GPT-6 Sol and Luna in the API: half the GPT-5.6 price, but check quality on your workload

General
1

If you run GPT-5.6 Sol, Terra or Luna in production, you now have a pricing decision to make. On September 22, OpenAI released GPT-6 Sol (gpt-6-sol) and GPT-6 Luna (gpt-6-luna) in the API, with per-token rates at roughly half those of their GPT-5.6 predecessors. The cut carries through the pricing rules both generations share, and one independent evaluation finds per-task cost falling by a similar margin. The open question is quality on your workload.

What OpenAI published

OpenAI’s API changelog describes both as reasoning models that accept text and image inputs and return text through the Responses and Chat Completions APIs. The model pages list a 1,050,000-token context window and 128,000 max output tokens for both. The knowledge cutoff is April 20, 2026 for Sol and May 18, 2026 for Luna.

OpenAI’s announcement, posted the same day in the Developer Community by forum moderator VeitB, says the models “build on Astra’s advances in alignment, showing improvements over their GPT-5.6 counterparts,” referring to GPT-6 Astra. That is a vendor claim. The models are also rolling out in ChatGPT Work and Codex, but not in ChatGPT Chat.

The numbers that matter

Standard rates per 1M tokens for prompts up to 272K input tokens:

  • GPT-6 Sol: $2 input, $0.20 cached input, $2.50 cache writes, $10 output.
  • GPT-6 Luna: $0.10 input, $0.01 cached input, $0.125 cache writes, $0.50 output.

Against the GPT-5.6 model pages (my calculations):

  • GPT-5.6 Sol is $4 input, $0.40 cached and $20 output. GPT-6 Sol is exactly half on all three.
  • GPT-5.6 Luna is $0.20, $0.02 and $1.20. GPT-6 Luna halves input and cached input and cuts output by about 58%.
  • GPT-5.6 Terra is $2, $0.20 and $12. GPT-6 Sol matches it on input and is about 17% cheaper on output. OpenAI has not named a GPT-6 Terra.

The announcement compares the new prices with “GPT-5.6 promotional pricing.” The GPT-5.6 Sol page says that promotion runs “at least through November 21, 2026,” and does not say what comes after. OpenAI’s docs and pricing pages don’t say whether the GPT-6 prices are permanent. Tibo (@thsottiaux), in an X post embedded in the announcement thread, says “We are also permanently reducing the API price by 50%”. VentureBeat reports that an OpenAI spokesperson confirmed the prices are permanent.

Pricing rules that carry over

The 272K rule is not new. Both generations’ pages price prompts over 272K input tokens at 2x input and 1.5x output “for the full request.” By my arithmetic, the saving holds above the threshold: GPT-6 Sol goes to $4 input and $15 output, against $8 and $30 for GPT-5.6 Sol.

It is still a cliff in absolute terms. On GPT-6 Sol, a request with 270K input and 10K output costs about $0.64; with 280K input it costs about $1.27.

The same holds for processing tiers. OpenAI’s pricing page lists Batch and Flex rates at half of Standard and Fast mode at 2x for both generations. GPT-6 Sol is $1 input and $5 output on Batch against $2 and $10 for GPT-5.6 Sol, and $4 and $20 on Fast against $8 and $40, so its Fast mode costs the same as GPT-5.6 Sol at Standard. The 10% regional-processing (data residency) uplift applies to eligible models released on or after March 5, 2026, which covers both generations. The model tables in OpenAI’s data residency guide show two differences, both against GPT-6. EU data residency for GPT-6 Sol and Luna is Standard-only, and UAE regional processing covers GPT-5.6 Luna but neither GPT-6 model. EU workloads on Batch, Flex or Fast cannot move tier for tier, and a UAE workload moving off GPT-5.6 Luna would lose in-region processing.

Tier 1 rate limits are 500 RPM and 500,000 TPM for both models. At Tier 5, Sol reaches 15,000 RPM and 40M TPM, and Luna reaches 30,000 RPM and 180M TPM. Both pages list the Free tier as not supported.

Per-task cost and quality

Artificial Analysis published its own evaluation on September 22. On its Intelligence Index (v4.3, per its chart) at max reasoning effort, it puts GPT-6 Sol at $1.06 per task against $1.99 for GPT-5.6 Sol, and GPT-6 Luna at $0.07 against $0.18. It attributes this to the price cut, since both models use “slightly more output tokens per task.” Its headline is that index scores “remain level with GPT-5.6,” with progress in some evaluations and regressions in others. It reports GDPval-AA drops for both models, larger for Sol (about 100 Elo) than for Luna (about 75). On its Coding Agent Index, run in OpenAI’s Codex harness, Sol gains 2 points and Luna drops 2. The evaluation was published before the September 25 image fix described below.

VentureBeat also reports OpenAI’s own per-task figures. OpenAI’s launch post was not accessible, so they are not quoted here.

Two caveats from OpenAI’s pages:

  • On September 25, the changelog recorded a fix for “a bug in image encoding that degraded image understanding” in both models. It recommends rerunning evaluations if your use cases involve image inputs.
  • Chat Completions supports function calling only with reasoning effort set to none.

What is not known

  • OpenAI’s pages cited here give no API retirement date for any GPT-5.6 model. The only statement is that they “remain available during the rollout,” and it refers to ChatGPT and Codex.
  • Per-task cost and quality at your reasoning effort, on your prompts.

My recommendation

  • If you use GPT-5.6 Sol, evaluate GPT-6 Sol at the reasoning effort you use today. Compare cost per completed task, not per-token rates.
  • If you use GPT-5.6 Luna, the output saving is the largest, but Artificial Analysis found Luna regressing on more evaluations. Run your own evaluation set before switching.
  • If you use Terra, try Sol first, since its per-token cost is close.
  • If you send images, only count evaluations run after the September 25 fix.

No migration deadline has been published, so you have time to evaluate before you switch.

Sources