Writing / Engineering notes

Is GPT-6.2 Astra Worth Four Times the Price of Sol?

Two Pac-Man games, one stubborn ghost, and a fourfold gap in API pricing. Here’s what early testers found and what I’ll try in Codex this week.

GPT-6.2 Astra and Sol launch graphic with two stars tracing arcs across a black sky
Read as Markdown

GPT-6.2 Astra built a $2.04 game of Pac-Man with one odd bug. The orange ghost often stayed home. In the same developer’s test, Sol built its version for 49 cents. Both finished in under seven minutes, and Astra earned the higher score. I like that the tester played long enough to notice what one ghost was doing. Read the tester’s notes on Hacker News or play the two versions on PacBench.

OpenAI released both models on November 17 in Codex, ChatGPT Work, and the API. Astra is the one OpenAI recommends for complex coding projects, frontends, and 3D design. Sol is its faster everyday option for jobs that can span several apps. What interests me as a developer is how often I’d reach for Astra when Sol costs a quarter as much per token. OpenAI’s launch announcement

Watch my 58-second Astra and Sol overview on YouTube.

What Astra’s extra effort looks like

My favorite example from the announcement is a paper lamp. OpenAI’s team started with photos of the curved balconies on Chicago’s Marina City towers. They used Astra to work through the design in CAD, then assembled the lamp by hand. The demo includes a website Astra built so you can design one yourself. A paper shade is a pretty unforgiving test of a 3D model. Sooner or later, the edges have to meet. See the lamp and the design process.

OpenAI also reports that Astra used 22% fewer output tokens than GPT-6 Astra on GDPval v2, its test of professional assignments, while scoring 6.2 percentage points higher. For an API user, that’s a potential saving on top of the new rate. OpenAI’s evaluation results

Where Sol saved money, and what it missed

Plotly co-founder Chris Parmer noticed that both models made fewer tool calls on his team’s data analytics exam. Sometimes, he wrote, a model “finished early” and missed the answer. You can see that tradeoff in Sol’s results. GPT-6.2 Sol answered 40 of 43 questions for $1.22. GPT-6.1 Sol had answered all 43 for $2.61 at the same per-token rates. Parmer’s comment and Plotly’s full results.

All three misses involved patterns the team had planted in the data. In one case, Sol never queried the part of the dataset that held the clue. Plotly ran each model once, so I wouldn’t pick a winner on three answers. It did give me a question to borrow for my own testing. When it says it’s done, what did it actually check?

Pricing and availability

Standard API prices in US dollars per million tokens
Token typeGPT-6.2 AstraGPT-6.2 Sol
Input$8.00$2.00
Cached input$0.40$0.10
Output$40.00$10.00

Compared with GPT-6 Astra, the new Astra charges 20% less for regular input and output and 60% less for cached input. A task that also takes fewer tokens saves money twice. Sol keeps GPT-6.1 Sol’s rates across all three categories. OpenAI API pricing

In ChatGPT Work and Codex, Plus, Pro, Enterprise, and Edu users have access to both models. The $500 Pro plan includes Ultrafast for either model. On the $200 Pro plan, Ultrafast starts with Sol. Eligible Enterprise and Edu users get Ultrafast once their workspace admin enables it. OpenAI’s availability details

In the API, use gpt-6.2-astra or gpt-6.2-sol. Both support Fast and Ultrafast at launch. For the first three months, OpenAI is cutting Ultrafast rates in half for API and Enterprise users, bringing them to three times the Standard rate. OpenAI Developers’ pricing update

If you’re using the API for a coding session, keep the project instructions at the start of the conversation and append each new request at the end. Keeping that shared prefix intact helps both models reuse cached input and carry reasoning into the next turn, which can reduce cost and latency. OpenAI’s guide to preserving the prompt prefix

What I’m testing in Codex this week

I’ll give Astra and Sol the same frontend change in separate copies of a real project. I want to see the first version each produces, then open both on my phone. Does the layout hold up? Can I actually use the controls? I’ll also give each model a bug to track down and look at which checks it runs before it hands the code back to me.

I’m particularly interested in the second prompt. If I have to point out a broken interaction, a quick first pass can turn into a long afternoon. I’ll share the prompts, the two versions side by side, and what I had to fix on my YouTube channel, @proflead.