Affiliate disclosure: the optional Novita comparison link below is sponsored. We may earn a commission at no extra cost to you.
Which API model should you test?
OpenAI’s current API pricing page lists GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna as flagship candidates. The chart and table use their exact catalog identifiers. Legacy GPT-6 Sol and GPT-5.6 variants are not presented as current default recommendations.
| Workload | First candidate to evaluate | Escalation check |
|---|---|---|
| Defined extraction, routing or classification | GPT-6 Luna | Repeated failures or correction time |
| Complex coding, computer use or professional work | GPT-6.1 Sol | Quality misses on your acceptance set |
| Harder tasks with expensive failed attempts | GPT-6 Astra | Whether better results justify total task cost |
| Pricing data: How we verify → | ||
These are evaluation starting points, not measured quality rankings. OpenAI describes Sol as a lower-cost candidate near Astra’s performance, but explicitly recommends comparing them on your tasks. Count accepted results, retries, latency and review effort.
Use the complete OpenAI pricing page and token calculator for rate selection. The API cost calculation guide explains workload arithmetic; Anthropic pricing provides a separate provider control.
Separate Standard rates from other charges
The headline chart compares recorded Standard token rates, not an all-in quote. Its source-verification dates are independent of this article’s update date.
Caching: uncached input, cache reads and cache writes are different billing dimensions. Stable prompt prefixes may help; a changing prefix may not reuse the same cached region.
Long context: the official rate card has separate short- and long-context columns. Sol’s model documentation says requests above its input threshold receive higher rates for the full request. Monthly aggregate tokens do not establish whether individual requests cross that threshold.
Service mode: Standard, Batch, Flex, Fast and Ultrafast have distinct rates and availability. Select the actual model and mode before estimating. Do not apply the Standard chart to a premium-mode request.
Tools and regions: tool-specific charges and eligible regional processing uplifts can be additional. Check the exact endpoint, tool and regional terms; the cheapest global text-token rate is not every deployment’s price.
What changed in Sol’s current documentation?
The current GPT-6.1 Sol model page specifies Ultrafast using model: "gpt-6.1-sol" with service_tier: "ultrafast" on the Responses API. Ultrafast is a service mode here, not evidence for a different model identifier.
It also states that Sol supports US and EU data residency, including Fast and Ultrafast, subject to eligibility. This replaces the older blanket EU Fast-mode restriction previously carried in this guide. Do not transfer the statement to every other model or account.
The same documentation directs tool-calling integrations to Responses. Chat Completions is supported without tool calling. Check feature and endpoint compatibility separately from the token rate.
Reduce cost per accepted result
- Route easy work to the smallest candidate that passes your tests.
- Move delay-tolerant jobs to an eligible discounted mode.
- Keep reusable context stable and trim unnecessary tool results.
- Cap output and repeated agent loops.
- Include retries, review and downstream failures in the cost metric.
For a separate open-model control, evaluate a supported Novita endpoint on the same acceptance set. Verify the exact hosted model and billing terms rather than assuming capability parity.
Subscriptions do not establish API entitlement
This guide prices developer API work. It does not promise any model, quota or API credit in a ChatGPT or Codex subscription. Check the ChatGPT subscription guide and exact current plan terms separately.
Historical launch articles document what was announced on their publication dates. They cannot establish current plan access, regional eligibility or a new verification clock.
FAQ
Is GPT-6.1 Sol Ultrafast still unpriced?
The current official rate card publishes Ultrafast rates. Check the chosen model, request context and region; the Standard chart is not an Ultrafast estimate.
Is Fast mode prohibited with EU residency?
Not as a blanket claim for Sol. Its current documentation explicitly includes US and EU residency with Fast and Ultrafast, subject to eligibility.
Does a ChatGPT subscription pay my API bill?
This guide does not establish that entitlement. Budget metered API work separately unless your exact current product terms explicitly provide a credit or other benefit.