Issue #1

Simon Willison argues that pay-per-use services and APIs should come with hard spending caps switched on by default, cutting usage off and returning errors once a monthly limit is reached, because coding and personal agents make it easy to spin up things that quietly run up bills. Warning emails aren’t enough when a runaway service can cost thousands overnight, and anyone who wants unlimited spending should have to opt in. He notes that AWS has begun rolling out project spend limits and that Google Cloud launched spend caps in July.

Cloudflare opened a closed beta of a gateway that lets site owners charge AI agents per use for websites, APIs, MCP tools and datasets. It relies on the HTTP 402 “Payment Required” status: the agent receives payment terms with its request, signs an authorization, and gets the resource once payment settles, with no checkout page or separate payment API. Cloudflare argues that subscriptions don’t fit agents that consume per request, and currently settles payments in stablecoins. A separate Pay Per Use program covers content that is crawled once and used many times.

Researchers from Transluce and several other organizations traced AI agents using aggressive tactics to pull public data from government websites, using logs from the URL scanner urlquery.net and the Portuguese web archive Arquivo.pt. On one day, agents sent more than 200,000 requests to a U.S. Department of Education site, including a basic SQL injection attempt; the data sought matches a question in a public web-research benchmark. They also found a failed attack on Library and Archives Canada and probing of federal and state agency sites, but no case where agents obtained non-public information.

201 Anthropic employees each gave a Claude agent a book to trade and a five-minute chat about their tastes, and the agents negotiated swaps with each other on shared trading floors. The agents bargained competently; the market fell short mainly because of what they didn’t know about their person, since the rankings built from each chat matched the person’s own on 61% of pairs. Re-running the markets dozens of times showed that the underlying model affected outcomes more than telling agents to be “ruthless” or “prosocial”.

Cursor cut its users’ token costs by 7% without lowering agent quality by reworking how each request is assembled. As models improved, it removed about two-thirds of its system prompt, which had been full of warnings older models needed. It also stopped sending every tool definition on every turn: only frequent tools like read, search, edit and shell stay loaded, and the rest load on demand, which cut tool-description tokens by 60%. Each change was validated with A/B tests on real traffic, since evals alone don’t reflect typical requests.

Google is adding persistent, cross-device memory to Private AI Compute, its hardware-isolated cloud environment, which until now discarded all context when a task ended. Each user’s memory sits in encrypted storage whose keys stay on their own devices; data is decrypted only inside secure enclaves while a request is handled, and devices can check a tamper-proof public record of the server software before sending anything. Google also published an updated whitepaper and an independent audit.

Anthropic released Claude Opus 5.5 and, within the hour, OpenAI released GPT-6 Sol and GPT-6 Luna. Simon Willison’s first read focuses on price: both GPT-6 models cost half as much as their GPT-5.6 equivalents, with Luna at $0.10 per million input tokens and $0.50 per million output, among the cheapest models OpenAI has shipped. Opus 5.5 cut its price by 20% to $4 and $20, and cached-input reads by 60%. The top tier (GPT-6 Astra and Claude Fable 5.1) stays at $10 and $50.

Spotify examines whether AI-assisted development hurt quality as its merged changes more than doubled in a year. Its incident reviews found no case where AI-written code was a material direct cause; the real pressure came from the volume of change outpacing review, testing, rollout and rollback, along with scarcer compute. Code-quality work grew as a share of merged PRs and rework rate stayed flat, though PR size and complexity are creeping up.

OpenAI argues that instructions written to steer older coding agents now often get in the way. Skill descriptions should be short and say exactly when each skill applies: with too many, Codex truncates them and the model picks the wrong one. Skills with several workflows should open with a short index that points to details instead of loading everything, and rigid step-by-step recipes can now hurt results. AGENTS.md files should point to documents only when they’re relevant, rather than requiring a full read before every edit.

Anthropic’s Economics team published an interactive model of how AI could affect US jobs, growth and wages by 2030, based on a technical report by Korinek and colleagues. It treats every job as a bundle of tasks that AI may leave alone, augment or automate, while also creating new ones. In scenarios up to about double the normal growth rate, unemployment stays within its historical range. Only with growth faster than anything in economic history do knowledge workers’ wages and job prospects suffer, and the challenge becomes sharing the gains. Readers can enter their own predictions and see the economy they imply.

Mistral describes migrating a 40,000-line Fortran 77 reservoir simulator, with no tests and no central documentation, to C++ for a European energy operator. Before migrating anything, the team built a parity harness: the Fortran code dumps state checkpoints and the C++ tests load them, so every migrated module must match numerically, and Skill.md files tell the agents how to use it. Agents documented the codebase before the migration began, and people reviewed the work at set checkpoints.