Back to blog
AI Tools10 min read

GPT-6.2 Astra and Sol: Price, Benchmarks and Which to Use

GPT-6.2 Sol costs $2/$10 per million tokens and Astra $8/$40. The benchmarks, what early testers say, and when to pick Sol, Astra, Fable 5.2 or Opus 5.5.

By Moe Lueker

Six weeks ago I gave Claude Opus 5.5 and GPT-6 Astra the same two game prompts, and Opus built the better game both times. Today OpenAI says GPT-6.2 fixes that. It shipped two models on November 17, 2026: GPT-6.2 Sol, the cheap one, and GPT-6.2 Astra, the big one.

I went through the launch post, the pricing page, the Artificial Analysis results and the first reviews from people who had it for weeks. The short version: Sol is the value story, Astra leads on 3D, and the people who used it most still have real complaints.

The short answer

GPT-6.2 is two models: Sol for value, Astra for the hard stuff

OpenAI released GPT-6.2 Sol (gpt-6.2-sol) and GPT-6.2 Astra (gpt-6.2-astra) on November 17, 2026, in ChatGPT Work, Codex and the API. Sol costs $2 in and $10 out per million tokens; Astra costs $8 and $40, which is 20% less than GPT-6 Astra. Use Sol for everyday work, Astra for hard projects and 3D, and keep Claude Opus 5.5 or Fable 5.2 where they already work for you.

Which GPT-6.2 model should you use?#

ModelUse it forWhy
GPT-6.2 SolEveryday work, routine coding, quick prototypesA quarter of Astra's price per token. On FrontierCode 2.0 it scored within 0.5 points of Astra for about a fifth of the cost (OpenAI). Arena called it the value story of the launch.
GPT-6.2 AstraHard projects, visual and interactive 3D buildsFirst on AA-3D at 1,318 Elo, ahead of every Claude model (Artificial Analysis). Early testers say it needs more steering than you'd hope.
Claude Fable 5.2The hardest open-ended work, creative buildsStill first on the Artificial Analysis Intelligence Index (AA). Same price as Astra. Arena found it more creative.
Claude Opus 5.5Daily coding if it already works for youHalf the price of Astra per token. Theo Browne, who tested GPT-6.2 early, still picks Opus 5.5 for his own developer work. It trails both GPT-6.2 models on 3D.

If you're switching from GPT-6 Astra, the answer is easy: GPT-6.2 Astra is cheaper and scores the same or higher on every test below. The real choice is between Sol and Astra, and between OpenAI and the Claude model you already use.

GPT-6.2 price: Sol vs Astra vs Claude#

Prices per million tokens on the standard API tier, from OpenAI's pricing page and Claude's pricing page, checked November 17, 2026.

ModelInputCached inputOutput
GPT-6.2 Sol (gpt-6.2-sol)$2$0.10$10
GPT-6.2 Astra (gpt-6.2-astra)$8$0.40$40
Claude Opus 5.5$4$0.20 (cache read)$20
Claude Fable 5.2$8$0.20 (cache read)$40
GPT-6 Astra (September)$10$1$50

Sol costs a quarter of Astra per token. Opus 5.5 sits halfway between.

List price per million tokens. Astra and Fable 5.2 cost the same per token.

Hover, tap or tab to a model to see all three of its prices.

What does one task cost?

Set how much the model reads and writes. Thinking counts as output on all four models, and agents reread a lot of the same context, which is where caching helps.

  • GPT-6.2 Sol$0.30
  • GPT-6.2 Astra$1.22
  • Claude Opus 5.5$0.61
  • Claude Fable 5.2$1.21
  • GPT-6 Astra$1.55
At these settings, Sol costs $0.30 and Astra $1.22. Same tokens, so Astra is 4.0 times the price. Real runs differ: models use different numbers of tokens for the same job.

Prices per million tokens from OpenAI API pricing (Standard tier) and Claude pricing, both checked November 17, 2026. Prices set by OpenAI and Anthropic; the task cost is my own arithmetic from those list prices. It leaves out cache-write fees, batch discounts and Ultrafast, and stays under 272K input tokens, above which OpenAI charges more.

Sol costs a quarter of Astra per token. Price per million tokens and the cost of one task for GPT-6.2 Sol, GPT-6.2 Astra, Claude Opus 5.5, Claude Fable 5.2 and GPT-6 Astra.

What the table doesn't say on its own:

  • Astra got cheaper. It's 20% below GPT-6 Astra on input and output, and 60% below on cached input. Sol keeps GPT-6.1 Sol's price.
  • Long prompts cost more on OpenAI. Above 272K input tokens, Sol goes to $4 in, $0.20 cached and $15 out. Astra goes to $16, $0.80 and $60.
  • Ultrafast is 3x for now. On the API, Ultrafast costs three times the standard rate through February 28, 2027, then six times from March 1. For Astra that's $24 in and $120 out at launch.
  • Astra and Fable 5.2 cost the same per token. So the Claude comparison that matters on price is Fable, not Opus.

GPT-6.2 cost per task vs intelligence, at every effort level#

Price per token is only half the bill. The other half is how many tokens a model spends thinking, and that changes with the effort setting. Artificial Analysis runs every effort level of every model through its Intelligence Index and reports what one task costs. Here are all four models on one chart:

Astra gets within a point of Fable 5.2 for a sixth of the cost

Each dot is one effort level, from low on the left to max on the right. Up and to the left is better: a higher score for less money. Thick lines are the new GPT-6.2 models.

35404550556065$0.2$0.5$1$2$5$10$20Cost per task (USD, log scale)IndexClaude Opus 5.5Claude Fable 5.2GPT-6.2 SolGPT-6.2 Astra
Compare at effort:
High effortIndexPer task
Claude Fable 5.261.8$9.08
GPT-6.2 Astra60.1$2.07
GPT-6.2 Sol54.9$0.56
Claude Opus 5.551.8$3.46

Hover, tap or tab to any dot for its numbers. Click a dot or a chip to compare all four models at that effort.

Data: Artificial Analysis model pages for GPT-6.2 Astra, GPT-6.2 Sol, Claude Fable 5.2 and Claude Opus 5.5, every effort level, read November 17, 2026. Intelligence Index v5; cost is AAโ€™s weighted average cost of one Index task at list prices. Tested by Artificial Analysis. Download the data (CSV).

Astra gets within a point of Fable 5.2 for a sixth of the cost. Artificial Analysis Intelligence Index against cost per task, at every effort level, for GPT-6.2 Astra, GPT-6.2 Sol, Claude Fable 5.2 and Claude Opus 5.5.
EffortGPT-6.2 AstraGPT-6.2 SolClaude Fable 5.2Claude Opus 5.5
Low45.7 for $0.6640.6 for $0.1853.9 for $3.1034.8 for $0.98
Medium57.0 for $1.6252.7 for $0.4156.4 for $3.8048.2 for $2.45
High60.1 for $2.0754.9 for $0.5661.8 for $9.0851.8 for $3.46
Xhigh61.6 for $2.5255.7 for $0.6562.2 for $14.8357.4 for $6.05
Max62.1 for $4.0057.3 for $1.0164.6 for $26.5459.1 for $10.04

Intelligence Index v5 score and cost per Index task, from the Artificial Analysis model pages, November 17, 2026.

What stands out:

  • Astra is the value pick at the top. Astra on xhigh scores 61.6 for $2.52 a task. Fable 5.2 on xhigh scores 62.2 for $14.83. That's within a point for about a sixth of the money.
  • Astra on high beats Opus 5.5 on max. 60.1 for $2.07, against 59.1 for $10.04.
  • Sol beats Opus 5.5 from low to high, for a fraction of the cost. On high, Sol scores 54.9 for $0.56 and Opus 51.8 for $3.46. Opus pulls ahead on xhigh and max, at about nine to ten times Sol's cost per task.
  • Max is rarely worth it on GPT-6.2. Astra gains half a point from xhigh to max, while the cost rises 59% and the time per task goes from 5.9 to 10.4 minutes.
  • Fable 5.2 still has the top score. 64.6 on max, but at $26.54 a task. You pay for the last few points.

The Claude rows are the setup Artificial Analysis labels "Default Fallback". If you're choosing a Claude effort level, my Claude Fable 5.2 vs Opus 5.5 post from November 9 goes deeper.

GPT-6.2 benchmarks#

Pick a tab. Each chart says who ran the test. The 3D and coding numbers come from Artificial Analysis, an outside lab. The last tab is OpenAI grading its own models.

Sol costs a quarter of Astra and lands within 60 Elo on 3D

Tested by Artificial Analysis. People voted blind on interactive 3D builds. The line shows each modelโ€™s error range: where lines overlap, the gap is too small to call.

Elo score, higher is better.

Fable 5.2 and Sol overlap, so that race is a tie. Opus 5.5 and Sonnet 5.5 are 2 points apart. Hover, tap or tab to a row for its range.

Source: Artificial Analysis AA-3D v1.0, checked November 17, 2026. 55 tasks in AAโ€™s own harness with no network access, blind votes from a recruited panel, every model at Max. ยฑ is AAโ€™s published error range.

GPT-6.2 benchmarks. AA-3D, the AA Coding Agent Index and OpenAI's own results, one tab each.

What stands out:

  • Astra leads 3D, but Sol is close. On AA-3D, Astra scored 1,318 Elo, Fable 5.2 1,271 and Sol 1,259. Fable and Sol's error ranges overlap, so that one is a tie. Opus 5.5, at 1,202, isn't even clearly ahead of Sonnet 5.5 at 1,204.
  • The outside coding check is mixed. The Artificial Analysis Coding Agent Index shows Astra flat against GPT-6 Astra and Sol down 3 points against GPT-6.1 Sol.
  • OpenAI's own gains are real but modest. On GDPval v2, Astra gained 6.2 points over GPT-6 Astra while using 22% fewer output tokens, and Sol gained 5.3 over GPT-6.1 Sol. On Agents' Last Exam, Astra beat GPT-6 Astra by 5.2 points and Fable 5.2 by 5.5.
  • One partner saw a big jump on 3D. Higgsfield told OpenAI that Astra hit 73% one-shot success in its 3D tests, where GPT-6 Astra needed three or more tries.

One more number worth knowing before you hand it a long job. OpenAI's own safety report found both models were more likely to leave problems out of their final reports. In a controlled coding test, Astra left out a blocker in 22.28% of trials, against 11.09% for GPT-6 Astra (OpenAI deployment safety). Codex's built-in instructions cut that down, but those don't apply to plain API calls. If you use the API, ask it to list what it could not finish, and check.

For the September baseline, my GPT-6 Astra launch post from September 3 has the original prices and benchmarks, and my GPT-6.1 Sol vs Astra breakdown from September 29 covers the Sol that 6.2 replaces.

What early testers say#

These are independent testers, not OpenAI's launch partners. Their verdict is more split than the launch post.

Every. The team behind Every had Astra for weeks. They liked its design defaults but found that its designs all start to look the same, its writing gets so terse it drops useful detail, and it sometimes stopped before finishing a full build. They called it the best OpenAI model yet for 3D, not the best model overall, and they didn't test Sol. Every's video review

Arena. Peter Gostev ran Astra and Sol against Fable 5.2 and Opus 5.5 on 3D scenes, UI, animation and games. His verdict: Astra is precise, Fable 5.2 is more creative, and Sol, near the top for a fraction of the price, is the surprise. In games, Astra had much better graphics and Fable slightly better mechanics. Arena's comparison, verdict at 42:16

Theo Browne. Theo tested Astra before launch and wrote a harsh review of it. He gives it credit in places, but his bottom line is blunt: "I would not choose this over Opus 5.5 for my work as a developer." Theo's post

Ethan Mollick. Mollick found Astra uneven and hard to steer. It still built him an ambitious space-elevator ride from scratch in Blender and Unity. Mollick's post

Put together, the pattern is clear. Astra makes the best-looking things, especially in 3D. The people who used it for weeks say it's harder to steer and doesn't always finish. That's exactly what I want to test.

Where to use GPT-6.2#

ChatGPT Work. Both models are in ChatGPT Work, not the regular chat model list (OpenAI's release note). New chats start on Sol at Light effort. To use Astra or turn the effort up, open Advanced and pick the model and effort there.

Plans. OpenAI's launch post lists Plus, Pro, Enterprise and Edu. Its Learn page also lists Business for Work. Enterprise and Edu admins may need to turn it on.

Codex. Update the Codex CLI first. OpenAI's launch thread tells people to update. Then pick the model from the model list.

API. Use gpt-6.2-sol or gpt-6.2-astra.

Ultrafast. In ChatGPT and Codex, the Pro 500 plan gets Ultrafast for both models and Pro 200 gets it for Sol. On the API it's available for both, at 3x the standard rate through February 28, 2027.

My own test: four escape rooms#

I'm running my own test next, and the results aren't in yet. I'm giving GPT-6.2 Sol, GPT-6.2 Astra, Claude Fable 5.2 and Claude Opus 5.5 the same prompt: build a 3D escape room I can actually escape. Then I play every room myself.

It checks the two things the early testers disagree on: does it look best, and does the door actually open? When it's done I'll add the rooms, the scores, the cost of each build and the video here.

Until then, my September game test of Claude Opus 5.5 vs GPT-6 Astra shows how I run these.

Quick answers#

Is GPT-6.2 in regular ChatGPT?

Not in the regular chat model list. OpenAI released GPT-6.2 Sol and Astra in ChatGPT Work, Codex and the API. OpenAI's help page points regular ChatGPT users to Work to pick these models. Work is on the Plus, Pro, Enterprise and Edu plans, and OpenAI's Learn page also lists Business. New Work chats start on Sol at Light effort; open Advanced to switch to Astra or change the effort.

GPT-6.2 Sol vs Astra: which is cheaper?

Sol, by a lot. On the API, Sol costs $2 per million input tokens, $0.10 cached and $10 output. Astra costs $8, $0.40 and $40, so Sol is a quarter of Astra's price per token. Above 272K input tokens both cost more: Sol $4, $0.20 and $15, Astra $16, $0.80 and $60.

Is GPT-6.2 better than Claude Fable 5.2?

It depends on the job. On Artificial Analysis's AA-3D test of interactive 3D builds, Astra scored 1,318 Elo against 1,271 for Fable 5.2. On the Artificial Analysis Intelligence Index at max effort, Fable 5.2 scores 64.6 and Astra 62.1, but Astra on xhigh gets within a point of Fable on xhigh for about a sixth of the cost per task. Astra and Fable 5.2 cost the same per token: $8 in and $40 out. Early testers at Arena called Astra more precise and Fable more creative.

Does GPT-6.2 beat Claude Opus 5.5 for coding?

Not clearly. Artificial Analysis found Astra flat and Sol down 3 points on its Coding Agent Index against the models they replace. Theo Browne, who tested it before launch, said he would not choose it over Opus 5.5 for his developer work. On 3D builds, both GPT-6.2 models scored well above Opus 5.5.

What does GPT-6.2 Ultrafast cost?

On the API, Ultrafast costs 3 times the standard rate through February 28, 2027, and 6 times from March 1, 2027. For Astra that is $24 in and $120 out per million tokens at launch. In ChatGPT and Codex, Ultrafast comes with Pro: the Pro 500 plan gets it for both models and Pro 200 for Sol.

How do I use GPT-6.2 in Codex?

Update the Codex CLI first, which OpenAI's launch thread tells people to do. Then pick gpt-6.2-sol or gpt-6.2-astra in the model list. Sol is the cheaper choice for routine changes. Astra is for hard or visual work.

What are the GPT-6.2 API model names?

gpt-6.2-sol and gpt-6.2-astra.

Published November 17, 2026, the day of release. Sources: OpenAI's GPT-6.2 launch post, OpenAI API pricing, ChatGPT release note, OpenAI deployment safety, Claude pricing, Artificial Analysis AA-3D and its launch article, Every, Arena, Theo Browne and Ethan Mollick. This post is not sponsored by OpenAI or Anthropic. I'll update it with my own test results.

Moe Lueker

Moe shares tool walkthroughs and lessons from real projects. Mechanical engineer, then venture capital, now building AI tools for creators and small businesses. More about Moe

gpt 6.2gpt 6.2 astragpt 6.2 solgpt 6.2 vs claudegpt 6.2 priceopenaiai benchmarksllm pricing