A family name is not a complete API quote. Record the versioned model ID, service tier, context band and billing categories before ranking Astra, Sol, Terra or Luna. This guide’s earlier same-week-launch story and customer migration anecdote had no supporting evidence and have been removed.
A dated comparison with a defined scope
On November 18, 2026, OpenAI’s API pricing page lists these Standard, short-context text rates in USD per million tokens:
| Exact model | Input | Output |
|---|---|---|
| gpt-6.2-astra | 8.00 | 40.00 |
| gpt-6.2-sol | 2.00 | 10.00 |
For an illustrative workload of 10 million uncached input tokens and 2 million billed output tokens, the corresponding token bills are 10 × 8 + 2 × 40 = 160 USD and 10 × 2 + 2 × 10 = 40 USD. This excludes tools, other modalities, cache charges and different service/context tiers. It is a price comparison, not a finding of equal task quality.
The cited table establishes those two rows. This article does not manufacture a four-way ranking by filling missing version-specific Terra or Luna terms from their names. Check the provider’s applicable row for the exact requested ID before adding it. Use the model catalog to discover candidates and the calculator to explore your own token mix; provider terms control the actual bill.
Compare completed work
Run a fixed set of representative tasks, recording total billed usage across tool calls and retries. Compare the cost per accepted task and the proportion completed within your deadline. If two models require different prompts or produce different-length answers, token prices alone will not describe your total cost.
Separate cached from uncached inputs and confirm how reasoning output is billed. Keep service-tier and context-band assumptions attached to the result. Do not compare a discounted background request with an interactive request as though their delivery terms match.
Correcting the old GPU calculation
The previous article also claimed a 117-billion-parameter estimate fit one H100 SXM. Its own arithmetic was 117 × 1.1 × 1.25 = 160.875 GB. NVIDIA specifies 80 GB for H100 SXM, so that conclusion was false.
Those parameter and overhead assumptions do not establish any named GPT model’s actual footprint, downloadable weights or deployment support. The calculation is retained here solely to explain the correction. Use the self-hosting worksheet only for a verified deployable checkpoint and tested configuration.