Short answer
GPT-6 Astra launched in preview on September 3, 2026 and opened to paid ChatGPT plans on September 4, with restrictions in areas such as cybersecurity. For most small and mid-size companies the model is not the bottleneck: MIT's research found 95% of AI pilots produced no measurable profit impact, and blamed integration, not model quality. What changes with each launch is the cost curve and the length of work an agent can do unattended. The right response is a simple model policy, a 20-example test set, and quarterly rather than weekly upgrades.
It has been a loud five weeks. Google released Gemini 3.7 Flash in August at half the per-token cost of the previous Flash model, and both ChatGPT and the Gemini app passed a billion monthly users. Anthropic shipped Claude Fable 5.1 on September 1. OpenAI released GPT-6 Astra to a limited preview on September 3 and opened it to paid plans the next day, calling it a generational leap and shipping it with restrictions in areas such as cybersecurity after its own July incident.
If you run a company, the temptation is to feel behind. Resist it. Here is what actually changes for a business with twenty people and a stack of ordinary workflows, and what does not.
What does not change: the model was never the bottleneck
A year ago MIT's NANDA initiative published a study that has aged well. Across 52 executive interviews, a survey of 153 leaders and 300 public deployments, it found that 95% of generative AI pilots produced no measurable impact on profit and loss. The authors did not blame the models. They blamed integration: generic tools that do not learn a company's workflow, and pilots that never get wired into the systems where work happens.
Every model since has been better than the ones in that study. The failure rate has not obviously moved, because the thing that fails is not the model. It is the absence of a clear process, clean inputs, an approval step and someone who owns the result. A more intelligent model reading an undocumented process produces a more confident wrong answer.
What does change: the cost curve
Each launch cycle moves prices down for the tier below the headline. Gemini 3.7 Flash halved the cost of the model most Google-based automations actually run on. Anthropic left Fable 5.1's headline prices alone but cut cache reads by 75%, which is the line item that dominates long agent runs. The practical effect is that workflows which were marginal a year ago, such as reading every inbound email and classifying it, or extracting line items from every supplier invoice, are now clearly worth doing even at small volumes.
So the question to ask on launch week is not "is the new model smarter?" It is "which workflow that we rejected on cost last year is now cheap enough to revisit?"
What also changes: how long an agent can run alone
The labs are competing on unattended work. Anthropic cites a customer whose Fable 5.1 job ran for 38 hours; OpenAI frames Astra around multi-step computer use. For a business this cuts both ways. Longer runs mean tasks like a monthly reconciliation or a research memo can be handed off overnight. They also mean the checkpoints matter more, because an agent that can work for hours can also go wrong for hours. Our rule has not changed: the agent may read, draft and prepare indefinitely; anything customer-facing or financial waits for a person.
A one-page model policy
This is the document we set up with clients, and it fits on a page.
| Task type | Model tier | Fallback | Approval |
|---|---|---|---|
| Classify, route, extract fields | Cheapest model that passes the test set | Second provider, same tier | None, unless confidence is low |
| Draft replies, quotes, summaries | Mid-tier | Second provider, same tier | Person approves before sending |
| Long research, multi-document analysis | Frontier model | Mid-tier with a warning | Person reviews the output |
Add four lines below the table: where prompts and logs live, who may change a model, the rule that every change runs against the same twenty real examples first, and a quarterly date to review. That last line is the one that saves the most money. Upgrading every time a lab holds a launch event is how teams spend a year re-testing and never shipping.
The twenty-example test
If you do nothing else this month, collect twenty real inputs for your most important workflow. Include the ugly ones: the invoice photographed at an angle, the inquiry written in two languages, the email with the actual question buried in paragraph four. Score how your current setup handles them. Then, and only then, try a new model. In our experience the result is usually that the mid-tier model was fine and the prompt needed work. Occasionally it is that the new model unlocks something. Either way you will know, which is more than the launch post can tell you.
Our own setup, for what it is worth
The assistant on this website runs on a fall-through chain: a primary model, an automatic fallback from a different provider, and a plain knowledge-base tier if both are unavailable. We did that not because we predicted this month's launches but because the model layer changes every quarter and the rest of the system should not have to.
Frequently asked questions
Should we switch our automations to GPT-6 Astra?
Not on launch week. Run your own twenty real examples through it and compare against what you use now. If accuracy on your inputs does not move, the upgrade is cost without benefit. If it does, check the pricing and the new restrictions before you switch.
Which model should a small business use?
Usually a tier below the headline model. Inbox triage, document extraction and drafting run well on mid-tier models from any of the three major labs. Reserve frontier models for the few steps that need judgment across long or messy inputs, and put a fallback behind whatever you choose.
Why do so many AI pilots fail if the models keep getting better?
Because the failures are rarely about the model. MIT's NANDA study, based on interviews, surveys and 300 public deployments, found the gap was in integration: generic tools do not learn a company's workflow, and pilots are rarely wired into the systems where the work actually happens.
What does a model policy look like for a company our size?
One page. Which tier of model each kind of task uses, which provider is the fallback, where prompts and logs live, who can change the model, and a rule that any change is tested on the same twenty examples first.
