Opus 5.5 leads the board; GPT-6 Sol and Luna halve the price

📰 News

Opus 5.5 sits at 58 — Fable 5.1 and GPT-6 Astra are at 53. $4 in, $20 out per million tokens, and the running bill is about 40% lower. Less of that padded Claude voice. It landed 90 minutes before OpenAI. → The Decoder
GPT-6 Sol and Luna are half price — Sol is $2/$10, Luna $0.10/$0.50. Scores stay near the old ones, 48 and 37. Fewer factual slips, but knowledge-work scores dropped. → The Decoder
Xiaomi leads the open models — MiMo-V2.6-Pro scores 46. $0.435 in, $0.87 out. A trillion parameters, 42 billion awake per request. Anthropic says Xiaomi pulled Claude’s answers to train it. → The Decoder
A hundred math problems in a month — Training started August 28. OpenAI also claims the Hodge conjecture. Advisors can discuss how results are shared, not how fast the lab moves. → The Decoder
Rules for AI that researches AI — OpenAI wants international standards before systems build the next model. Fully automatic is not here, they say. Measure it, report incidents, keep a person in charge. → The Decoder
The neurons stay in the dish — A thin layer copied from how lab cells spread a signal. The pitch is video that’s 5× faster and 80% cheaper. They never name the base model. → The Decoder
Laptop-sized models, in Transformers — llama.cpp GGUF files load through the usual API. Apple silicon first. → Hugging Face

BuilderIO/agent-native — TypeScript. The agent and the screen call the same actions.
trycua/cua — Desktops, drivers, and benches for agents that click around a computer.
akitaonrails/ai-memory — Memory as markdown. Switch from Claude to Codex and the notes come along.
anthropics/financial-services — Claude drafts pitches and research. A person still signs.
coder/coder — Dev boxes on your own servers. The agent loop stays there; API keys stay out.

Yes, Jev Is Insane, But There’s A Catch — Two Minute Papers. Jev returns probabilities, not paragraphs. Cheap, with a hole in it.
GPT-6 SOL AND LUNA ARE OUT!!! — Matthew Berman. A first look at the half-price pair.
Did Elon catch up? (Grok 4.7 is here) — Matthew Berman. Whether cheap Grok 4.7 closed on Claude and GPT-6.

Jev only decides — $0.042 per million input tokens. Output is free. Yes, a choice, or a score. No reason attached.
llm-typesafe — Simon Willison’s plugin so the LLM CLI can ask Jev. “Refund this” comes back 0.99.
Nobody is reading the code — Specs, tickets, and PRs are all Claude. Management says shipping isn’t the bottleneck. People hit enter for 12 hours.
ChatGPT, via the ad collector — A post claiming other-site behavior now reaches ChatGPT. The Lobsters thread is the real argument.