Yiming pushed "long-termism over leaderboard-chasing" — no distilling others' outputs for short-term rankings. The company is even enforcing it via API detection. 36Kr
- China's first fully domestic 100,000-card AI supercluster goes live. Marking a major step for domestic compute infrastructure as the country builds out a national "one network" of compute resources. 36Kr
- Zhongke Huisi debuts three dexterous hand products. The new embodied-intelligence company is pushing dexterous hands from mere "grasping" toward real-task operation — with modular fingertips and data gloves building a flywheel of operation data. 36Kr
- Google orders AI core staff back to Silicon Valley. The company is pulling key AI employees into office-based roles in the Bay Area, and reportedly spending another $1.5B to acquire an AI coding team. QbitAI
Trend Watch
The cost curve is bending upward. Between Amazon reportedly burning $1.8M/day on Claude, DeepSeek raising prices, and DRAM/memory shortages that suppliers say will last into 2027, cheap tokens are no longer guaranteed. Expect pricing pressure to ripple through every layer of the stack — apps, agents, and cloud bills alike.
Agents are moving up the value chain. From chat assistants to full software engineering (Muse Code) and on to embodied operation (dexterous hands that perform real tasks), the definition of "what AI can actually do" keeps expanding — and so does the engineering depth required to make it reliable.
Disclaimer: This briefing is compiled from public sources and is for informational purposes only. It does not constitute investment advice.