11/20 2026
338
OpenAI Launches GPT-6.2 Dual Models—Is the More Powerful One Actually Cheaper?
In the early hours of November 18, Beijing time, OpenAI officially unveiled two new models: GPT-6.2 Astra and GPT-6.2 Sol. One is built for the most complex professional tasks, while the other handles everyday work and multi-step agent workflows. In this release, the more capable model actually became cheaper. Astra's API pricing is 20% lower than that of the previous-generation GPT-6 Astra for both input and output tokens, with cached input dropping by 60%—to $8 per million input tokens and $40 per million output tokens. Sol directly inherits the previous generation's pricing—$2 for input and $10 for output, with no increase at all. What matters more is the cost of completing a task, not just the price per token. Across several core benchmarks published by OpenAI, this decline is even more striking.
01 Astra: More Capable, Lower Bill
Astra's positioning is clear: complex professional work that requires multi-step reasoning and the ability to call on different tools.
On Agents' Last Exam, a benchmark for multi-step professional workflows, Astra scores 5.2 points higher than the previous-generation GPT-6 Astra—yet the cost of completing a single task is less than one-third of its predecessor's.
Compared with competitor Claude Fable 5.2, Astra leads by 5.5 points, at a task cost of only about 5% of the competitor's.
Scientific research is a key scenario OpenAI emphasized for Astra this time. On the Terminal-Bench-Science 0.1 test, Astra's score is 9 percentage points higher than the previous generation's, with task costs also less than one-third.
OpenAI's official page states: "We believe GPT-6.2 Astra is the world's best model for scientists."
In practical work, Astra's improvements are equally tangible. Give it a screenshot of a PowerPoint slide, and it can replicate the original layout, fonts, and color scheme to build a new page directly;
give it a financial report, and it can generate an Excel workbook complete with revenue forecasts and valuation analysis. On the Harvey LAB-AA legal document test, Astra's task pass rate is more than twice that of Claude Fable 5.2, while costing only about one-sixth of the competitor's.
02 Sol: Good Enough for Everyday Work—and Absurdly Cheap
Sol follows a completely different logic. It doesn't aim to be the strongest in every dimension, but rather to minimize cost at a performance level that's good enough.
On Cognition's FrontierCode 2.0 programming test, Sol scores less than 0.5 points behind Astra, yet its task cost is only about one-fifth of Astra's.

Image source: Internet
In OpenAI's in-house LifeVal everyday task evaluation, Sol's score even surpasses that of the previous-generation flagship GPT-6 Astra, at about one-seventh of the cost.
For coding, handling multi-step workflows, and everyday writing, Sol's results are not much different from Astra's—but you pay an order of magnitude less.
For the vast majority of ordinary users and small teams, professional-grade output that was previously unaffordable can now meet most needs with Sol.
03 What's Truly Worth Noticing About This Release
OpenAI did not focus this time on how many more benchmark records were broken. Astra ranks second in third-party comprehensive intelligence indices, behind Claude Fable 5.2; on the coding agent index, Astra is essentially flat compared with the previous generation, and Sol even dropped 3 points.
Feedback from the first batch of developers after the release has also been mixed. Some see no reason not to use models with stronger benchmark scores and lower prices; others believe the smartest AI isn't necessarily the most practical AI, and that in real-world collaborative experience, the new generation hasn't opened up a must-switch gap.

Image source: Internet
But one thing is certain: for the same money, you can buy more. Astra brings complex analysis, research assistance, and legal document processing—previously affordable only to large institutions—down to a cost range that small teams and even individuals can try.
Sol pushes the cost of using AI for everyday programming, writing, and multi-step tasks down to a level close to casual use. Models becoming stronger and cheaper ultimately benefits not the benchmark leaderboards, but everyone who needs AI to get work done.
#Tech #OpenAI