TODAY IN FIVE LINES
● Haiku 5.5: $0.10 / $0.50 under 100K tokens. Five times that above it
● GPT-6: Intelligent UI reaches Free and Go globally today
● Chip debt: ~$90B sought across Broadcom, Oracle and SpaceX in one week
● China: 16 models in September, 44-day release cycle, down from 125
● Power: Google–Constellation 3,590 MW, about a quarter new nuclear
The price floor got set twice in one day
Anthropic released Claude Haiku 5.5 on 7 October at $0.10 input and $0.50 output per million tokens. GPT-6 Luna is priced at exactly the same two numbers. When two labs independently land on identical pricing to the cent, that is a floor, not a coincidence.
The caveat worth carrying: Haiku's rate applies below 100,000 tokens. Above it, $0.50 and $2.50. Anthropic's own realistic workload estimate is about 75% savings, not the 90% headline. Sonnet 5.5 cache reads also quietly halved to $0.10.
Full pricing and benchmarks →
ChatGPT stopped answering in prose
GPT-6 Sol went to paid tiers on 7 October; GPT-6 Luna reached Free and Go globally today. Replies now render interactive charts, editable graphs, forms, buttons and small tools. Streaming thinking cuts average wait by a claimed 44%.
Google shipped the same idea as Neural Expressive in May 2026. The open question is what happens to source citations inside a rendered visual — nobody has explained it.
What changed and what to check →
Roughly $90 billion of chip debt in a week
Broadcom is arranging $50B+ in private credit for OpenAI's Nexus chip. Oracle wants an off-balance-sheet vehicle to buy and lease back hardware for a 1GW site. SpaceX is after ~$40B for Nvidia silicon. Apollo appears in all three.
The structures, and why they are private →
Also this week
| What |
The number |
| Chinese model releases, September |
16 models. Average cycle 44 days, down from 125 |
| Google–Constellation power deal |
3,590 MW, about 25% new nuclear capacity |
| Etched valuation talks |
$40–50B, roughly double its September round |
| South Korea sovereign AI |
~$3.49B, plus a probe after AI-driven hacks hit 7 financial firms |
| Meta child-safety enforcement |
33.2M pieces actioned in H1 2026, with a new ad-destination classifier |
| Surface RTX Spark Dev Box |
$5,999, preorders open, shipping November |
Three worth reading past the headline
OpenAI's 372 maths results. OpenAI says an unreleased model produced 372 new mathematical results, nearly all from a single prompt to one agent. It has not released the prompts or per-problem compute times, so the claim cannot currently be assessed by anyone outside the company.
Meta and Sierra published the Personal Agent Protocol, an open standard for AI shopping agents, with Walmart and Shopify backing it. An open standard with that much retail weight behind it tends to become the standard.
A Cisco and Carnegie Mellon study found eight of thirteen models recommended more expensive options to profiles signalling higher wealth, even when the user explicitly asked for the cheapest. That is a procurement problem, not just a research finding.
One safety story
Common Sense Media rated ChatGPT for Teens an "Unacceptable Risk" after testing more than 4,000 prompts, reporting that crisis-prompt hotline referrals fell from 33% to 23%. OpenAI disputes the methodology. Separately, OpenAI's first teen-use report put average teen usage at under 15 minutes a day. Both figures are worth treating as contested until the test set is published.
FAQ
What is the cheapest capable model right now?
Claude Haiku 5.5 and GPT-6 Luna are tied at $0.10 input and $0.50 output per million tokens. On published benchmarks Haiku 5.5 leads, but Haiku's rate only holds below 100,000 tokens per prompt.
Do I have GPT-6 yet?
If you pay for ChatGPT, since 7 October. If you are on Free or Go, since today, 8 October, rolling out globally.
How fast are Chinese labs shipping?
Sixteen models in September, on an average release cycle of 44 days — down from 125 days. Nikkei reported the figures.
Should I switch to Haiku 5.5 today?
Check your typical prompt size first. Under 100,000 tokens the economics are compelling for classification, extraction and routing. For long-horizon agent work Sonnet 5.5 still scores 70.6% on Terminal-Bench against Haiku's 39.2% at maximum effort.
Sources