A field guide

Cheap to attempt, expensive to trust.

OpenAI shipped a model that can operate software for forty minutes at a stretch and still fails most of the way through six out of ten professional automations. Seven businesses that work anyway, because in each one checking the answer costs less than producing it. Each also gets the specific reason it might not.

by Dave Brown

GPT-6 Astra shipped on September 3rd and the argument started before the docs finished loading. OpenAI’s president said he personally thinks we’re there. The company’s own launch materials pointedly did not say that, settling for “most intelligent and aligned model” and “world’s best computer use model.” Everyone else has spent the week relitigating what AGI means.

That argument is unfalsifiable and it is not the one that pays. There is a narrower question sitting underneath it with an actual answer: what work is now worth doing that wasn’t worth doing last month? Answering it requires no position on machine consciousness, only two numbers: what a unit of work costs, and what it costs to check it.

The honest read of the launch is that this is not a general leap. Independent evaluation puts Astra roughly level with its own predecessor on broad intelligence and behind Anthropic’s Fable 5.1. What moved, and moved hard, is the ability to sit in front of bad software and keep going: longer horizons, fewer wasted tokens, a hallucination rate that fell by roughly half, and task times that dropped by nearly half as well.

That is a strange, specific capability, and it favors a strange, specific class of business. What follows is seven of them, with the same discipline as everything else I write: every idea gets a matching paragraph on how it dies, and most of them have a funded competitor already.

The seductive version of this thesis is “intelligence too cheap to meter, spin up a thousand interns, print money.” Astra costs two and a half times what the model it replaced cost, per token, and the independent measurement of cost per completed task came back higher. Anyone budgeting off the long-run price curve should check how far away the long run is.

Every row below carries a confidence level, since that is the column that changes what you do about it.

Two rows deserve to come out of the table, because people are getting both of them wrong in opposite directions.

• Read the harness gap as a lesson rather than a scandal. A 37-point swing from changing how reasoning state and context are managed tells you that a large fraction of usable capability sits in the scaffolding around the model rather than inside it. The same trick previously moved the prior model from 13.3% to 38.3%. For anyone building here that is good news, because scaffolding is the part you control, and it is why “wait for the next model” is a weaker strategy than it sounds.
• Nobody has a reliable cost model yet. One widely circulated estimate had per-task cost falling roughly 57% on computer-use work because Astra burns fewer tokens getting there. An independent evaluator measured cost per task across a broader mix and found it about 75% higher, because the price increase outweighed the efficiency gain. Both are probably right about their own task mix. That they are both circulating means your unit economics are something you will have to measure for yourself.

A 41% success rate sounds like a reason not to build. It isn’t, provided you never sell the agent’s raw output. You sell verified output, and the economics of the whole business reduce to one ratio: what it costs to attempt the work, divided by what it costs to know whether the attempt succeeded.

Checking is nearly free in some workflows. A test suite passes. A checksum matches. An invoice either duplicates an earlier one or it does not. A form submits and returns a confirmation number. Where that holds, a 41% success rate becomes a scheduling problem: run it three times, keep whatever verifies, and delivered accuracy lands near 80% at three times the compute cost, which is still a rounding error against the human alternative. Where checking instead requires a licensed professional to read the output carefully, you have not built a business. You have built a way to generate work for a licensed professional.

This is also why the long-horizon research matters more than the launch benchmarks. Errors in agent workflows are positively correlated across steps rather than independent, so failure probability grows faster than the naive compounding math predicts. Current frontier agents are close to perfect on tasks a skilled human would finish in under four minutes and succeed less than a tenth of the time on tasks that would take a human more than four hours. Success rates of 40–50% on short tasks fall below 10% when those same tasks are embedded in a long interaction history.

The design implication is unambiguous and almost nobody’s pitch deck reflects it: decompose until every unit fits inside the reliable horizon, then verify at every seam. Sixty independently checkable four-minute units will finish work that one four-hour autonomous run will not.

Everything in Parts three through five passes that filter. Whether each one passes the harder tests in Part six is a different question, and for at least two of them the answer is probably no.

Most organizations leave money on the floor in amounts too small to justify a human going to get it. Nobody pays an analyst $300 to chase a $137 duplicate charge. The entire market here is that threshold, the point below which recovery costs more than it returns, and it just moved.

An enormous amount of white-collar labor is human beings acting as adapters between systems that refuse to talk to each other. That work survived every previous automation wave for one reason: integration required either an API that didn’t exist or brittle scripting that broke on the next UI release. A model that can look at a screen and adapt changes the failure mode from “breaks” to “costs a retry.”

Pick one awful administrative workflow and sell the finished work. We handle everything after a home inspector finishes an inspection. We run the back office for independent adjusters. We process supplier invoices for electrical contractors. We do the post-close file for small title agencies. The customer never sees a product and never learns a prompt. They email operations@yourcompany.com and completed work comes back.

You are selling exactly what an outsourcing firm sells, which is delivered outcomes, except that the labor pool is tokens. That arbitrage is measurable rather than theoretical: traditional outsourcers run at roughly 25–30% gross margins, and AI-native operators claim blended margins around 64% at a 70/30 split of machine to human work. Treat that specific figure with suspicion, since it comes from people selling the model, but the direction gets corroborated by the incumbents’ own behavior. One of the large consultancies paid $3.3B for an outsourcing firm to buy its way in, and early adopters report operating-margin expansion of 750 to 1,000 basis points.

The structural advantage is that a services wrapper absorbs the 41%. When the agent fails, a human on your side fixes it and the customer never knows, so the failure shows up as an internal cost line rather than a broken product experience. No other shape on this list has that property, which is why this is the one idea here that does not need the next model to be better.

Five objections come up whenever this is presented, with the most honest response I have to each. One of them I have no answer to at all.

Scored against all of that, the list separates cleanly. The middle column is the one that decides whether you have a business; the right column is what you’re betting on.

Idea 07 is the outlier and it’s worth saying plainly. It needs no capability improvement, no price decline, no access tier you can’t get, and no thesis about AGI, only a boring workflow and a willingness to run an unglamorous company. It is the one item here that survives if the argument in Part two turns out to be wrong. Idea 01 has the largest ceiling, available only to whoever builds the collections half alongside the search half.

The indicators first. These would move this from an interesting capability to a settled market, roughly in the order they would have to happen.

The test second. For anyone seriously considering one of these, this sequence produces better information than any amount of market sizing, and it fits in a month of evenings.

A real capability shipped, in a narrower band than the coverage suggests. Long-horizon computer use got substantially better and substantially faster, and hallucination roughly halved. General reasoning did not move, one economically-relevant benchmark went backwards, the flagship number depended on a harness, and tokens got more expensive.

The businesses that work in this window are the unglamorous ones: somebody already pays for the manual version, the output can be checked by a machine for pennies, the work decomposes into short units, and being wrong is survivable. Most of the seven above fail at least one of those, and which ones fail depends on details of your own workflow that no market map contains. The thirty days in Part seven cost less than one bad quarter.

The most interesting thing about this whole shift is that almost nobody arguing about it has done the division. Cost per accepted unit against the loaded cost of the human currently doing it: that single number settles more of this than a week of thinkpieces, and it comes out different for every workflow.

I’m doing a small number of informal reads. Send me the workflow you’re thinking about, including what it is, who does it now, how long it takes them, and how you’d check the output, and I’ll tell you which of the seven shapes above it fits and whether the verification step is cheap enough for it to be a business. No pressure; everything here is yours to use either way.

A note on sourcing: benchmark figures come from OpenAI’s launch materials and from independent evaluators (Artificial Analysis, ARC Prize, and the Agents’ Last Exam leaderboard), and where those disagree I’ve given both numbers rather than the flattering one. The reliability findings on long-horizon agents come from recent academic work on error correlation and task-length scaling. Market and funding figures come from press coverage and, in the case of the outsourcing margin numbers in Part five, from vendors with an obvious interest in them; treat those as directional. Benchmark scores in a week-old model release move fast, and access tiers were still shifting while I wrote this, so verify anything you plan to build on. The seven ideas, the scoring, and the thirty-day test are argument rather than research. I have no inside information about any company named here, no revenue figures from anyone in a position to know them, and no way to check the Part five margin claims against an audited filing, because none has been published.