OpenAI has a new answer to the question every AI lab has been asking since last week: what do you do when your rival takes the top spot? On Monday, November 17, the company released GPT-6.2 Astra and GPT-6.2 Sol, two models aimed squarely at Anthropic’s Claude Fable 5.2, which had claimed the lead on the independent Artificial Analysis Intelligence Index just days earlier.
Astra is the flagship, built for demanding professional work such as documents, slides, spreadsheets, and coding. Sol is the everyday model, positioned as the go-to for routine tasks on the same paid plan. Both are available now to ChatGPT Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex, and through the API.
The headline numbers tell a familiar story for this race. Astra is cheaper and nearly as capable as the model ahead of it, while Sol is dramatically cheaper and good enough for most work.
Astra takes second place, but costs a sixth of the leader per task
Independent testing from Artificial Analysis puts GPT-6.2 Astra at 62 on its Intelligence Index v5.0.1 at maximum reasoning effort. That is second overall, behind Claude Fable 5.2 at 65 and ahead of Google’s Gemini 4 Carbon at 60 and Claude Opus 5.5 at 59. Astra improves five points on GPT-6 Astra, which scored 57.
The gap that matters more to buyers is cost. Artificial Analysis measures Astra at about $4 per Intelligence Index task, against $26.54 for Fable 5.2 at its maximum setting. Fable needs its xhigh setting, at roughly $15 per task, to land about level with Astra. Sol, at $1.01 per task, scores 57, level with GPT-6 Astra at roughly 15 percent of that model’s cost per task.
OpenAI’s own benchmarks make the same point more aggressively. The company says Astra scores 5.2 points above GPT-6 Astra on Agents’ Last Exam at less than a third of the cost per task, and 5.5 points ahead of Fable 5.2 at about 5 percent of the cost. On an internal knowledge-work test, GDPval-v2, it gains 6.2 percentage points while using 22 percent fewer output tokens. On FrontierMath’s Erdos problems, it solved 7 of 68, against a previous best of 3.
Where Astra actually leads
Astra is not second everywhere. It leads the Vals Science Index, which tests how well models analyze scientific data and carry out research tasks, and OpenAI calls it “the best model in the world for scientists,” citing a new high on Terminal-Bench-Science, nine percentage points above GPT-6 Astra at less than a third of the cost.
It also leads Artificial Analysis’s AA-3D test, which uses human graders to rate interactive 3D experiences, with 1,318 Elo against Fable 5.2’s 1,271 and Sol’s 1,259. On AA-Robotics, a research preview of real-world planning and action, Astra scores 79 and Sol 78, ahead of Fable 5.2 at 75.
The picture is less flattering in two places. Astra’s coding-agent score is flat with GPT-6 Astra at 53, and Sol actually lost three points against GPT-6.1 Sol. And both models hallucinate a lot on Artificial Analysis’s Omniscience test at maximum effort, Astra at 59 percent and Sol at 55 percent. Those are numbers worth remembering before handing either model a task where a wrong answer is worse than no answer.
What it costs
Pricing is where OpenAI is pressing hardest. Astra costs $8 per million input tokens and $40 per million output tokens, 20 percent below GPT-6 Astra, and cached input drops from $1.00 to $0.40 per million tokens. Sol keeps GPT-6.1 Sol’s pricing of $2 input and $10 output. At roughly 88 rupees to the dollar, Astra’s output token price works out to about 3,520 rupees per million tokens, against about 880 rupees for Sol.
| Model | Input / 1M tokens | Cached input / 1M | Output / 1M tokens |
|---|---|---|---|
| GPT-6.2 Astra | $8 | $0.40 | $40 |
| GPT-6.2 Sol | $2 | $0.10 | $10 |
| Claude Fable 5.2 | $8 | $0.20 | $40 |
| Claude Opus 5.5 | $4 | $0.20 | $20 |
Astra and Fable now have identical standard input and output rates, so the choice between them comes down to benchmark strengths and cost per task, where Astra’s advantage is large. One caveat for heavy users: when input passes 272,000 tokens, OpenAI doubles the input and cached-read rates and raises output by 50 percent for the whole request, taking Astra to $16 input and $60 output.
There is also a safety note OpenAI chose to highlight itself. Astra is the first model it has released externally to reach the “High” threshold for AI self-improvement, its measure of whether a model can meaningfully accelerate AI research. That is a capability claim as much as a warning label, and it will draw scrutiny from evaluators in the weeks ahead.
What this means for the model race
The pattern of the last month is now clear. Anthropic set the capability bar with Claude Fable 5.2, Google answered on speed and price with Gemini 4 Hydrogen, and OpenAI has now answered on cost per task. Nobody holds the lead on every measure, and the gaps between the top three labs are measured in single benchmark points, not leaps.
For users, the practical effect is that the mid-tier model is now the interesting one. Sol matches the previous flagship’s index score at a quarter of Astra’s price, and Astra matches Fable’s capability closely enough that the $22-per-task gap will decide most buying decisions. OpenAI’s bet is that most professional work does not need the absolute best score, it needs a good score at a price that survives a monthly invoice.
That bet is being placed while OpenAI’s own finances are under the microscope. The company posted its first $10 billion revenue quarter in the third quarter of 2026 and still lost $10.9 billion, so cheaper models that keep users inside its ecosystem are not just a product decision, they are a margin decision too. The next independent benchmarks, and the first real-world reports from developers, will show whether Astra’s cost advantage holds up outside the test suite.
