OpenAI released two new models on November 17, 2026, and the interesting part is not that they are faster or smarter. It is that the company split its flagship line into two distinct jobs and cut the price of the expensive one. GPT-6.2 Astra is the model for work that has to be right, the long professional tasks where a mistake costs real money. GPT-6.2 Sol is the model for everything else, the daily volume that makes up most of what people actually do with AI. The split matters more than the benchmark charts, because it tells you how the people building these systems now think about the difference between a ten-dollar task and a ten-thousand-dollar task.
Two Models, Two Jobs
Astra is aimed at multistep professional work. OpenAI points to slideshows, spreadsheets, legal research, scientific work in a terminal, and front-end builds that have to look like a human made them. Sol is aimed at the routine, the drafting and summarizing and answering that fills a workday. Both models use fewer output tokens than their predecessors for the same result, which is the company's way of saying the models got more efficient rather than merely more capable.
That framing is a change from a few years ago, when every release was a single flagship and the marketing was about one number on one chart. Now the pitch is about fit. The model you reach for depends on the cost of being wrong, not on which one wins a leaderboard.
What Astra Buys You
The benchmark that matters for professional work is long-horizon tasks, the kind that take a model dozens of steps and a lot of tokens to finish. On Agents' Last Exam, which tests exactly that, Astra scored 5.2 points above GPT-6 Astra while costing less than a third as much per task. OpenAI also reports a new high on Terminal-Bench-Science and on the Vals Science Index, which asks models to interpret real biological data and reason about how a simulated system works.
The customer quotes tell a more practical story. A legal technology company reports that Astra used 62 percent fewer tokens than an earlier model on complex legal work and scored higher on eight of ten research tasks across tax, mergers, and litigation. A documentation platform says the model completed more than three times as many legal benchmark tasks while using 34.5 percent fewer output tokens. A workspace software company says Astra edged out its production model in blind testing while using about 40 percent fewer tokens and 30 percent fewer review rounds.
Those numbers describe a model that finishes the job with less back and forth. For a firm that bills by the hour or a team that reviews every AI output, fewer review rounds is the whole game.
What It Costs
Pricing is where the announcement gets concrete. Astra's API prices are 20 percent lower than GPT-6 Astra's, at eight dollars per million input tokens, forty cents per million cached input tokens, and forty dollars per million output tokens. Sol keeps the prices of the model it replaces.
Run that against a realistic workload. A support desk that sends a thousand tickets a month through a model, with a few hundred tokens of context each, pays a number small enough to fit in a marketing budget. A research team running the same volume through Astra pays more, and should, because the task justifies it. The price gap between the two models is now wide enough that routing work between them is a genuine cost decision rather than a rounding error.
Sol Is the Quiet Story
Sol will handle more total work than Astra, and OpenAI seems to know it. The company says Sol comes within half a point of Astra on FrontierCode 2.0 at about one-fifth the cost, and that it outscores an earlier flagship on LifeVal, an evaluation of everyday tasks, at about a seventh of the cost.
That is the shape of a product meant to be used constantly. Most business AI spending is not on the hard problems. It is on drafting, rewriting, summarizing, classifying, and explaining. A model that is nearly as good as the flagship at a fifth of the price will win that volume on economics alone.
The Safety Part Is the Real News
The part of the announcement that will matter in five years is a single line. Astra is the first externally released model to meet OpenAI's High threshold for AI self-improvement, meaning the model is good enough at AI research to potentially accelerate the next generation of AI research.
OpenAI's response is telling. It shipped tighter safeguards for high-risk cybersecurity help, and it routes approved defenders through a program called Daybreak where refusals are reduced. It also says Astra and Sol meet the Critical threshold for cybersecurity capability under its Preparedness Framework.
This is the trade that the industry is now negotiating in public. Capability and control move together, and the releasing company has to build the control layer at the same time as the model. Any business evaluating these tools should ask what safeguards travel with them, not just what the model can do.
What This Means for a Small Business
The practical takeaway is that the two-tier structure is now real and affordable. You no longer have to choose between a cheap model that fails on hard work and an expensive model you use for everything. You can route the routine volume to an efficient model and hand the hard cases to the frontier one, and the price difference makes that worth doing deliberately.
Start by listing the tasks you actually run this month and sorting them by the cost of a wrong answer. Put the low-stakes volume on the cheaper model and measure. Put the high-stakes work on the stronger model and measure that separately. Most teams find the split saves money without a visible drop in quality, because the expensive model was doing work the cheap one could have handled all along.
The Bottom Line
GPT-6.2 Astra and Sol are less a single upgrade than a reorganization of the model lineup around what work actually costs. Astra is for the tasks where getting it right is worth real money. Sol is for everything else, and it is close enough to the flagship to take the volume. The price cut makes the split worth doing on purpose.
If you want help deciding which of your own tasks belong on which model, talk to us at Crowe Endeavors and we will sort your workload the same way.
Sources
"GPT-6.2 Astra and Sol: OpenAI's New Models, Read and Highlighted" (YouTube, November 2026) — https://www.youtube.com/watch?v=_DxOKTrhjWI
"Introducing GPT-6.2 Astra and Sol" (OpenAI, November 17, 2026) — https://openai.com/index/introducing-gpt-6-2-astra-sol/