Skip to content
Claudexia TeamMODELS

GPT-6 Sol and Luna: price, API, and when they beat Astra

GPT-6 Sol and Luna launched on 22 September 2026: OpenAI charges $2 / $10 and $0.10 / $0.50 per 1M tokens, Claudexia $0.38 / $1.86 and $0.08 / $0.30.

GPT-6 Sol and GPT-6 Luna were released on 22 September 2026. Sol is built for complex code and agent workflows, Luna for narrow, high-volume tasks. Both cost far less than GPT-6 Astra: 5 and 100 times less at OpenAI, and roughly 2 times and 9.75 to 13 times less in the Claudexia catalog. The choice comes down to two questions. What does one answer cost at your volume, and do you need reasoning at all? On Sol and Luna the none level switches it off.

What these models are

According to the OpenAI changelog, both models appeared on 22 September 2026. They are reasoning models: they accept text and images, answer in text, and work through the Responses API and Chat Completions. On the model pages OpenAI describes GPT-6 Sol as built for complex coding and agentic workflows, and GPT-6 Luna as its most efficient model for focused, high-volume tasks. Sol's knowledge cutoff is 20 April 2026, Luna's is 18 May 2026.

Three days after release OpenAI fixed an image encoding bug that made both models worse at understanding pictures (entry of 25 September). The company advises rerunning evaluations and retrying affected workflows. If you tested Sol or Luna on images in late September, revisit those results.

Here is where things stand now. The GPT-6 Sol page points to GPT-6.1 Sol as the newer Sol model, which came out on 29 September. In the model-choice block on the catalog page on 19 November 2026, OpenAI names GPT-6.2 Astra, GPT-6.2 Sol and GPT-6 Luna, and there is no GPT-6 Sol card. Luna is still recommended. Sol has been followed by two newer models, GPT-6.1 Sol and GPT-6.2 Sol.

Parameters

ParameterGPT-6 SolGPT-6 Luna
API IDgpt-6-solgpt-6-luna
Context at OpenAI1,050,000 tokens1,050,000 tokens
Maximum output128,000 tokens128,000 tokens
Knowledge cutoff20 April 202618 May 2026
reasoning_effort levelsnone, low, medium (default), high, xhigh, maxnone, low, medium (default), high, xhigh, max
Context and output in the Claudexia catalog1,000,000 and 128,0001,000,000 and 128,000

Of GPT-6 Sol, Luna, Astra, GPT-6.1 Sol and GPT-6.2 Sol, only the first two have a none level. For GPT-6 Astra OpenAI lists low, medium, high, xhigh and max. GPT-6.1 Sol has the same list without none, and GPT-6.2 Sol rejects a request with none. For short answers, classification, and chat where every second of latency shows to the user, skipping reasoning is useful.

Function calling comes with a caveat. According to both model pages, built-in tools and function calling are meant for the Responses API, and in Chat Completions function calling works only when reasoning_effort is none. For clients that speak only Chat Completions, that is the single path to tool calling on Sol and Luna in OpenAI's documentation. How the Claudexia gateway handles reasoning_effort and functions on these models, we did not check.

Price at OpenAI and on Claudexia

Prices per 1M tokens. OpenAI: the model pages, Standard tier for requests up to 272K input tokens. Claudexia: the live price list for the Premium channel on 19 November 2026.

Per 1M tokensSol, OpenAISol, ClaudexiaLuna, OpenAILuna, Claudexia
Input$2.00$0.38$0.10$0.08
Cache read$0.20$0.12$0.01$0.10
Cache write (OpenAI: 30 minutes)$2.50$0.46$0.125$0.34
Output$10.00$1.86$0.50$0.30

The OpenAI cache-write rates in the table are for a minimum retention of 30 minutes. In the Responses API you can choose 12 hours instead, and the write then costs 2.5 times the ordinary input rate: $5.00 for Sol and $0.25 for Luna. The Claudexia price list has a single write rate with no lifetime, and we did not check how the gateway passes that parameter.

Two cells in the table behave differently from OpenAI. At OpenAI a cache read costs 10% of the input price. For gpt-6-luna our price list puts the cache read above the plain input price, $0.10 against $0.08, so caching saves nothing on that model. For gpt-6-sol a cache read costs $0.12 against $0.38 for input, so caching pays off, but the discount is smaller than at OpenAI. The price list gives no reason for either, and we do not know it.

Three more notes on price.

  • Long prompts. At OpenAI a prompt longer than 272K input tokens is billed at double the input and cache rates and at 1.5 times the output rate, for the whole request. The Claudexia catalog has a single rate per model. Whether the gateway applies such a surcharge, we did not check.
  • OpenAI discounts. The Sol and Luna pages say Batch and Flex cost 50% of Standard, and Fast mode costs twice as much. Claudexia has no Batch API, so the Batch discount is unavailable through the gateway. We did not check Flex support or pricing through Claudexia.
  • Channel. Both models exist only in Premium. On 19 November 2026 Cheap holds seven models: Claude Haiku 5.5, Opus 5.5, Sonnet 5.5 and four image models, with no text GPT models among them. A key that has only Cheap gets 403 model_not_in_channel, as the channels documentation describes.

Now an example in numbers. One thousand requests, each with 3,000 input tokens and 500 output tokens, with no caching and no reasoning tokens.

ModelClaudexiaOpenAI
gpt-6-luna$0.39$0.55
gpt-6-sol$2.07$11.00
gpt-6-astra$4.29$55.00

This is arithmetic on price lists, not a measurement. Reasoning tokens in real requests add to the output, and the bill will be higher. Take your own numbers from the usage field in the responses.

When Sol and Luna beat Astra

The first argument is obvious: price. It is worth counting how many attempts Sol can afford, though. On Claudexia one Astra run from the example costs $4.29 and one Sol run costs $2.07, so two Sol runs ($4.14) cost about the same as one Astra run. At OpenAI the gap is wider: Astra costs exactly 5 times more on both input and output. If every run uses the same number of tokens, Sol stays cheaper as long as it needs four runs or fewer for a task Astra solves in one. Five Sol runs ($55.00) cost the same as one Astra run, and the saving is gone. Which model solves your task on the first try, only a run on your data will show. The method takes 20 minutes and is described in Public benchmarks miss your task.

The second argument: reasoning can be switched off. For extracting fields, tags and short summaries, the none level turns reasoning off, and with it the spend on reasoning tokens. The lowest level on Astra in OpenAI's list is low.

The third argument concerns Luna. OpenAI built the Decisions API around it: the v1/decisions endpoint turns text and images into typed answers, the beta started on 6 October, general availability came on 22 October, and one request can carry up to 400 questions. Whether v1/decisions works through Claudexia, we did not check, so count this point only if you call OpenAI directly.

When you need Astra

OpenAI describes GPT-6 Astra as a model for complex coding and professional work, and GPT-6.1 Sol as a model with performance close to Astra at a lower price. We would set it up like this: give the bulk and simple work to Luna, the main work to Sol, and switch on Astra for a session where Sol has failed twice. More on Astra is in our Russian article "GPT-6 Astra: what it is, price and how to use it".

GPT-6.2 Astra and GPT-6.2 Sol: what is missing from the catalog

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on 17 November 2026. According to the model pages Astra costs $8 and $40 per 1M tokens and Sol costs $2 and $10, with 1,050,000 tokens of context and up to 128,000 tokens of output. The changelog warns that tool calling requires the Responses API.

On 19 November 2026 these models are not on Claudexia. We checked in two ways. The live price list (/api/public/pricing) returns 64 models in Premium and 7 in Cheap, and GET /v1/models returns the same 64 IDs, with neither gpt-6.2-astra nor gpt-6.2-sol among them. For what is known about these models and how they relate to the rest of the lineup, see GPT-6 and GPT-5.6 compared. Our Russian articles on Codex model choice and on ChatGPT 6 access cover them too.

What to use instead of GPT-6.2 Sol

The catalog has gpt-6.1-sol at $0.12 / $0.58, and at OpenAI it costs the same as GPT-6 Sol, $2 / $10. So on Claudexia gpt-6-sol costs about 3.2 times more than gpt-6.1-sol, on input and on output alike. Picking gpt-6-sol makes sense in two cases: you need the none level on a Sol-class model, or your pipeline is already tuned on it. In every other case, start with gpt-6.1-sol.

How to call it through Claudexia

The address for the OpenAI SDK and compatible clients is https://api.claudexia.tech/v1. The key starts with sk_cdx_, and it is best kept in an environment variable. If you already have an OpenAI integration, read Moving to the OpenAI-compatible format.

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.claudexia.tech/v1",
    api_key=os.environ["CLAUDEXIA_API_KEY"],
)

reply = client.chat.completions.create(
    model="gpt-6-luna",
    messages=[{"role": "user", "content": "Pull the payment date and amount out of this email."}],
)
print(reply.choices[0].message.content)

Signing up at claudexia.tech takes a login and password, or Google, GitHub or Telegram, and there is no email field. The minimum top-up is $1, paid in rubles through SBP (Russia's Faster Payments System), with a Russian card, or in crypto. A key on the Premium channel opens both Luna and Sol, and the Claude models too, for example Claude Sonnet 5.5. Only a chat request validates the key itself: GET /v1/models returns 200 even without a key.

FAQ

How does GPT-6 Sol differ from GPT-6.1 Sol and GPT-6.2 Sol?

At OpenAI all three cost the same, $2 per 1M input tokens and $10 per 1M output tokens. They differ in knowledge cutoff (20 April, 30 April and 1 June 2026) and in reasoning levels: GPT-6 Sol supports none, the other two do not. Of the three, Claudexia carries gpt-6-sol and gpt-6.1-sol.

Is GPT-6.2 on Claudexia?

Not on 19 November 2026: it is in neither the price list nor the GET /v1/models response. The current list is on the models page.

What are the context and output limits of Sol and Luna?

OpenAI states 1,050,000 tokens of context and up to 128,000 tokens of output for both. The Claudexia catalog lists 1,000,000 and 128,000.

Can I use Sol or Luna on the Cheap channel?

No. On 19 November 2026 Cheap has no text GPT models, and a key that has only Cheap gets 403 model_not_in_channel. You need a key on the Premium channel.

What should I pick for Codex?

If you want Sol, start with gpt-6.1-sol and keep Luna for mechanical edits. Profiles, limits and a monthly token estimate are in our Russian article on Codex model choice.