GPT-6 Sol and Luna Pricing: $2 / $10 and $0.10 / $0.50 per Million (2026)

OpenAI’s cheaper GPT-6 tiers are out, at half the price of the models they replace. Here’s the full rate card, what each tier is for, and where the open models land next to them.

Three weeks after GPT-6 Astra launched at $10 in and $50 out, OpenAI filled in the rest of the family. GPT-6 Sol and GPT-6 Luna shipped on September 22, 2026, in the API, in Codex, and in ChatGPT Work, and both are priced at half of the GPT-5.6 tiers they replace. Sol is the everyday coding and agent model at $2 per million input tokens and $10 output. Luna is the volume model at $0.10 and $0.50.

That changes the math on the Astra pricing we worked through earlier this month: Astra is now 5x Sol rather than 2.5x, and the gap between OpenAI’s cheapest tier and the open-weight models has closed to almost nothing on paper. Here’s the rate card, the benchmark claims, and the comparison that decides where traffic should go.

The GPT-6 family price list

Astra’s fuller card (Fast mode at 2x, long-context rates over 272K, batch and flex at half price) is in the Astra pricing post. OpenAI’s announcement for Sol and Luna gives the standard rates and the 90% cache discount; it doesn’t state context windows, output limits, or long-context surcharges for the two new tiers, so check the model pages before you assume Astra’s fine print carries over.

Two things worth noticing in that table. Sol’s new price is what the old mid-tier used to cost: GPT-5.6 Terra was $2 in and $12 out, so Sol now sits where Terra was, with a lower output rate. And Luna at $0.10 input is priced under the cheapest open-weight Flash rates in the table below, which OpenAI’s bottom tier hadn’t been before.

What the benchmarks claim

OpenAI’s launch numbers, its own runs, with the reasoning-effort setting it used in parentheses:

The claim OpenAI is making is cost per result, not raw score. Sol at maximum effort lands within about a point of Claude Fable 5 on DeepSWE at what OpenAI says is roughly 80% lower cost, and it beats Astra on AutomationBench when Astra is run at low effort, at a quarter of the cost per task. Luna at max effort is two points behind Sol on DeepSWE at a twentieth of the price. OpenAI also says Sol makes “about half as many mistakes” as GPT-5.6 Sol on its factuality evaluation.

One caution before you compare across launches. OpenAI’s Astra announcement three weeks ago listed GPT-5.6 Sol at 72.7 on DeepSWE, and third-party tables carry the same figure. The new announcement reports GPT-6 Sol at 68.8. Either the harness or the settings differ between the two tables, so don’t read that as the new model scoring lower than the old one; read it as two OpenAI tables that don’t line up, and wait for independent numbers.

Where the open models land

This is the comparison most teams care about. The Sol tier is priced against the open-weight flagships now, and the Luna tier is priced under the open-weight Flash class. Here’s the field on Yotta AI Gateway, flat rates with no peak windows, plus the two DeepSeek and Z.ai Flash models at their vendors’ rates:

Every DeepSWE figure is the vendor’s own run at its own settings, so the column is a rough ordering, not a ranking. Read it that way and three things stand out.

Sol is priced like an open flagship now, not above them. At $2 / $10 it’s cheaper than Kimi K3 on both sides and within a dollar of GLM 5.3 and Qwen 3.8-Max on input. The output rate is still the gap: $10 against $4.40 for GLM 5.3 and $2.97 for DeepSeek V4 Pro, so on generation-heavy traffic the open flagships are still two to three times cheaper. On DeepSWE, Sol’s 68.8 sits between Kimi K3’s 67.5 and GLM 5.3’s 66.9, which is to say the open flagships and OpenAI’s mid-tier are now peers on the coding benchmark they all report. The GLM 5.3 vs DeepSeek V4.1 Flash comparison covers the open side of that in detail.

Luna is the cheapest model in this table, and it’s OpenAI’s. $0.10 input undercuts GLM 5.3 Flash’s $0.15 and DeepSeek V4.1 Flash’s $0.30 peak rate, and the $0.50 output matches GLM Flash exactly. On DeepSWE Luna’s 66.6 is a few points above GLM 5.3 Flash’s 63.4 and a few below V4.1 Flash’s 74.2. If those numbers hold up independently, the open Flash class no longer wins on price by default; it wins on open weights, self-hosting, peak-free flat pricing on the Gateway, and, for DeepSeek, a cache-hit rate ($0.003 off-peak) that’s a third of Luna’s $0.01.

Astra is now a 5x premium over Sol. At Astra’s launch the gap to the tier below was 2.5x. With Sol at half price it’s 5x on both input and output, and OpenAI’s own AutomationBench table shows Sol beating Astra-at-low-effort at a quarter of the cost. The case for Astra is the top end (74.1 on DeepSWE, 72.6 on OSWorld at launch) on the tasks where the last few points matter, which is the same case the Astra launch post made three weeks ago, now with a cheaper fallback one model string below it.

What it means for a routing table

The practical version. Sol takes the slot that GPT-5.6 Sol and Terra used to share: the default for coding agents and repeated analytical work where you want OpenAI’s stack. Luna takes summarization, extraction, classification, and any high-volume call where a Flash-class model is enough. Astra stays on the hard tail.

The open models compete at two of those three slots. Against Sol, GLM 5.3 and DeepSeek V4 Pro run the same class of work at a fraction of the output rate, and Kimi K3 and Qwen 3.8-Max are the multimodal, long-context alternatives; all of them are on Yotta AI Gateway behind one OpenAI-compatible key, so an A/B against gpt-6-sol is a base URL and a model string. Against Luna, DeepSeek V4 Flash on the Gateway at $0.33 / $0.99 and the V4.1 and GLM Flash models at their vendors are the same price class, and the deciding factors are cache behavior, peak windows, and whether you want weights you can host. The best OpenAI API alternatives post and the Best Chinese LLMs roundup cover the field if you’re building that table from scratch.

Nothing here replaces Astra for the work Astra is for. What changed on September 22 is the price of everything below it, and that’s the part of the bill that’s most of the volume.

Availability

Both models are live in the API as gpt-6-sol and gpt-6-luna. In ChatGPT, Sol and Luna are available in ChatGPT Work and in Codex for Plus, Pro, Business, Enterprise, and Edu users, and Free and Go users can try Luna in the desktop app. OpenAI said the rollout would proceed gradually on launch day. Neither is in the regular chat model picker at launch. Neither is on Yotta AI Gateway; the Gateway carries the open models in the table above, and Astra, Sol, and Luna are called through OpenAI directly.

Frequently asked questions

How much does GPT-6 Sol cost? $2 per million input tokens and $10 per million output on OpenAI’s API, with cached input at $0.20. That’s half of GPT-5.6 Sol’s $4 / $20.

How much does GPT-6 Luna cost? $0.10 per million input and $0.50 per million output, with cached input at $0.01. GPT-5.6 Luna was $0.20 / $1.20.

What’s the difference between GPT-6 Sol, Luna, and Astra? Price and target work. Luna ($0.10 / $0.50) is for high-volume routine tasks like summarization and extraction. Sol ($2 / $10) is for coding, debugging, data analysis, and agent work. Astra ($10 / $50) is for the hardest multi-step projects, computer use, and hard science and math. Astra is 5x Sol and 100x Luna on both input and output.

Is GPT-6 Sol better than GPT-5.6 Sol? OpenAI says it makes about half as many mistakes on its factuality evaluation and costs half as much. OpenAI’s DeepSWE figure for GPT-6 Sol (68.8) is lower than the GPT-5.6 Sol figure it published at the Astra launch (72.7), which most likely reflects different test settings; treat cross-launch benchmark comparisons as unresolved until independent numbers arrive.

How does GPT-6 Sol compare to open models like GLM 5.3 or Kimi K3? On price, Sol’s $2 / $10 is cheaper than Kimi K3’s $3 / $15 and close to GLM 5.3 ($1.40 / $4.40) and Qwen 3.8-Max ($1.50 / $4.50) on input, but two to three times more on output than GLM 5.3 or DeepSeek V4 Pro. On DeepSWE, each vendor’s own run puts Sol at 68.8, Kimi K3 at 67.5, and GLM 5.3 at 66.9.

Is GPT-6 Luna cheaper than DeepSeek or GLM Flash? On list input price, yes: $0.10 against $0.15 for GLM 5.3 Flash and $0.30 peak for DeepSeek V4.1 Flash. Output is $0.50, the same as GLM Flash and under DeepSeek’s $1.20 peak. DeepSeek’s off-peak and cache-hit rates ($0.15 / $0.60 and $0.003) still undercut Luna for cache-heavy traffic outside its peak window.

Are GPT-6 Sol and Luna in ChatGPT? Yes, in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu plans, and Luna for Free and Go users in the desktop app. They aren’t in the standard chat model picker at launch.

Are GPT-6 Sol or Luna on Yotta AI Gateway? No. The Gateway carries GLM 5.3, GLM 5.2, Qwen 3.8-Max, Qwen3.8-27B, DeepSeek V4 Pro and V4 Flash, Kimi K3, and Grok 4.6 at flat rates behind one OpenAI-compatible key. OpenAI models are called through OpenAI directly, and a routing layer can send traffic to both.

Bottom line

GPT-6 Sol at $2 / $10 and GPT-6 Luna at $0.10 / $0.50 halve the cost of OpenAI’s everyday tiers and push Astra out to a 5x premium. Sol is now priced like an open-weight flagship and scores like one on the coding benchmark they all report; Luna is priced under the open Flash class on input and level with it on output. The open models keep what OpenAI can’t sell: weights you can host, flat pricing with no peak clock on the Gateway, and, on the DeepSeek side, cache pricing a third of Luna’s. The right move is the same one it was three weeks ago, with cheaper inputs on both sides: put the OpenAI tiers and the open models in one routing table, run your own traffic through it, and let cost per result decide. The open side of that table is on Yotta AI Gateway, one API key away.