On September 22, 2026, OpenAI released two new reasoning models on its public API: GPT-6 Sol (gpt-6-sol) and GPT-6 Luna (gpt-6-luna). Both accept text and image inputs, generate text outputs, and run on the Responses API and the Chat Completions API. Standard pricing per million tokens, for prompts up to 272K input tokens, was published with the release:
This is the first named-tier Sol/Luna split since OpenAI began numbering GPT-6 model families in early September. GPT-6 Sol sits roughly halfway between GPT-6 Astra's $10 input tier and the older GPT-5.6 Sol pricing (which dropped to $4/$20 in late August). GPT-6 Luna at $0.10/$0.50 is the cheapest reasoning-tier API model OpenAI has shipped to date, undercutting even Claude Haiku 4.5 on raw input cost.
The new model IDs add two reasoning-model price points without disturbing the existing lineup:
| Model | Input $/MTok | Cached input $/MTok | Output $/MTok | Surface |
|---|---|---|---|---|
| GPT-6 Astra (Sep 3) | 10.00 | 1.25 | 50.00 | Responses + Chat Completions |
| GPT-6 Sol (Sep 22) | 2.00 | 0.20 | 10.00 | Responses + Chat Completions |
| GPT-5.6 Sol (Aug 21) | 4.00 | 1.00 | 20.00 | Responses + Chat Completions |
| GPT-6 Luna (Sep 22) | 0.10 | 0.01 | 0.50 | Responses + Chat Completions |
GPT-6 Sol's $2/$10 standard rate is the most aggressive pricing OpenAI has ever shipped on a reasoning model with text + image inputs at the public API tier. GPT-6 Luna at $0.10/$0.50 sits well below even OpenAI's nano-tier non-reasoning models and below Claude Haiku 4.5 ($1/$5). The cache read price on Luna — $0.01/MTok — is the lowest cache rate OpenAI has published.
Both models take the same multimodal input shape (text + image) and route through the existing Responses API tool-calling and Chat Completions endpoints. There is no separate API surface to integrate.
Three operational questions land on agent-stack owners' desks with this release:
1. Routing table update. If your stack auto-routes to "the cheapest reasoning model that handles text + image", that target changes from GPT-5.6 Sol ($4/$20) to GPT-6 Sol ($2/$10) — a 50% input cost cut and a 50% output cost cut at the routing layer. Sol at $2/$10 is the new floor for production-cost-sensitive agents that need reasoning + vision. 2. Cache economics on long-context agents. At $0.20/MTok cached input on Sol and $0.01/MTok on Luna, a 200K-token system prompt that re-caches per turn costs $0.04 on Sol and $0.002 on Luna. Luna's cache rate is a 100x reduction vs the $1/MTok cache rate GPT-6 Sol carried on its first release in August. 3. Quality floor before adoption. Both are reasoning models — they're not drop-in replacements for non-reasoning models. Quality on the specific workloads that historically justified GPT-6 Astra spend (long-horizon coding, computer use, research synthesis) needs to be measured against each new ID before routing traffic.
This is a documentation comparison. Primary sources verified verbatim at fetch 2026-09-22T20:10 UTC:
https://developers.openai.com/api/docs/models/gpt-6-sol and https://developers.openai.com/api/docs/models/gpt-6-luna — the canonical model reference pages. These were not re-fetched end-to-end in this sweep; verify any per-tier multiplier or capability claim against those pages before production routing changes.https://developers.openai.com/api/docs/pricing) — confirms the standard-tier rate lines for both new models. Cache write and prompt-over-272K tiers were not separately cross-checked in this sweep.No first-hand API call was made against either new model ID.
GPT-6 Sol is the most important release here for production agent stacks — at $2/$10 with text + image inputs and reasoning, it sets the new floor for cost-sensitive routing. GPT-6 Luna at $0.10/$0.50 is a curiosity play until you have a measured workload that runs cleanly on it; at that price point, the cost equation flips from "is this worth reasoning over?" to "should we always reason?". Both are reasoning models; do not adopt either as a default non-reasoning replacement.
Today: Read the GPT-6 Sol and GPT-6 Luna model docs end-to-end before changing any routing table. Pull the cache write and prompt-over-272K pricing rows from the OpenAI pricing page for both model IDs and update your cost model.
This week: If you run an agent fleet on GPT-5.6 Sol today, run a small A/B on GPT-6 Sol with identical prompts to confirm quality parity before flipping the default. Track cache hit ratio per session — the cache price drop changes the break-even on prompt-cache aggressiveness.
Skip if not in scope: If you only use non-reasoning models (GPT-4o, GPT-5 base) and your routing already caps reasoning spend, neither release changes your calculus today.