Two frontier models in one week

I've been using Anthropic's Fable 5.1 and OpenAI's GPT-6 Astra. Both feel more capable, but I'm using up my subscription allowances faster too. This week, I explain the different ways we pay for AI and how I'd compare whether the extra spending is worthwhile.

I also look at Google's attempt to keep AI test questions hidden from its developers, and how an assistant can use a company account without holding that account's digital access key. Then I leave AI behind for a mouse study on aging eyes and a documentary-style audiobook I enjoyed, Leon Neyfakh's Fiasco: The Battle for Boston.

Read the full issue: https://quanta-bits-newsletter.beehiiv.com/p/quanta-bits-september-5-2026
Subscribe to Quanta Bits: https://quanta-bits-newsletter.beehiiv.com/subscribe

Transcript

Copy Transcript Copied

SPEAKER_00 0:01

Hey, I'm Reza Markovadi. Welcome back to QuantoBits. Last weekend we took Coda or Doc down to Narraganset on the Rhode Island coast. A little sun, the Point Judith Lighthouse, and a good rock for him to sit on. I love being able to do these trips from Boston. Two new AI models arrived this week and I've been using both. I want to spend a little bit time on what using them actually involves. Antropic, the company behind the Cloud Assistant, released Fable 5.1 on September 1st. Two days later, OpenAI, the company behind ChatGPT, released GPT-6 Astra. These are AI models, the underlying software that interprets your request and works out a response. They're described as frontier models because they're among the most capable available. Both companies say the new models can handle harder work with less help. That includes computer tasks that take several steps rather than answering a single question. I've been using both for software development. In my limited experience, they are noticeably better than their predecessors, and Astro has been particularly good at carrying out tasks on my computer. I'm also using my subscription allowances faster, especially with Fable. I'll explain that in a minute because a lower advertised cost can be misleading if you're paying a different way. One example in Anthropics announcement came from Millennium, an investment firm. It says Fable traced the rare software crash to a defect that its engineers and earlier AI models hadn't explained. Software that fails only occasionally can be difficult to diagnose. You have to work out what was different when it failed, and whether changing something actually prevents the failure. I can see why a company would pay more for help with that kind of a problem. Finding a plausible explanation would be a start. I would still want engineers to reproduce the problem and test the repair. A millennium's account is a customer testimonial chosen for a product launch. I haven't found independent confirmation of the headline performance results in these announcements by Anthropic and OpenAI. There is also a change in how AI performance gets tested. A benchmark is a standardization of tasks used to compare AI models. The scores can look better if a model has seen the questions before. Google DeepMind, Google's AI research division, is piloting a different arrangement with outside evaluators. They test one of Google's Gemini AI models inside a protected computer setup. The evaluators can submit questions without Google developers seeing them. They get access to run tests without receiving a copy of Google's model to takeaway. The model itself still has to receive the questions to answer them. The aim is to reduce the opportunity to prepare a model for those particular test questions. It's a pilot and it doesn't guarantee the questions are good measures of your work. I would use those results to choose what to try, then check the model on a job I understand. Now above what you pay, you can pay by usage or buy a subscription. If software connects directly to an AI model to do a job, that connection is called an API. Say a company builds a program that sends documents to the model for processing. The company pays for the material the model reads and produces. That material is counted in small pieces called tokens. A longer document or answer generally means more tokens to pay for. Anthropic says its new Fable model can be cheaper for some of that work because it has reduced the charge for reusing material the model has already processed. Instructions repeated across many requests are one example. The savings depends on the work being done. A subscription is different. You pay a recurring fee for Cloud Pro or Max and the plan includes a limited amount of use. It doesn't include unlimited use of every model. On Max, Fable can use up to half your weekly allowance. Once you reach that Fable limit, you can wait for your allowance to reset or pay for extra credits to keep using Fable. Those credits cost money on top of your subscription. On Pro, Fable requires credits from the start. Business plans have their own rules, which I've linked in the newsletter. That's why I separate the advertised savings from what I'm experiencing. A lower charge for some kind of usage doesn't mean your subscription includes more work. Asking a model to think longer or giving it more material to process also affects how far your allowance goes. One request can also launch other AI assistants to help, each adding usage. Where the software supports it, I will set a spending limit for the whole job, including those helpers. Telling an AI assistant to stay within the budget doesn't guarantee it will stop spending there. For my own work, spending more on AI could be worthwhile if it finishes jobs I previously had to do myself. In the newsletter, I mentioned Sarah Freyer, OpenAI's chief financial officer. In an account of how her finance team uses AI, she includes employee time and rework when judging what an AI job costs. That's the comparison I would want to make too. Taking monthly spending report, I might ask two models to explain why expenses increased using the same records. I would check whether each explanation agrees with the records, whether it missed a large expense, and how much time I spent correcting the report. A confident paragraph blaming the wrong expense isn't a finished job, however quickly the right. I would rather pay a little bit more for work I can use than pay less and spend the afternoon repairing it. Giving both models the same job should also help me decide which one fits that work. I don't need to move everything to one model because it won't attest. Cost is one concern. Giving an AI assistant access to company accounts is another. Once the assistant can send messages or change records, a mistake can affect other people before you have checked the work. WorkOS, a company that makes login and access software, describes one approach through a service card relay. For example, suppose an AI assistant needs to post an update in Slack, the Workplace Messaging App. Slack requires a digital key to authorize that request. Relay keeps the Slack key, adds it to the assistant's request, and passes the request to Slack. The assistant can get the update posted without holding the Slack key itself. That reduces the chance of the assistant exposing that particular key in its output or working files. But it could still send the wrong message. And the assistant still needs a separate key to use relay, which also needs protection. So I would still limit which channels it can post to and which actions need my approval. Protecting the key and limiting what someone can do with it are both necessary. Relay doesn't establish all those permission rules for you. Let me leave AI for a moment and go to this week's meanwhile story. Researchers at the University of California Irvine reported an experiment last year that improved vision in older mice. The work concerns the retina, the light-sensitive tissue at the back of the eye. As the mice aged, they produced less of certain fats that helped maintain that tissue. After the researchers injected one of those fats into the mice's eyes, their vision improved and some signs of aging in the eye tissue were reduced. They also tested DHA, a better known omega-3 fat, and it didn't produce the same improvement in this experiment. So this was a specific treatment tested in mice, not evidence that taking a supplement improves human vision. There is still a lot to establish before we know whether something like this could help people. And for after hours, I finished Fiasco The Battle for Boston by journalist Leon Neifach. The audiobook edition came out in February 2025. It's about 5.5 hours and it feels much more like a documentary podcast than someone reading a book aloud. It follows the effort to end segregation in Boston's public schools in the 1970s. A court ordered busing, which meant transporting children between neighborhoods to integrate schools, and the series explores the resistance that followed. I've lived here a long time and I hadn't understood this part of the city's history. The interviews and old recordings are a big part of why I enjoyed it. You're hearing people involved in the events, alongside the explanation of what happened. It also covers singer James Brown's televised Boston concert after Martin Luther King Jr.'s assassination and the efforts to keep the city calm that night. I liked how quickly it moved, although it lost some steam towards the end. Still 5 stars from me, I would happily listen to more history told this way. If you enjoy documentary podcasts, this is a good one to consider, especially if you know Boston, but haven't spent much time with its history. That's it for this week. The full issue with all the links and sources is in your inbox at sh newsletter or at quanta dash bits newsletter.behive.com. I'm Reza Marikabadi, thanks for listening. See you next week.