GPT-6.2 Astra and Sol: Features, Benchmarks, Tests, and Pricing
OpenAI's GPT-6.2 Astra and Sol cut the cost per task sharply. We cover what's new, the benchmarks, which tier to use, pricing, and a hands-on test against GPT-6 Astra.
November has been a crowded month for frontier models. Anthropic shipped Claude Fable 5.2, Google has Gemini 4 Carbon in the field, and today, after a week of speculation, OpenAI answered with two models at once: GPT-6.2 Astra and GPT-6.2 Sol.
GPT-6.2 Astra is OpenAI's new frontier model for hard professional work, such as decks, financial workbooks, legal drafting, CAD, and long agent runs. GPT-6.2 Sol is the cheaper everyday model for coding, writing, and agent loops, and it lands close to Astra on coding for about a quarter of the token price.
I just talked about their differences, but both cost the same or less per task than the models they replace: in the API, Astra's input and output prices are 20% lower than GPT-6 Astra's, and Sol keeps GPT-6.1 Sol's prices.
In this article, I'll cover everything new with GPT-6.2 Astra and Sol, looking at the new features, exploring the benchmarks, and putting Astra through its paces with 2 hands-on tests against GPT-6 Astra.
As you can see, new models are being released at a fast clip. To keep up with the AI and data space, definitely subscribe to our weekly newsletter, The Median.
TL;DR
GPT-6.2 Astra is OpenAI's new flagship for hard professional work; GPT-6.2 Sol is the everyday model.
The headline gain is cost per task: Astra does more for far less than GPT-6 Astra.
Sol comes close to Astra on coding at a fraction of the price, so it should be your default.
Claude Fable 5.2 still leads several coding and reasoning benchmarks, so OpenAI doesn't sweep.
In our hands-on tests, Astra was far cheaper than GPT-6 Astra, but GPT-6 Astra read a messy dataset more sensibly.
Switch if you pay for GPT-6 Astra today; the savings alone justify it.
What Are GPT-6.2 Astra and Sol?
GPT-6.2 Astra replaces GPT-6 Astra at the top, and Sol replaces GPT-6.1 Sol one tier down. Two things set them apart from the last generation:
Slightly smarter, much cheaper per task. OpenAI's own capability index puts Astra just 1.7 points above GPT-6 Astra, but its cost per task fell by more than half on several benchmarks.
A new risk rating. Astra is the first model OpenAI has released that meets its High threshold for AI self-improvement, so it ships with OpenAI's strictest safeguards yet.
Here's the cost story on one slide, built by GPT-6.2 Astra. (Improved slides are one feature. We will get to that.)
GPT-6.2 Astra and Sol Key Features
Most of what's new targets professional output and efficiency rather than raw intelligence. Here are the four changes that matter.
Build slides, workbooks, and documents that follow your template
GPT-6.2 Astra can take a slide template, as an original file or even a screenshot, and match its layout, colors, and typography down to the font size. OpenAI trained it on the conventions of fields like law, sales, finance, and accounting, with titles that state the main takeaway and less filler in the body copy.
I put this new feature to the test throughout this article, including the slide above. Astra wrote each slide as HTML, which I rendered to an image without changes.
The same training shows up in spreadsheets. On Artificial Analysis's AA-AnalystAgent, which asks models to find figures in real spreadsheets and reports, Astra answered 62.5% of questions correctly on all five independent attempts, a new high according to OpenAI. ClickUp reports that Astra edged out its production model in blind testing of the apps, dashboards, and slides ClickUp Brain builds, with about 30% fewer review rounds.
Work in CAD and simulation tools with computer use
GPT-6.2 Astra's computer use now works down to individual parts and connections in CAD and simulation tools. We expect this is a big deal for mechanical and product engineers. OpenAI's demos include a car's chassis dynamics simulated with a Python solver and animated in Blender, a 1:8-scale RC buggy modeled part by part from public photos and build manuals, and a paper lamp designed from reference photos through CAD revisions.
We can't easily build a CAD project of our own to test this, so here is OpenAI's driving simulation demo. Astra built the 3D car, wrote the Python solver for the tires and suspension, and animated the result in Blender:
Both models reach better results with fewer output tokens, which is where most of the cost savings come from. On OpenAI's internal GDPval v2, GPT-6.2 Astra scores 6.2 points higher than GPT-6 Astra while using 22% fewer output tokens. Astra also chooses its tools better, including when to use a plugin instead of computer use, and Basis reports its average time for long-horizon accounting work fell by half.
The system card also has better news on prompt injection, where hidden instructions in an email or web page try to hijack the agent. Astra's slide sums up the safety trade-offs:
Slide by GPT-6.2 Astra
Write more secure code, within tighter safeguards
GPT-6.2 Astra and Sol find and exploit vulnerabilities more effectively, and OpenAI added training signals for writing secure code and patches that preserve intended behavior. Both meet the Critical threshold for cybersecurity under OpenAI's Preparedness Framework, as GPT-6 Astra did.
The trade-off is more refusals. OpenAI launched both models with tighter safeguards that also restrict some legitimate vulnerability research and security testing. Approved defenders still use GPT-6 Astra and GPT-6 Cyber, which became available two weeks ago, through OpenAI Daybreak.
How Do GPT-6.2 Astra and Sol Perform on the Benchmarks?
GPT-6.2 Astra leads OpenAI's professional-work and science benchmarks, but Claude Fable 5.2 leads several coding and reasoning results.
In the following tables, all scores are vendor-reported. I made the best score in each row bold.
Professional work and agentic workflows
On Agents' Last Exam v1.1, which tests long, multistep professional workflows, GPT-6.2 Astra scores 78.3% at $2.89 per task. GPT-6 Astra scores 73.1% at $8.99 per task, and Claude Fable 5.2 scores 72.8% at $60.05.
Gemini 4 Carbon comes closest at 77.4%, but costs $4.60 per task. The rest of the professional-work results follow the same pattern:
Benchmark
What it measures
GPT-6.2 Astra
GPT-6.2 Sol
GPT-6 Astra
GPT-6.1 Sol
Claude Fable 5.2
Agents' Last Exam v1.1
Long-horizon professional workflows
78.3%
73.0%
73.1%
72.5%
72.8%
GDPval v2 (OpenAI internal)
Multistep consulting, finance, and design assignments
50.4%
46.9%
44.2%
41.5%
44.8%
Harvey LAB-AA v1.1
Legal deliverables that meet every criterion
15.0%
11.1%
8.6%
6.9%
7.2%
AutomationBench 1.0.6
Business workflow automation
52.4%
41.2%
41.4%
36.1%
51.4%
LifeVal (OpenAI)
Challenging everyday tasks
86.1%
83.8%
79.8%
79.5%
n/a
Astra leads every row here among the models in the table, and doubles Fable 5.2 on Harvey's legal benchmark. Gemini 4 Carbon still edges it on AutomationBench with 53.9%, and on GDPval v2, Carbon's 48.8% sits between Astra and Sol.
OpenAI's GDPval v2 chart shows where the savings come from: for the same number of output tokens, the new models score higher than the ones they replace.
Coding and security
Coding is where Sol earns its place: on FrontierCode 2.0, it scores 43.8% at $1.29 per task, against Astra's 44.2% at $5.97.
Benchmark
What it measures
GPT-6.2 Astra
GPT-6.2 Sol
GPT-6 Astra
GPT-6.1 Sol
Claude Fable 5.2
FrontierCode 2.0
Code changes that meet maintainers' standards for merging
44.2%
43.8%
42.8%
39.0%
49.1%
SecureSWE (OpenAI internal)
Functional and secure code for security-sensitive components
65.7%
55.6%
48.7%
49.7%
n/a
PatchBench
Security patches
79.7%
75.6%
70.9%
71.0%
n/a
ExploitBench Internal Port
Arbitrary code execution rate across V8 and Chrome
53.5%
50.5%
31.5%
21.5%
n/a
Claude Fable 5.2 keeps the lead on FrontierCode 2.0, and Claude Opus 5.5 at max effort (45.5%) also beats both GPT-6.2 models there. The security rows are where the generational jump is clearest, though OpenAI didn't report Fable 5.2 on any of them.
OpenAI's FrontierCode chart makes Sol's case: at max effort, it reaches Astra's score for about a fifth of the cost per task.
Science, CAD, and general reasoning
On Terminal-Bench-Science 0.1, which tests scientific work in a terminal, GPT-6.2 Astra scores 77.1%, 9 points above GPT-6 Astra at less than a third of the cost per task.
Benchmark
What it measures
GPT-6.2 Astra
GPT-6.2 Sol
GPT-6 Astra
GPT-6.1 Sol
Claude Fable 5.2
Terminal-Bench-Science 0.1
Scientific research workflows in a terminal
77.1%
68.0%
68.1%
57.0%
73.3%*
Vals Science Index v1
Interpreting biological data and running experiments
68.4%
58.5%
65.1%
59.6%
63.7%
BenchCAD 2.1 Core
CAD modeling tasks
82.5%
73.1%
62.6%
52.6%
83.5%
Artificial Analysis Intelligence Index v5.0.1
Composite of 9 evaluations (index score)
62.1
57.3
56.9
54.6
64.6
GPQA Diamond (no tools)
Graduate-level science questions
97.0%
96.0%
96.0%
n/a
n/a
Fable 5.2 still leads the Intelligence Index and BenchCAD, but at $26.54 and $50.77 per task against Astra's $4.00 and $5.43. If you need the absolute best CAD output, Fable is still the pick; if you need it many times a day, Astra gets you 99% of the way for about a tenth of the price.
On Terminal-Bench-Science, Astra reaches the top score at $7.15 per task at max effort, against $23.80 for GPT-6 Astra and $45.21 for Claude Fable 5.2.
Which Tier Should You Use?
GPT-6.2 Sol should be your default, and GPT-6.2 Astra the model you escalate to. Sol costs a quarter of Astra's token price, and on coding it gives up less than a point.
GPT-6.2 Astra made the same case against itself in its own deck:
Slide by GPT-6.2 Astra
Both models expose five reasoning efforts in OpenAI's charts, from low to max, and the effort setting matters as much as the tier. On Agents' Last Exam, Sol at high effort (72.0% at $0.69 per task) lands about a point behind GPT-6 Astra at max effort (73.1% at $8.99) for one-thirteenth of the cost. Ultrafast mode is available for both models in the API and for Pro 500 subscribers in ChatGPT.
Use case
Tier
Why
Everyday coding, refactors, and code review
Sol
Within 0.5 points of Astra on FrontierCode 2.0 at about one-fifth of the cost
High-volume agent loops and automations
Sol
$2 / $10 per million tokens, unchanged from GPT-6.1 Sol
Client-ready decks, financial workbooks, and memos
Astra
Template matching and the GDPval v2 lead are Astra-specific
Legal drafting and research
Astra
15.0% on Harvey LAB-AA against Sol's 11.1%
CAD, simulation, and scientific analysis
Astra
About 9 points ahead of Sol on BenchCAD and Terminal-Bench-Science
Testing GPT-6.2 Astra: Hands-On Examples
The benchmarks above are OpenAI's own, so I ran 2 tests against GPT-6.2 Astra and its predecessor, GPT-6 Astra, with identical prompts and identical tooling. For the dashboard test, both models ran in OpenCode with a pinned tool surface (reading, searching, and editing files only), the same high reasoning effort, one attempt, and a fresh session.
Test 1: An honest dashboard from messy data
The task: build a single-file HTML dashboard from a CSV of earnings and employment outcomes for 173 U.S. college majors, for a careers advisor comparing fields for clients. It had to compare categories on earnings and employment, surface the highest- and lowest-earning majors, embed the data instead of fetching it, and fit one 1440×900 screen.
The real test is in the data, which has problems the prompt never mentions:
The top earner, Petroleum Engineering at $110,000, rests on a survey sample of 36 people, and 76 of the 173 majors have samples under 100.
The lowest earner, Library Science at $22,000, rests on a sample of 2.
Food Science is missing its graduate totals.
Categories range from a single major (Interdisciplinary, 12,296 graduates) to 13 majors with 1.3 million graduates (Business), so how you average them matters.
It probes 2 launch claims at once: Astra's better analytical workbooks and its efficiency, with fewer output tokens for better results. Here's what GPT-6.2 Astra produced:
The main findings:
Both models shipped a working dashboard on the first try. Each embedded all 173 majors, drew its own SVG charts without a charting library, fit one screen, and loaded without errors.
They read the same data differently, and it reorders the categories. GPT-6.2 Astra ranked each category by the median of its majors' medians, which counts a major with a few thousand graduates the same as one with 200,000. GPT-6 Astra weighted each major by its number of graduates. The result: Health comes 10th at $35,000 on GPT-6.2 Astra's dashboard and 4th at $40.6k on GPT-6 Astra's, because nursing, with 209,394 graduates and a $48,000 median, counts for what it is.
GPT-6 Astra also put sample sizes up front. Its main table shows that the top 4 earners rest on samples of 36, 7, 3, and 16. GPT-6.2 Astra shows them only after you select a major, with a one-line warning about unstable rankings.
GPT-6.2 Astra was cheaper and leaner. It cost $1.79 against $2.93 for GPT-6 Astra, with 27,821 output tokens against 35,298.
Let's take a closer look at GPT-6 Astra's version. Compare its category chart with GPT-6.2 Astra's above: Health sits 4th here and 10th there, and Biology & Life Science drops from 6th to 11th.
Neither method is wrong as math, and GPT-6.2 Astra labels its chart as a median of major medians. But a careers advisor wants to know where graduates actually land, and only the graduate-weighted version answers that.
Test 2: A slide deck from a template screenshot
The task: turn the research notes behind this article into a 5-slide deck for engineering managers deciding whether to switch, in the style of one template slide. The models got the template only as a screenshot, and the prompt asked them to match it "down to the font sizes" without naming the font or any size.
It tests OpenAI's slide claim directly. Because the models had to see an image, each ran as a single API call through OpenRouter at the same high reasoning effort, and I added GPT-6.2 Sol, since decks are everyday work. This is the template they saw, which I designed for the test:
The main findings:
All 3 models rebuilt the template from the screenshot alone. Each identified the font as Poppins and kept the green accent bar, the kicker line, the grey stat card, the source line, and the page number. Every title states a takeaway, like the template's.
All 3 matched the template closely enough that I couldn't tell them apart by eye. Checked in the browser, only GPT-6.2 Sol hit the template's font sizes exactly. Astra, the model OpenAI says matches templates "down to the font size", was off by a pixel or two.
GPT-6.2 Astra was the most economical of the two Astras. It cost $0.22 and 4,968 output tokens, against $0.59 and 11,247 for GPT-6 Astra. Sol cost $0.07.
Astra's 5 slides are the captioned ones you've seen in this article, and every figure on them matches my notes. GPT-6 Astra's deck was just as usable, so on this task the generations don't separate.
So is GPT-6.2 Astra a real step up? On cost, yes: it finished both tests for 39% to 63% less, which matches OpenAI's efficiency claim. On judgment, not quite: the older model was the one that weighted the data the way a careers advisor needs it, while the slides came out about even.
One run per task can't rank two models, but it fits OpenAI's own Blocker Deception finding: check what GPT-6.2 Astra leaves out, not just what it builds.
GPT-6.2 Astra and Sol Pricing and Availability
GPT-6.2 Astra is 20% cheaper than GPT-6 Astra on input and output tokens and 60% cheaper on cached input, while GPT-6.2 Sol keeps GPT-6.1 Sol's prices. Standard API rates per million tokens:
Model
Input
Cached input
Output
GPT-6.2 Astra
$8.00
$0.40
$40.00
GPT-6.2 Sol
$2.00
$0.10
$10.00
The lower prices are only part of the saving. Because both models use fewer output tokens, the cost per task falls further, which is why Astra's Agents' Last Exam run costs less than a third of GPT-6 Astra's. OpenAI also recommends keeping prompt prefixes unchanged and appending new instructions, to get more cache hits and preserve reasoning across turns.
Astra's own summary of the new prices:
Slide by GPT-6.2 Astra
In ChatGPT, both models are available to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex from launch day. Enterprise and Edu users get access once their workspace admin turns it on.
Ultrafast is available with either model for Pro 500 subscribers, and Pro 200 subscribers get it for the first time, starting with GPT-6.2 Sol. OpenAI's announcement doesn't mention the Free plan.
How to Get Access to GPT-6.2 Astra and Sol?
Both models are generally available in the OpenAI API as gpt-6.2-astra and gpt-6.2-sol, with Ultrafast available for both at launch. A minimal request with the OpenAI Python SDK looks like this:
from openai import OpenAI
client = OpenAI()
response = client.responses.create(
model="gpt-6.2-sol",
reasoning={"effort":"high"},input="Summarize this quarterly report in 5 bullet points: ...",)print(response.output_text)Powered by
Both models are also listed on OpenRouter (openai/gpt-6.2-astra and openai/gpt-6.2-sol) and in the OpenCode model catalog.
GPT-6.2 is a price cut that OpenAI has turned into a capability release. Astra beats GPT-6 Astra on every professional-work benchmark OpenAI reports, at a third of the cost per task, and Sol gets close enough on coding that most teams should start there.
It doesn't settle the overall race. Claude Fable 5.2 still leads on FrontierCode and the Intelligence Index, Gemini 4 Carbon leads AutomationBench, and in our tests the older Astra read a messy dataset more sensibly. But I'd switch from GPT-6 Astra today for the savings.
Slide by GPT-6.2 Astra
If you're keen to learn more about building with OpenAI's models, I recommend checking out our Working with the OpenAI API course.
Author
Josef Waples
I'm a data science editor with contributions to research articles in scientific journals. I'm especially interested in linear algebra, statistics, R, and the like.
FAQs
What is GPT-6.2 Astra and how is it different from GPT-6.2 Sol?⌄
GPT-6.2 Astra is OpenAI's frontier model for hard professional work like decks, financial workbooks, legal drafting, and CAD. GPT-6.2 Sol is the cheaper everyday model for coding, writing, and agent loops. Sol costs a quarter of Astra's token price and comes within 0.5 points of Astra on FrontierCode 2.0, while Astra leads on professional and science benchmarks.
How much does GPT-6.2 Astra cost in the API?⌄
GPT-6.2 Astra costs $8 per million input tokens, $0.40 per million cached input tokens, and $40 per million output tokens. That is 20% below GPT-6 Astra on input and output and 60% below on cached input. GPT-6.2 Sol costs $2, $0.10, and $10, the same as GPT-6.1 Sol.
Is GPT-6.2 Astra better than GPT-6 Astra?⌄
On OpenAI's benchmarks, yes: GPT-6.2 Astra beats GPT-6 Astra on every professional-work result, such as Agents' Last Exam (78.3% vs 73.1%), at about a third of the cost per task. In our 2 hands-on tests it was cheaper too, but GPT-6 Astra weighted a messy dataset more sensibly.
Where can I access GPT-6.2 Astra and GPT-6.2 Sol?⌄
Both models are available in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu users, with Enterprise and Edu access turned on by a workspace admin. Developers can call gpt-6.2-astra and gpt-6.2-sol in the OpenAI API, and both are listed on OpenRouter.
Should I use GPT-6.2 Sol or GPT-6.2 Astra for coding?⌄
Use GPT-6.2 Sol for most coding. On FrontierCode 2.0 it scores 43.8% against Astra's 44.2% at max effort, for $1.29 per task instead of $5.97. Switch to Astra for security-sensitive code, where it leads Sol on SecureSWE, or for CAD and scientific work.
OpenAI's GPT-6.1 Sol claims GPT-6 Astra-level capability at a fifth of the task cost, with a 1,050,000-token context window and $2/$10 per million token pricing.
OpenAI's GPT-6 Astra tops computer use, coding, and math benchmarks. Full breakdown of features, scores vs Claude and Gemini, pricing, and how to access it.
Start your journey developing AI-powered applications with the OpenAI API. Learn about the functionality that underpins popular AI applications like ChatGPT.
November has been a crowded month for frontier models. Anthropic shipped Claude Fable 5.2, Google has Gemini 4 Carbon in the field, and today, after a week of speculation, OpenAI answered with two models at once: GPT-6.2 Astra and GPT-6.2 Sol.
GPT-6.2 Astra is OpenAI's new frontier model for hard professional work, such as decks, financial workbooks, legal drafting, CAD, and long agent runs. GPT-6.2 Sol is the cheaper everyday model for coding, writing, and agent loops, and it lands close to Astra on coding for about a quarter of the token price.
I just talked about their differences, but both cost the same or less per task than the models they replace: in the API, Astra's input and output prices are 20% lower than GPT-6 Astra's, and Sol keeps GPT-6.1 Sol's prices.
In this article, I'll cover everything new with GPT-6.2 Astra and Sol, looking at the new features, exploring the benchmarks, and putting Astra through its paces with 2 hands-on tests against GPT-6 Astra.
As you can see, new models are being released at a fast clip. To keep up with the AI and data space, definitely subscribe to our weekly newsletter, The Median.
TL;DR
What Are GPT-6.2 Astra and Sol?
GPT-6.2 Astra replaces GPT-6 Astra at the top, and Sol replaces GPT-6.1 Sol one tier down. Two things set them apart from the last generation:
Here's the cost story on one slide, built by GPT-6.2 Astra. (Improved slides are one feature. We will get to that.)
GPT-6.2 Astra and Sol Key Features
Most of what's new targets professional output and efficiency rather than raw intelligence. Here are the four changes that matter.
Build slides, workbooks, and documents that follow your template
GPT-6.2 Astra can take a slide template, as an original file or even a screenshot, and match its layout, colors, and typography down to the font size. OpenAI trained it on the conventions of fields like law, sales, finance, and accounting, with titles that state the main takeaway and less filler in the body copy.
I put this new feature to the test throughout this article, including the slide above. Astra wrote each slide as HTML, which I rendered to an image without changes.
The same training shows up in spreadsheets. On Artificial Analysis's AA-AnalystAgent, which asks models to find figures in real spreadsheets and reports, Astra answered 62.5% of questions correctly on all five independent attempts, a new high according to OpenAI. ClickUp reports that Astra edged out its production model in blind testing of the apps, dashboards, and slides ClickUp Brain builds, with about 30% fewer review rounds.
Work in CAD and simulation tools with computer use
GPT-6.2 Astra's computer use now works down to individual parts and connections in CAD and simulation tools. We expect this is a big deal for mechanical and product engineers. OpenAI's demos include a car's chassis dynamics simulated with a Python solver and animated in Blender, a 1:8-scale RC buggy modeled part by part from public photos and build manuals, and a paper lamp designed from reference photos through CAD revisions.
We can't easily build a CAD project of our own to test this, so here is OpenAI's driving simulation demo. Astra built the 3D car, wrote the Python solver for the tires and suspension, and animated the result in Blender:
Source: OpenAI
Get more done per token in agent workflows
Both models reach better results with fewer output tokens, which is where most of the cost savings come from. On OpenAI's internal GDPval v2, GPT-6.2 Astra scores 6.2 points higher than GPT-6 Astra while using 22% fewer output tokens. Astra also chooses its tools better, including when to use a plugin instead of computer use, and Basis reports its average time for long-horizon accounting work fell by half.
The system card also has better news on prompt injection, where hidden instructions in an email or web page try to hijack the agent. Astra's slide sums up the safety trade-offs:
Slide by GPT-6.2 Astra
Write more secure code, within tighter safeguards
GPT-6.2 Astra and Sol find and exploit vulnerabilities more effectively, and OpenAI added training signals for writing secure code and patches that preserve intended behavior. Both meet the Critical threshold for cybersecurity under OpenAI's Preparedness Framework, as GPT-6 Astra did.
The trade-off is more refusals. OpenAI launched both models with tighter safeguards that also restrict some legitimate vulnerability research and security testing. Approved defenders still use GPT-6 Astra and GPT-6 Cyber, which became available two weeks ago, through OpenAI Daybreak.
How Do GPT-6.2 Astra and Sol Perform on the Benchmarks?
GPT-6.2 Astra leads OpenAI's professional-work and science benchmarks, but Claude Fable 5.2 leads several coding and reasoning results.
In the following tables, all scores are vendor-reported. I made the best score in each row bold.
Professional work and agentic workflows
On Agents' Last Exam v1.1, which tests long, multistep professional workflows, GPT-6.2 Astra scores 78.3% at $2.89 per task. GPT-6 Astra scores 73.1% at $8.99 per task, and Claude Fable 5.2 scores 72.8% at $60.05.
Gemini 4 Carbon comes closest at 77.4%, but costs $4.60 per task. The rest of the professional-work results follow the same pattern:
Astra leads every row here among the models in the table, and doubles Fable 5.2 on Harvey's legal benchmark. Gemini 4 Carbon still edges it on AutomationBench with 53.9%, and on GDPval v2, Carbon's 48.8% sits between Astra and Sol.
OpenAI's GDPval v2 chart shows where the savings come from: for the same number of output tokens, the new models score higher than the ones they replace.
Coding and security
Coding is where Sol earns its place: on FrontierCode 2.0, it scores 43.8% at $1.29 per task, against Astra's 44.2% at $5.97.
Claude Fable 5.2 keeps the lead on FrontierCode 2.0, and Claude Opus 5.5 at max effort (45.5%) also beats both GPT-6.2 models there. The security rows are where the generational jump is clearest, though OpenAI didn't report Fable 5.2 on any of them.
OpenAI's FrontierCode chart makes Sol's case: at max effort, it reaches Astra's score for about a fifth of the cost per task.
Science, CAD, and general reasoning
On Terminal-Bench-Science 0.1, which tests scientific work in a terminal, GPT-6.2 Astra scores 77.1%, 9 points above GPT-6 Astra at less than a third of the cost per task.
Fable 5.2 still leads the Intelligence Index and BenchCAD, but at $26.54 and $50.77 per task against Astra's $4.00 and $5.43. If you need the absolute best CAD output, Fable is still the pick; if you need it many times a day, Astra gets you 99% of the way for about a tenth of the price.
On Terminal-Bench-Science, Astra reaches the top score at $7.15 per task at max effort, against $23.80 for GPT-6 Astra and $45.21 for Claude Fable 5.2.
Which Tier Should You Use?
GPT-6.2 Sol should be your default, and GPT-6.2 Astra the model you escalate to. Sol costs a quarter of Astra's token price, and on coding it gives up less than a point.
GPT-6.2 Astra made the same case against itself in its own deck:
Slide by GPT-6.2 Astra
Both models expose five reasoning efforts in OpenAI's charts, from low to max, and the effort setting matters as much as the tier. On Agents' Last Exam, Sol at high effort (72.0% at $0.69 per task) lands about a point behind GPT-6 Astra at max effort (73.1% at $8.99) for one-thirteenth of the cost. Ultrafast mode is available for both models in the API and for Pro 500 subscribers in ChatGPT.
Testing GPT-6.2 Astra: Hands-On Examples
The benchmarks above are OpenAI's own, so I ran 2 tests against GPT-6.2 Astra and its predecessor, GPT-6 Astra, with identical prompts and identical tooling. For the dashboard test, both models ran in OpenCode with a pinned tool surface (reading, searching, and editing files only), the same high reasoning effort, one attempt, and a fresh session.
Test 1: An honest dashboard from messy data
The task: build a single-file HTML dashboard from a CSV of earnings and employment outcomes for 173 U.S. college majors, for a careers advisor comparing fields for clients. It had to compare categories on earnings and employment, surface the highest- and lowest-earning majors, embed the data instead of fetching it, and fit one 1440×900 screen.
The real test is in the data, which has problems the prompt never mentions:
It probes 2 launch claims at once: Astra's better analytical workbooks and its efficiency, with fewer output tokens for better results. Here's what GPT-6.2 Astra produced:
The main findings:
Let's take a closer look at GPT-6 Astra's version. Compare its category chart with GPT-6.2 Astra's above: Health sits 4th here and 10th there, and Biology & Life Science drops from 6th to 11th.
Neither method is wrong as math, and GPT-6.2 Astra labels its chart as a median of major medians. But a careers advisor wants to know where graduates actually land, and only the graduate-weighted version answers that.
Test 2: A slide deck from a template screenshot
The task: turn the research notes behind this article into a 5-slide deck for engineering managers deciding whether to switch, in the style of one template slide. The models got the template only as a screenshot, and the prompt asked them to match it "down to the font sizes" without naming the font or any size.
It tests OpenAI's slide claim directly. Because the models had to see an image, each ran as a single API call through OpenRouter at the same high reasoning effort, and I added GPT-6.2 Sol, since decks are everyday work. This is the template they saw, which I designed for the test:
The main findings:
Astra's 5 slides are the captioned ones you've seen in this article, and every figure on them matches my notes. GPT-6 Astra's deck was just as usable, so on this task the generations don't separate.
So is GPT-6.2 Astra a real step up? On cost, yes: it finished both tests for 39% to 63% less, which matches OpenAI's efficiency claim. On judgment, not quite: the older model was the one that weighted the data the way a careers advisor needs it, while the slides came out about even.
One run per task can't rank two models, but it fits OpenAI's own Blocker Deception finding: check what GPT-6.2 Astra leaves out, not just what it builds.
GPT-6.2 Astra and Sol Pricing and Availability
GPT-6.2 Astra is 20% cheaper than GPT-6 Astra on input and output tokens and 60% cheaper on cached input, while GPT-6.2 Sol keeps GPT-6.1 Sol's prices. Standard API rates per million tokens:
The lower prices are only part of the saving. Because both models use fewer output tokens, the cost per task falls further, which is why Astra's Agents' Last Exam run costs less than a third of GPT-6 Astra's. OpenAI also recommends keeping prompt prefixes unchanged and appending new instructions, to get more cache hits and preserve reasoning across turns.
Astra's own summary of the new prices:
Slide by GPT-6.2 Astra
In ChatGPT, both models are available to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex from launch day. Enterprise and Edu users get access once their workspace admin turns it on.
Ultrafast is available with either model for Pro 500 subscribers, and Pro 200 subscribers get it for the first time, starting with GPT-6.2 Sol. OpenAI's announcement doesn't mention the Free plan.
How to Get Access to GPT-6.2 Astra and Sol?
Both models are generally available in the OpenAI API as
gpt-6.2-astraandgpt-6.2-sol, with Ultrafast available for both at launch. A minimal request with the OpenAI Python SDK looks like this:Both models are also listed on OpenRouter (
openai/gpt-6.2-astraandopenai/gpt-6.2-sol) and in the OpenCode model catalog.For a full build with the previous generation, see our GPT-6 Astra API tutorial and GPT-6.1 Sol tutorial; switching the model ID is the main change.
Final Thoughts
GPT-6.2 is a price cut that OpenAI has turned into a capability release. Astra beats GPT-6 Astra on every professional-work benchmark OpenAI reports, at a third of the cost per task, and Sol gets close enough on coding that most teams should start there.
It doesn't settle the overall race. Claude Fable 5.2 still leads on FrontierCode and the Intelligence Index, Gemini 4 Carbon leads AutomationBench, and in our tests the older Astra read a messy dataset more sensibly. But I'd switch from GPT-6 Astra today for the savings.
Slide by GPT-6.2 Astra
If you're keen to learn more about building with OpenAI's models, I recommend checking out our Working with the OpenAI API course.