A model price war, and 50 tons of books fed to a scanner

Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war — Willison's tally of a busy week: Anthropic dropped Opus 5.5 while OpenAI shipped two GPT-6 tiers, and "GPT-6 Sol and Luna are half the price of their GPT-5.6 equivalents".

Halving inference cost changes what's economical to put in a loop — bigger eval suites, more retries, agents you'd previously have budgeted out of existence. Worth re-pricing your workloads before you re-architect them. — Simon Willison

Introducing GPT-6 Sol and Luna — the official launch post for the two tiers: "frontier intelligence to everyday work with different balances of capability and cost." Model selection is now a three-axis problem — capability, latency, price. Read the API limits and run your own evals before migrating anything that pages someone at 3am. — openai.com

CASE STUDY

Perplexity trusts GPT-6 Astra with end-to-end systems — Perplexity has Astra writing communications, changing software, and monitoring production systems, and "checks in much less frequently than with earlier models."

Evidence that longer-horizon agent loops survive contact with production — assuming you've solved permissioning and blast radius, which the post does not solve for you. — openai.com

Ringg's AI agents resolve up to 65% of customer calls with OpenAI — multilingual agents across voice, chat, WhatsApp and web on GPT-5.6, claiming 65% call resolution and 90% lower cost vs GPT-4.1. Vendor-supplied numbers, but rare published baselines for voice-agent deflection — useful as a target to beat, not a promise. — openai.com

Japanese used bookstores see 5x sales surge as books are bought by the ton — used bookstores report a 5x sales surge from bulk buyers, including one 50-ton shipment to the US for scanning and destruction.

Physical scan-and-shred as a training data pipeline raises the obvious provenance and copyright questions, and suggests the cheap web corpus is no longer enough. — speckx

Humanity's Last Invention — Richard Socher of Recursive — the You.com CEO and NLP veteran has spun out a startup aimed at recursive self-improvement, "already worth $5B". The valuation is concrete; the technical approach, from this interview, is not. — latent.space

DEBATE

Roundtables: Could AI really kill us all? — a panel on extinction risk, prompted by the fact that "employees at the world's leading AI labs are saying there's a real possibility that advanced AI could destroy humanity." Mostly familiar ground, but useful for separating lab rhetoric from the failure modes you can actually engineer against. — MIT Technology Review

OpenSpec – A lightweight and configurable AI spec framework — a framework for writing down what your AI system is supposed to do, in a structured form. No benchmarks or technical detail on the landing page, so treat it as a look-and-see if you're currently specifying agent behaviour in a Notion doc. — etoxin

AI-generated posters don't have to be horrible — a design critique arguing the failure mode of AI event posters is taste, not the model. Short on worked examples, but a fair reminder that the prompt is the easy part. — ereiamjh