GPT-6.2 Astra Benchmark Results for Data Analysis

Nov 17, 2026: OpenAI's new frontier model, tested on real-world data analysis the day it launched.

At Plotly, we keep a close eye on model releases. Our agentic analytics products, Plotly Studio Desktop and Plotly Studio Embedded, and our open source libraries Plotly and Dash work with any model, and our customers look to us for recommendations.

Today OpenAI released GPT-6.2 Astra, which its announcement calls "our new frontier model for your hardest projects", alongside GPT-6.2 Sol, "our efficient, versatile model for everyday work" (our GPT-6.2 Sol results are here). OpenAI says GPT-6.2 Astra "costs significantly less per task than GPT-6 Astra, driven by better token efficiency, a 20% discount on input and output tokens, and a deeper 60% discount on cached input tokens." On OpenRouter it's $8 per million input tokens, $0.40 per million cached input tokens and $40 per million output tokens, down from $10, $1 and $50 for GPT-6 Astra, with a context window of 1.05 million tokens and up to 128,000 output tokens. The weights are closed, and OpenAI's announcement doesn't say what GPT-6.2 Astra was built from or how it relates to GPT-6 Astra in training.

We ran it through our data analytics exam. GPT-6.2 Astra passed 41 of 43 questions (95%, an A) for $4.11 a full run, about a fifth of GPT-6 Astra's $19.97. That's about 10 cents a question, or about $96 per 1,000 questions. Its median question took 12.4 seconds, about half of GPT-6 Astra's 26.9, which makes it the fastest model we've tested that scores 41 or more.

Data analytics benchmark cost-accuracy frontier with every model labelled and GPT-6.2 Astra circled at 95% for $4.11

GPT-6.2 Astra vs. GPT-6 Astra, GPT-6.1 Sol and GPT-6.2 Sol

GPT-6.2 Astra

GPT-6 Astra

GPT-6.1 Sol

GPT-6.2 Sol

Score

41 / 43

43 / 43

43 / 43

40 / 43

Analytics questions

20 / 20

20 / 20

20 / 20

20 / 20

Ambiguous questions

5 / 5

5 / 5

5 / 5

5 / 5

Impossible requests

5 / 5

5 / 5

5 / 5

5 / 5

Data quality

3 / 3

3 / 3

3 / 3

3 / 3

Statistics traps

5 / 5

5 / 5

5 / 5

5 / 5

Hidden needles

3 / 5

5 / 5

5 / 5

2 / 5

Open weights

No

No

No

No

Cost of one full run

$4.11

$19.97

$2.61

$1.22

Per person per year*

$239

$1,161

$152

$71

Median time per question

12.4 seconds

26.9 seconds

17.6 seconds

10.5 seconds

*1,000 people each asking 10 questions a day, 250 working days a year, at OpenRouter's prices when each model ran.

GPT-6.2 Astra ties six other models at 41, including GPT-6 Sol, Claude Opus 5.5 and Gemini 4 Argon, and 8 of the 53 models on our board score higher. GPT-6 Astra and GPT-6.1 Sol both scored 43. Each model gets a single run on this benchmark, and in our repeat runs of other models scores moved by up to three points between runs, so a question or two either way is within normal run-to-run variation.

It isn't on the cost-accuracy frontier, because GPT-6.1 Sol scores higher for less, at $2.61. It is on the speed-accuracy frontier, which is now Mercury 3, GPT-6.2 Sol, GPT-6.2 Astra and GPT-6.1 Sol: no model we've tested is both faster and more accurate.

Why it costs less

There are two reasons, of about equal weight: lower prices and less work. Re-pricing GPT-6 Astra's own run at GPT-6.2 Astra's rates would cut it from $19.97 to about $11, roughly 40% less. The rest of the saving, from about $11 to $4.11, comes from GPT-6.2 Astra doing less work to answer the same 43 questions.

GPT-6 Astra

GPT-6.2 Astra

Price per million input / output tokens

$10 / $50

$8 / $40

Price per million cached input tokens

$1.00

$0.40

Tool calls

266

148

Model calls

249

171

Input tokens

9.1 million

4.4 million

Output tokens

78,000

34,000

Input tokens read from cache

93%

98%

Cost of one full run

$19.97

$4.11

Totals for the full exam of 43 questions. The saving is across the board, not only on the hardest questions. On the 38 questions without a hidden pattern, GPT-6.2 Astra made about 2.8 tool calls per question against 4.7 for GPT-6 Astra, about 40% fewer, and still answered every one correctly. On the five hidden-pattern questions it searched much less, running about 9 database queries per question against about 15, and that's where both of its misses are.

What it got right

GPT-6.2 Astra was perfect on five of the six kinds of question. It got all 20 straightforward analytics questions, it asked about or stated its assumptions on all five ambiguous questions, and it turned down all five impossible requests instead of inventing an answer. It caught all three data quality problems we planted in the warehouse, and it avoided all five statistics traps, where the obvious calculation gives a misleading answer.

It's efficient, too. It made about 4 model calls per question and wrote about 800 output tokens per question, and 98% of the input tokens it read came from the provider's cache. One tool call hit a transient error and was retried immediately, and every question finished.

What it got wrong

It missed two questions, and we checked each one by hand against the transcript and the answer key. Both were genuine misses, not grading problems.

Both were hidden patterns, the hardest part of our exam. Each one points toward something unusual in the data without naming it, and the model has to go looking. In one, GPT-6.2 Astra never isolated the pattern we planted. In the other, it found the right place to look but argued against the actual cause. Across the five hidden-pattern questions it ran about 9 queries each, against a median of 25 on our board.

Speed

GPT-6.2 Astra's median question took 12.4 seconds, and the whole exam about 13 minutes, against 26.9 seconds and about 32 minutes for GPT-6 Astra. Only four models we've tested have a faster median: Mercury 3, gpt-oss-120b, GPT-6.2 Sol and Nemotron 3.5 Super VL, and none of them scores as high. Both GPT-6.2 models join the speed-accuracy frontier, shown here.

Accuracy vs. median time per question for every model on the board, with the speed-accuracy frontier

Which OpenAI model to choose

For data analysis, GPT-6.1 Sol is still the cheapest model we've tested with a perfect score, at $2.61 a run. GPT-6.2 Astra is the fastest of the top scorers, at 12.4 seconds a question against 17.6, and costs about a fifth of GPT-6 Astra. GPT-6.2 Sol scored one point lower for under a third of the cost, and it's a little faster still (more in our GPT-6.2 Sol post).

OpenAI's models over time

We've tested 16 OpenAI models, from GPT-4o in November 2024 to the two released today, shown here with every other model in grey. GPT-4o passed 20 of 43 for $5.83; GPT-6.2 Sol passes twice as many for about a fifth of the cost. Speed has moved around more: GPT-4o was quick at 7 seconds a question but got fewer than half right, GPT-5 took 82, and the GPT-6.2 models, at 10.5 and 12.4 seconds, are the fastest OpenAI models we've tested that score an A.

We wrote up GPT-6.1 Terra, OpenAI's mid-tier model, in October (GPT-6.1 Terra results).

Accuracy vs. cost of all 16 OpenAI models we've tested: GPT-6.2 Astra and GPT-6.2 Sol drawn solid, the earlier OpenAI models outlined and labelled, and all other models in grey
Accuracy vs. median time per question for all 16 OpenAI models we've tested: GPT-6.2 Astra and GPT-6.2 Sol drawn solid, the earlier OpenAI models outlined and labelled, and all other models in grey

The Test Harness

We run every model in Plotly Studio Embedded, a drop-in analytics agent that runs on Dash apps. This harness is tuned to run data analytics queries and make charts with multi-turn tool calling, with less emphasis on writing text. It's meant to put the data front and center.

Claude Sonnet 5.5 in Plotly Studio Embedded answering "Which is our best-performing wind farm?"

Frequently asked questions

How good is GPT-6.2 Astra at data analysis?

On PlotlyDataAnalyticsBench it passed 41 of 43 questions (95%, an A). It was perfect on analytics questions, ambiguous questions, impossible requests, data quality problems and statistics traps, and found three of the five hidden patterns.

How much does GPT-6.2 Astra cost?

$8 per million input tokens, $0.40 per million cached input tokens and $40 per million output tokens. Our whole exam cost $4.11, about 10 cents a question, or about $239 per person per year at 10 questions a day, against $1,161 for GPT-6 Astra.

Why is GPT-6.2 Astra cheaper than GPT-6 Astra?

Lower prices and less work, in about equal parts. Its token prices are 20% lower and its cached input price 60% lower, and on our exam it made about half as many tool calls and read about half as many input tokens as GPT-6 Astra.

Is GPT-6.2 Astra faster than GPT-6 Astra?

Yes. Its median question took 12.4 seconds against 26.9, and the whole exam about 13 minutes against about 32. It's the fastest model we've tested that scores 41 or more.

Is GPT-6.2 Astra open source?

No. Its weights are closed. It's available through OpenAI's API, in ChatGPT Work and Codex, and on OpenRouter. For how open-weight models compare, see our open vs. closed models post.

Should I use GPT-6.2 Astra or GPT-6.2 Sol?

On our exam they're close: GPT-6.2 Astra scored 41 and GPT-6.2 Sol 40. GPT-6.2 Sol cost $1.22 against $4.11 and had a median of 10.5 seconds against 12.4 (GPT-6.2 Sol results).

About the benchmark

We run new models through the same exam as they're released, and we'll keep sharing what we find. We ran GPT-6.2 Astra through OpenRouter on the day it arrived there, with the same harness and grading as every other model.

To see all 53 models on one chart and in a full results table, along with cost, speed and how the exam works, see Introducing DataAnalyticsBench.

The exam is private and we're sharing it with researchers. We're keeping it private so that the answer key doesn't make its way into the training set. For the same reason, the data is fabricated: the warehouse is a synthetic wind-energy fleet, built to look like a real operator's data, so its answers don't exist anywhere else.

If you'd like access, reach out to Chris Parmer at Plotly.

bluesly
x
instagram
youtube
medium
fcaebook

Product

© 2026
Plotly. All rights reserved.
Cookie Preferences
aicpa/soc2