Should I switch AI models now that prices have dropped?

Look at what you run, not at the headline. This week OpenAI's smaller GPT-6 models launched at half their predecessors' prices and Anthropic cut Opus 5.5's list price by 20%. For routine, repeatable work, the cheapest model that does the job well is now very cheap. Keep the expensive models for the tasks where you have seen them make a difference.

Can I let an AI make decisions on a long project?

Not the ones that set the direction. A new benchmark measured how often AI agents choose the better path at the forks of a long task, and the best model got it right 59.7% of the time, with extra thinking time barely helping. Let AI do the work between decisions, and make the direction-setting calls yourself.

This Week in AI: Intelligence Got Cheaper. Judgment Didn't.

The price of AI fell again this week. The quality of its judgment did not rise with it. Anthropic and OpenAI both released cheaper models within an hour of each other, while new research measured how often the best AI picks the right path at the decisions that shape a long task: 59.7% of the time. For a one-person business, those two facts together say where to spend and where to stay involved. Buy the cheap execution. Keep the decisions.

A price war, in plain numbers

On 22 September Anthropic released Claude Opus 5.5 and OpenAI released GPT-6 Sol and GPT-6 Luna. Simon Willison’s price table has the numbers: Opus 5.5 is 20% cheaper than the model it replaces, and Sol and Luna are each half the price of the previous generation. Luna, OpenAI’s smallest, costs ten cents per million tokens of input. Ars Technica summed it up as “a little more for a lot less money.”

What that means for you: the routine work — drafting, summarising, sorting, first passes — just got cheaper to hand off on a model you do not need to think about. The price cut does not tell you which tasks deserve the expensive model. That is still a decision you make per task, from what you have seen each model do on your own work.

Where AI still gets it wrong: the forks

The Tasteful Agent (115 upvotes on Hugging Face) tested something the usual benchmarks skip. On long tasks, AI makes choices along the way — which approach to try, which draft to build on. The researchers collected these forks automatically from real AI work and checked how often the model chose the path that turned out better. The best model: 59.7%. When the evidence that settles the choice only shows up later, it gets worse. Giving the model more time to think barely helps.

This is The Reliance Calibration Dial in one number. Trust should be set per kind of output, from what you actually observe — and here is an observation: AI is reliable at doing the steps and much less reliable at choosing the direction. Let it write the three versions. You pick which one the business builds on.

Your AI’s memory is moving to the cloud

Google DeepMind announced secure server-side memory for its Private AI Compute platform on 23 September: personal AI that remembers your context across devices, stored encrypted, with the keys kept on your device and processing inside protected hardware. Google backs the claims with an updated whitepaper and an independent security audit; the announcement reads as capability being rolled out, not something every user has yet.

Memory is what turns a stateless tool into something that learns your business, so this direction matters. It also raises the question The Data Boundary exists to answer before you need it: what are you willing to put into a memory you do not host? Client names, pricing, contracts and health details each deserve their own answer. Decide the categories now, while the memory is still empty.

A self-improving system can learn the wrong lesson

Google Research’s RRSI paper (181 upvotes) studied AI setups that rewrite their own instructions to get better over time. Left alone, they improve mostly on the tasks they practised and much less on anything new: up to 14.1 points better on familiar tasks, up to 4.7 on unfamiliar ones, even after the researchers added guardrails to curb the memorising.

This is the honest version of “AI that learns your business.” A system that improves from your corrections does compound — but check it on work it has not seen before, not only on the tasks you corrected it on. That is the point of the 90-day test: improvement you cannot see on new work is familiarity, not capability.

“The lack of anything”

Simon Willison quoted a TikTok creator on why AI-written scripts are easy to spot. Past the familiar tics, the problem is “the lack of anything” — polished sentences with no point of view behind them. That is The Distinctiveness Drift described by an audience member: AI smooths your work toward the category average, and your audience notices before you do. Use AI for structure and speed. The opinion has to be yours.

The throughline

Cheaper models make execution a commodity. What they do not make cheaper is knowing which path to take, what to keep private, and whether your system is actually getting better on new work. Those are the parts of the business that compound when you own them — and decay when you rent them. That is the argument behind marketing-grade decays, engineering-grade compounds.

More on building AI into a one-person business you actually own: curiochat.ai/solopreneur/.