Today's thread is a scissors: the price of intelligence is collapsing while the price of the hardware that makes it is exploding. OpenAI cut its API rates in half, Xiaomi trained the best open-weights model on the planet in six days for less than $3 million, Alibaba showed off a chip built for 500,000-card clusters, and AMD crossed a trillion dollars in market value in a single session.
OpenAI launches GPT-6 Sol and Luna at half the price
- Released September 22, GPT-6 Sol costs $2 per million input tokens and $10 per million output, down 50% from $4/$20; GPT-6 Luna costs $0.10 and $0.50, a 58% cut on output. Both carry a 1.1-million-token context window, and OpenAI says these are permanent prices, not promotional rates.
- Sol scores 68.8% on DeepSWE 1.1 and 60.5% on OSWorld 2.0, matching Opus 5 on computer use at roughly 80% lower cost per task and beating it on AutomationBench at 9% of the cost. Luna hits 66.6% on coding for 93% less per task. Both ship API-first, plus ChatGPT Work and Codex for paid plans.
Why it mattersFrontier intelligence got 50% cheaper overnight. The products that were too expensive to run last week — agents that work for hours, support bots that read entire case files — just moved into budget, and that math is hardest to ignore in markets where the software has to earn its keep in dollars.
Xiaomi's MiMo-V2.6-Pro is now the best open model in the world
- MiMo-V2.6-Pro — 1.02 trillion total parameters with 42 billion active per request — scored 46 on the Artificial Analysis Intelligence Index, first place among open-weights models, ahead of DeepSeek V4.1 (36) and level with xAI's closed Grok 4.7. It takes text, image, audio and video, with a 1-million-token context window.
- Xiaomi says it trained the Pro in six days across 30 reinforcement-learning steps and about 750,000 trajectories for roughly $2.62 million; the smaller Flash (310 billion parameters, 15 billion active) cost about $850,000. Both are MIT-licensed and free to download from Hugging Face, at $0.435 per million input tokens for Pro.
Why it mattersA phone maker reached the open frontier for the price of a Miami apartment. When the best open model on earth is a free download under an MIT licence, any company that cannot or will not send its data abroad can run frontier-grade AI on its own servers — and pay nobody for the privilege.
Alibaba unveils the Zhenwu V900, 'China's most powerful AI chip'
- Shown Tuesday at the Apsara conference in Hangzhou, the V900 accelerator carries 216 GB of on-chip memory and 1,200 GB/s of chip-to-chip bandwidth, up from 144 GB and 800 GB/s on the previous M890, with roughly triple the performance. Mass production was pulled forward to Q1 2027 from Q3 2027.
- Alibaba says it scales to superclusters of up to 500,000 cards and is built to train Qwen models of 5 to 10 trillion parameters, against the 2.4 trillion of today's Qwen3.8-Max. It has already shipped over 650,000 Zhenwu chips to more than 400 customers and targets over 20 GW of data-center capacity by 2032.
Why it mattersThis is China answering Nvidia's export controls with its own silicon, announced precisely as Washington and Beijing sit down to talk about AI. If the V900 works as claimed, the cheap inference that reaches Latin America over the next two years may well be running on Chinese hardware.
AMD crosses $1 trillion, beating Intel to the club
- On Monday September 21 AMD passed a trillion dollars in market value, with shares around $615 after a gain of nearly 10% on the day, 29% over the past month and 21% of that in five sessions. Intel, which rose over 12% the same day, still sits at roughly $648 billion.
- The move rode a Nasdaq up more than 2% on semiconductor and AI names, with Meta up 11.4%. AMD now trades at a price-to-earnings ratio near 158, against a five-year median of 102 — GuruFocus puts the stock 117% above its estimated intrinsic value of $283.63.
Why it mattersThe market is no longer paying only for Nvidia: the second source is now worth a trillion too. That is a bet that AI demand will outrun any single supplier — and a valuation with no room for a bad quarter, which matters to every LatAm saver whose pension or ETF is quietly long semiconductors.
The big picture
Intelligence got cheap and iron got expensive, and that scissors is the whole story of the day. OpenAI charges half of what it did on Monday, Xiaomi gives away the best open model on the planet, and the margin has migrated downward — to the silicon and the electricity, where AMD is worth a trillion and Alibaba is buying 20 gigawatts. For Latin America the reading is unusually direct: building on top of AI has never been cheaper, and building underneath it has never been further out of reach.
Go deeper
Sources & further reading
- VentureBeat — OpenAI releases GPT-6 Sol and Luna
- MarkTechPost — GPT-6 Sol and Luna pricing and benchmarks
- VentureBeat — Xiaomi MiMo-V2.6-Pro tops open weights
- Tom's Hardware — Alibaba unveils Zhenwu V900
- Quartz — Alibaba unveils Zhenwu V900 AI chip
- Yahoo Finance — AMD beats Intel to the trillion-dollar club
- GuruFocus — AMD surpasses $1 trillion market cap