GPT-6.2 Astra and Soul Pricing: API Costs vs Claude Fable 5.2
GPT-6.2 Astra and Soul pricing broken down: API token costs, ChatGPT plan access, and how they stack up against Anthropic's Claude Fable 5.2.

What do GPT-6.2 Astra and Soul cost through the API?
GPT-6.2 Astra costs $8 per million input tokens and $40 per million output tokens through the API. That’s 20% cheaper than the previous Astra generation, and it happens to match Anthropic’s pricing for Claude Fable 5.2 exactly. Soul, the smaller everyday model, stays flat at $2 input and $10 output per million tokens. Both models are live now in ChatGPT (Plus, Pro, Enterprise, Edu) and in Codex, as well as through the API.
TL;DR
- Astra API pricing lands at $8/$40 per million tokens, a 20% cut from the last Astra version and an exact match for what Anthropic charges on Claude Fable 5.2.
- Soul pricing stays flat at $2 input and $10 output per million tokens, making it the budget option for everyday tasks rather than hard reasoning or design work.
- The real pitch is efficiency, not raw intelligence: on one benchmark run, Astra cost roughly $4 per task compared to about $26 for Fable 5.2, even though the two models charge identical API rates.
- Coding performance didn’t move much: Astra scored 53 on a coding benchmark inside Codex, the same as the prior version, while Fable 5.2 inside Claude Code scored 58.
- Access is already open to anyone on a paid ChatGPT plan or using the API, no waitlist, no special tier.
- Early testers praised the cost-to-quality ratio but flagged usability problems like short responses, tasks that stop early, and instructions that are harder to follow than the previous Astra.
- 3D design, CAD, and science tasks are where Astra pulls ahead on a cost basis, with one science benchmark showing Astra scoring close to 78% at under $10 per task, far outperforming Fable 5.2 on that specific cost curve.
How does Astra’s pricing compare to Claude Fable 5.2?
On paper, there’s no difference. Astra and Fable 5.2 both run $8 per million input tokens and $40 per million output tokens. That parity matters because it removes price as a tiebreaker between the two labs’ flagship models. What separates them instead is token efficiency: how many tokens each model burns to finish a given job.
That’s where the gap shows up. On an aggregated benchmark suite, running the full test came out to about $4 per task on Astra versus about $26 per task on Fable 5.2. Since the per-token rate is identical, the entire cost difference comes down to Astra using far fewer tokens to reach an answer. OpenAI’s framing for this release wasn’t “smarter,” it was “cheaper to run the same work.” On a general intelligence benchmark aggregator, Astra scored 62, moving it into second place behind Fable 5.2’s 65 and ahead of Gemini 4 Carbon and Claude Opus 5.5. A five-point jump over the prior Astra model, but still short of the top spot.
Coding is the exception to the efficiency story. Astra scored 53 on a coding benchmark run inside Codex, identical to the previous Astra generation. Soul actually dropped three points on the same test. Fable 5.2 inside Claude Code scored 58. So Astra didn’t get cheaper at coding by using fewer tokens and better, it’s simply unchanged, while remaining behind Fable on raw coding score.
How much does Soul cost, and who is it for?
Soul holds steady at $2 per million input tokens and $10 per million output tokens. That makes it the clear budget pick for high-volume, lower-stakes work, the equivalent slot Anthropic fills with a cheaper tier beneath its flagship Opus-class model. If you’re running large batches of simple tasks, summarization, routine code review, basic content generation, Soul is the one to reach for, since the per-token cost is a quarter of Astra’s input rate and a quarter of its output rate.
The tradeoff shows up in testing. In one direct comparison, a Soul run used around 14,000 tokens and cost about 14 cents, finishing in roughly a minute. A “Pro” variant of Soul used far more tokens and cost close to $14, taking about five minutes. Both produced reasonable output, but neither ranked near the top when judged against other models on quality. Soul is built for jobs where speed and cost matter more than squeezing out the best possible result.
Is GPT-6.2 Astra worth switching to?
If you’re already paying for Fable 5.2 or Opus 5.5 and you’re happy with the output, there’s little financial reason to switch. The per-token price is the same as Fable 5.2, and Astra isn’t ahead on general intelligence benchmarks or on coding. Several early testers, including people who build AI tools and products for a living, reported that Astra’s responses felt too compressed, that it would stop in the middle of longer builds, and that its explanations were harder to follow than the previous Astra version. One tester wouldn’t even publish a review because the experience was rough enough that he couldn’t recommend it. Others called the model “uneven,” delivering strong results on big, ambitious projects while struggling with smaller, more routine asks.
Where Astra does look genuinely strong is 3D design, CAD modeling, and science-oriented tasks. It topped a science benchmark and a robotics-focused benchmark, and multiple reviewers independently praised its ability to go from photos or sketches to fully modeled, assembled 3D objects, down to individual parts and mechanical connections. If your work touches CAD, simulation, game environments, or scientific computation, Astra’s combination of price and capability in that narrow lane is hard to beat right now. For general writing, coding, or anything requiring sustained back-and-forth reasoning, testers consistently pointed toward Fable 5.2 or Opus 5.5 instead.
Soul is worth trying as a default for everyday chat and lighter tasks precisely because it’s a quarter of Astra’s price and still live on every paid ChatGPT plan already. If a task feels like it needs more depth, stepping up to Astra costs more per token but may still land cheaper overall than a slower, token-heavy competitor, depending on the job.
What does the pricing say about where OpenAI is positioning these models?
The decision to match Fable 5.2’s exact API rate, rather than undercut or overprice it, signals that OpenAI isn’t trying to win purely on sticker price. The pitch is efficiency: same rate card, fewer tokens needed to finish comparable work, which pushes effective cost per task down even when the headline number per million tokens is identical. That’s a different competitive lever than the price wars seen in earlier model generations, where labs undercut each other on raw per-token pricing to win developer attention.
It also means the Astra/Soul split mirrors the two-tier structure most labs now use: a frontier-priced flagship for hard problems, and a cheaper, faster model for the bulk of day-to-day requests. Anthropic runs the same pattern with Opus-class and cheaper tiers. For developers building products on these APIs, the practical upshot is that choosing between models is now as much about tokens-per-task and task type as it is about the published per-million-token rate.
Frequently Asked Questions
How much does GPT-6.2 Astra cost per million tokens?
Astra costs $8 per million input tokens and $40 per million output tokens through the API, a 20% reduction from the previous Astra version.
How much does GPT-6.2 Soul cost?
Soul costs $2 per million input tokens and $10 per million output tokens, unchanged from its prior pricing.
Is GPT-6.2 Astra pricing the same as Claude Fable 5.2?
Yes. Both charge $8 per million input tokens and $40 per million output tokens, making per-token API pricing identical between the two flagship models.
Do I need a special plan to access Astra and Soul?
No. Both models are available now to anyone on a paid ChatGPT plan (Plus, Pro, Enterprise, or Edu), inside Codex, and through the standard API, with no separate waitlist.
Is Astra actually cheaper to run than Fable 5.2 in practice?

On one benchmark run, yes: Astra completed the full test for about $4 compared to roughly $26 for Fable 5.2, because it used significantly fewer tokens to reach its answers despite identical per-token pricing.


