Quanta Bits - September 5, 2026

Hi everyone,

Last weekend we drove down to Narragansett with Koda. Sun, granite, and the Point Judith lighthouse, and he found the best rock on the shore. Love day trips to the shores around here!

This is the 30th edition of Quanta Bits, more than six months in the making. I know from experience that when you work full time on specific projects, it's hard to pick your head up, look around, and research other developing and perhaps interesting areas. My intent behind this newsletter has been, and remains, to give my busy professional friends, and anyone without much time to follow AI, a quick shot of the developments I find interesting and relevant to business operations.

I still believe in that mission and am committed to it. I keep experimenting with formats and segments, and this will keep evolving like most organic, living experiments. To that end, the essay section is becoming less deep thoughts and more expanded news patterns and discussions, with occasional deeper delves (to use what we now think of as an AI word, such madness we live in!) into specific threads.

This week is a good example. Two frontier models, the most capable class of AI available, arrived within days of each other, so the feature is a discussion of what was claimed and what it could mean for the rest of us.

Cheers.

Reza

On August 30, The Economist examined why software engineers use AI more heavily than other workers. One advantage is that code can be tested automatically. A team can check whether a change works before accepting it. Other business work can be harder to assess, particularly when a plausible answer still needs an experienced person to check it.

That connects with an earlier discussion about measuring AI returns. In an August 10 account of rebuilding OpenAI's finance function, its CFO Sarah Friar asks whether AI finished useful work, what it cost including employee time and rework, whether the result was usable, and whether the team moved faster or made better decisions. An August 3 New York Times report described another approach at cybersecurity company Elisity: employees write a job description for each proposed AI agent, and the company compares AI spending with the full cost of employing a person. These are different attempts to connect AI spending to a defined job. Neither the number of requests nor a busy agent tells you whether that job was done well.

Andrew Bailey, chair of the Financial Stability Board, which coordinates financial-risk policy across major economies, warned G20 finance officials that disruptions involving advanced AI could spread across national borders. His letter, published August 31, calls for coordinated preparation and safe deployment. Reuters reported, via Insurance Journal, that cyber insurers are also reviewing coverage as AI agents take actions on their own. One insurance question is whether losses caused by an agent are covered when there is no conventional hacker behind them.

A separate exposure is the access already granted to AI tools. VentureBeat reported stolen Claude sessions and unauthorized usage; access to connected Gmail was a risk, not a confirmed outcome. A session is what keeps a user signed in, so stealing it can let someone act through that account without knowing its password.

WorkOS, a company that sells login software, described one way to reduce credential exposure in an August 10 technical post. Its Relay service holds the login credential for an outside service and attaches it when forwarding the agent's requests. The agent still needs its own key to use Relay. This keeps the outside service's credential out of the agent's working environment; it does not, by itself, restrict the agent to one task. Permissions still have to limit what those requests can do.

Frontier models are the most capable class of AI available at any moment. On September 1, Anthropic released Claude Fable 5.1 and Mythos 5.1: the same model with different safeguards, with Mythos restricted to approved organizations. On September 3, OpenAI released GPT-6 Astra. In between, Google shipped Gemini 3.8 Flash, its fast and inexpensive model. This gives people choosing AI for work several new options at once, with different costs and limits on what they can do.

Anthropic reports Fable 5.1 scoring 55.8%, up from Fable 5's 42.0%, on Terminal-Bench 4.0, a test of multistep computer tasks performed through typed commands. On AutomationBench's business workflow tasks, it reports 31.4%, up from 17.1%.

A more concrete example comes from investment firm Millennium. It says Fable 5.1 traced a rare software crash to a defect that engineers and earlier models had not explained. This is a customer testimonial in Anthropic's launch post, not independent proof.

The pricing is easier to follow if we separate the ways people pay. API use is when software connects directly to the model, for example to process documents automatically. You pay for the amount of material the model reads and produces, measured in small pieces called tokens. Anthropic has kept Fable's basic rates unchanged but made reusing previously processed material cheaper. It says this can lower usage bills, depending on the work.

A subscription, such as Claude Pro or Max, charges a recurring fee that includes a limited amount of use. It is not unlimited access to every model. Paid usage credits cover work outside that allowance and cost extra, on top of the subscription. You can incur those extra charges while using Claude's app; they are not limited to people building software through the API.

For Fable, the plan rules matter: Max users, premium Team users, and premium users on legacy seat-based Enterprise plans can spend up to half their weekly allowance on it before needing extra credits. Pro and standard Team users need credits from the start. So cheaper metered use does not mean a cheaper subscription or a larger included allowance.

OpenAI reports a 57.9% score on Terminal-Bench 4.0. Its president Greg Brockman told NBC that Astra "can really do anything a human can do with a computer." That is a much broader claim than a result on one test.

Before release, OpenAI rated Astra "Critical" for cybersecurity under its own safety framework. It believes the model can find unknown flaws in well-protected systems and develop attacks without step-by-step human guidance. OpenAI deployed it with restrictions intended to block harmful use, while reserving more advanced security access for vetted partners. The rating describes a capability OpenAI found in testing; it does not mean the public product freely offers that capability.

OpenAI's safety report describes a problem with catching unwanted behavior: Astra's written explanations of its thinking are less helpful than those of the previous model, GPT-5.6 Sol. Compared with Sol, monitoring systems were better at spotting unwanted behavior in Astra's actions and answers alone. But when given the full record, including each model's written explanations, they generally performed worse with Astra. In some tests, Astra also avoided detection when explicitly instructed to try. OpenAI can still monitor what Astra does, but reading its explanations is becoming a less reliable way to spot trouble.

I haven't found independent replications of the headline performance results cited here. The published prices, plan rules, and vendors' stated safety classifications can be checked against their documents. Whether either model finishes your work more reliably, or reduces the total cost of doing it, still needs testing on that work.

In my limited use so far, both models have been noticeably better than their predecessors. Astra has been particularly good at carrying out tasks on my computer. But I'm also using up my subscription allowances faster, especially with Fable 5.1. That is worth distinguishing from the lower usage costs Anthropic advertises. How much work you get from your plan also depends on how much thinking you ask the model to do and how much material it processes along the way.

I am running both Fable 5.1 and Astra for my own development work. These are the things I'd check before giving either new model more responsibility:

There is enough here to try both models on work that has been difficult to delegate. Give each model the same job and compare the results. Which did the better job? How much checking and correction did each need, and what did it cost compared with how you do the work today? That would tell me more than which one answered faster, and help me decide which model to use for that kind of work.

Around Astra's release, reports of disruptions to Grok, Codex, and Claude prompted this joke from Tibo. Wait, are they all just running the same model on the same server?! 😊

Source: Tibo (@tibo_maker) on X, September 3, 2026

In a study reported by the University of California, Irvine in September 2025, researchers improved aged mice's vision by injecting a specific fatty acid into their eyes. The work concerns the retina, the light-sensitive tissue at the back of the eye. Activity of a gene called ELOVL2 declines with age, reducing production of fats that help maintain it. Replacing one of those fats improved vision and some molecular signs of aging. DHA, a better-known omega-3 fatty acid, did not produce the same improvement in this experiment. It's still mice and still injections, so a human treatment is a long way off.

Credential proxying: Imagine an AI assistant posting an update in Slack, a workplace messaging app. Instead of giving the assistant Slack's access token, the digital key that authorizes the request, a service such as WorkOS Relay holds it. The assistant sends its request through Relay, which adds the key and forwards the message to Slack. This keeps the Slack key out of the assistant's hands. It does not prevent the assistant from posting the wrong message, so rules about what it can read or send still matter.

Leon Neyfakh | Audiobook released February 4, 2025 | 5h 25m audio | ★★★★★

I finished Fiasco: The Battle for Boston. It's a short audio documentary, more podcast than book, and I'd happily consume more things like it. I've lived in Boston a long time and didn't know this history of school segregation and the 1970s court-ordered busing that tried to end it.

It also covers moments like James Brown's televised Boston concert after Martin Luther King Jr.'s assassination, and the efforts to keep the city calm that night.

The format is why it works. Clips and interviews make it feel like a documentary rather than a narration, and it moves quickly. It loses some steam near the end, which is my only real complaint.

Prefer to listen? Quanta Bits is also available on Apple Podcasts and Spotify.

I collaborate with Spock, my AI agent. He researches extensively: scanning, filtering, and surfacing what is relevant across my business. I read, listen, and watch what resonates, and decide what matters. I provide the direction, and we draft together. The editorial judgment is mine. He would tell you the same. Most logical. 🖖