GPT-6 Sol and Luna: Complete Guide, Pricing and Benchmarks

OpenAI filled out the GPT-6 generation on 22 September 2026, about 90 minutes after Anthropic shipped Claude Opus 5.5. Until today the only GPT-6 model was GPT-6 Astra, the $10/$50 flagship from 3 September. For everyday work, developers were still using the GPT-5.6 Sol, Terra and Luna models. GPT-6 Sol and GPT-6 Luna now take the mid-tier and cheap-tier slots. They cost about half as much as the models they replace, and they score about the same, better on some benchmarks and worse on others.

This guide covers what each model is, the exact API specifications and prices from OpenAI's developer documentation, what OpenAI's benchmarks show and what independent testing found, where you can use the models, a working API quickstart, and how to choose between Sol, Luna and Astra. The numbers come from OpenAI's model reference pages and changelog, checked on launch day. Benchmark figures are attributed to whoever published them.

GPT-6 Sol vs GPT-5.6 Sol: clearing up the naming

OpenAI reused the tier names across generations, which makes searches for "Sol pricing" easy to get wrong. There are now two different models called Sol:

• GPT-5.6 Sol (gpt-5.6-sol) is the older flagship of the GPT-5.6 family, which also includes Terra and Luna. It is still in the API at $4/$20 per million tokens. That is itself a promotional price, which OpenAI's changelog says runs "at least through November 21, 2026". Its knowledge cutoff is 16 February 2026. Our GPT-5.6 Sol, Terra and Luna guide covers it.
• GPT-6 Sol (gpt-6-sol) is the new mid-tier GPT-6 model released today. It costs $2/$10 and has a knowledge cutoff of 20 April 2026. It sits below GPT-6 Astra, where GPT-5.6 Sol used to sit at the top of its own family.

The same applies to Luna: gpt-5.6-luna ($0.20/$1.20) and gpt-6-luna ($0.10/$0.50) are separate models. There is no GPT-6 Terra. OpenAI's GPT-6 family is Astra, Sol and Luna, and OpenAI has not announced a GPT-6 Terra. If a benchmark table or pricing page just says "Sol", check the model ID before you trust the number.

OpenAI and the press compare GPT-6 Sol with GPT-5.6 Sol, not GPT-5.6 Terra. The prices bear that out: GPT-6 Sol is exactly half of GPT-5.6 Sol's current rate. In practice, GPT-6 Sol is the upgrade path for teams running GPT-5.6 Sol who do not need to pay for Astra.

What are GPT-6 Sol and GPT-6 Luna?

Both are reasoning models with text and image input and text output. OpenAI's model selection guide puts them this way. Astra is for the hardest end-to-end work. Sol is for demanding reasoning, coding and agent tasks that need strong analysis but not maximum intelligence. Luna is for "efficient, repeatable work at scale", where latency and cost matter most. OpenAI's launch post, as quoted by TechCrunch, describes Luna's jobs as "summarizing documents, extracting information, or answering quick questions."

OpenAI's summary line: "GPT-6 Astra introduced a new generation of intelligence; these models extend its benefits by making that intelligence more efficient and accessible."

Full specifications

Source: OpenAI's GPT-6 Sol and GPT-6 Luna model reference pages.

Three details stand out:

• Both models keep the none reasoning effort. Astra dropped it. That means you can switch thinking off completely on Sol and Luna for latency-sensitive calls. On Astra, every call spends reasoning tokens.
• Luna has a newer knowledge cutoff than Sol (18 May against 20 April 2026) and than Astra (30 April). It is unusual for the cheapest model in a family to know the most recent facts, but that is what the docs say.
• The context window is the same as every other current OpenAI model. Like the GPT-5.6 family and Astra, both have 1.05M tokens with input capped at 922K. The generational change is in price and behaviour, not context.

How much does GPT-6 Sol cost?

GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens, with cached input at $0.20. OpenAI's API changelog for 22 September gives the same figures. A company spokesperson told VentureBeat that the pricing is permanent, not a launch promotion. OpenAI says better inference and caching let it cut prices while raising capability.

How much does GPT-6 Luna cost?

GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output tokens, with cached input at $0.01. The output price fell by more than the input price: from $1.20 to $0.50, a 58% cut, against 50% on input.

Full GPT-6 pricing table

The price cuts, model by model

One detail matters for anyone budgeting a migration. GPT-5.6 Sol's $4/$20 is a promotional rate announced on 21 August and guaranteed only "at least through November 21, 2026". Against GPT-5.6 Sol's pre-promotion list price, GPT-6 Sol's saving is bigger than 50%. If you are staying on GPT-5.6 Sol, plan for that price going back up at some point.

What it costs in practice

Take a workload of 100M input tokens and 20M output tokens a month, with no caching:

• GPT-6 Luna: $10 + $10 = $20 (GPT-5.6 Luna: $44)
• GPT-6 Sol: $200 + $200 = $400 (GPT-5.6 Sol at promo rates: $800)
• GPT-6 Astra: $1,000 + $1,000 = $2,000

With 80% of input tokens served from cache, the Sol bill falls to $40 + $16 + $200 = $256. Output tokens then make up most of the cost. Reasoning tokens are also billed as output, so the effort level you choose is now the biggest cost control you have.

The 272K long-context cliff applies to Sol and Luna exactly as it does to Astra: go over it by one token and the whole request is repriced. On Sol, a 270K-input / 10K-output request costs $0.64. A 275K-input / 10K-output request costs $1.25. Our GPT-6 Astra pricing breakdown explains the cliff in detail, and the same rule covers the whole GPT-6 family.

For comparison with Anthropic, VentureBeat notes that GPT-6 Sol matches Claude Sonnet 5's $2/$10 and costs half of Claude Opus 5.5's $4/$20. We compare those two directly in Claude Opus 5.5 vs GPT-6 Sol.

How do GPT-6 Sol and Luna perform on benchmarks?

There are two views here, and they point in slightly different directions.

What OpenAI reported

These are vendor-reported scores from OpenAI's launch post, as reproduced by VentureBeat and the Handy AI launch breakdown. The OpenAI post itself blocked automated fetching on launch day, so we cross-checked each number across at least two outlets.

OpenAI's own table shows GPT-6 Sol scoring below GPT-5.6 Sol on DeepSWE and OSWorld. OpenAI's argument is about price-performance, not raw capability: similar scores to much more expensive Anthropic models at a fraction of the cost per task, plus a large drop in factual errors and deception. 9to5Mac's summary of the launch says OpenAI claims GPT-6 Luna at max effort can beat GPT-5.6 Sol at medium effort for about a tenth of the cost.

What independent testing found

Artificial Analysis published results on launch day. Its summary: both models "push the cost efficiency frontier by halving cost relative to GPT-5.6 Sol and Luna", while Intelligence Index and Coding Agent Index scores "remain level with GPT-5.6, with progress in some evaluations and regressions in others." The detailed figures below come from Artificial Analysis's results as reported by OfficeChai:

On the same index, Claude Opus 5.5 leads at 58 and GPT-6 Astra scores 53, so Sol is clearly a step below the flagships. Artificial Analysis updates its index over time, so scores from this report should not be compared with figures from earlier write-ups, including numbers in our Astra launch coverage.

Taken together, the independent data supports three conclusions:

• Sol is about as capable as GPT-5.6 Sol and costs about half as much per task. It gains 1 point on intelligence and 2 on coding, and costs 47% less per task.
• Sol is much less likely to make things up. The hallucination rate fell from 92% to 60%. Part of that comes from Sol declining more questions (it attempted 83%, against 99% for GPT-5.6 Sol), so accuracy on the questions it answers went down slightly.
• Luna is the same intelligence for less than half the per-task cost, but it slipped slightly on coding. Its Coding Agent Index fell 2 points, and Luna uses more output tokens than its predecessor (about 51K against 41K on the index). Hacker News commenters picked up on the same coding regression within an hour of launch.

The GDPval regression (about 100 Elo on Sol) matters if your workload is professional document work such as spreadsheets, memos and slide decks. Test those tasks before switching.

Where can you use GPT-6 Sol and Luna?

TechCrunch and 9to5Mac report that the ChatGPT rollout is gradual over launch day, so the model picker may not show them straight away. OpenAI has not announced a deprecation date for GPT-5.6 Sol or Luna in the API. Its deprecations page lists none as of 22 September. In Codex, though, users on Hacker News report sessions being switched to the new models automatically. If you depend on GPT-5.6 behaviour in Codex, pin the model explicitly.

How do you call GPT-6 Sol and Luna from the API?

Use the Responses API. It is the path OpenAI recommends for reasoning models and it supports every hosted tool. This Python example follows the pattern in OpenAI's reasoning guide:

from openai import OpenAI

client = OpenAI()  # reads OPENAI_API_KEY

# GPT-6 Sol for a coding / agentic task
response = client.responses.create(
    model="gpt-6-sol",
    reasoning={"effort": "high"},
    input=[{"role": "user", "content": "Refactor this function to remove the N+1 query: ..."}],
)
print(response.output_text)

# GPT-6 Luna for high-volume extraction, thinking switched off
summary = client.responses.create(
    model="gpt-6-luna",
    reasoning={"effort": "none"},
    input=[{"role": "user", "content": "Extract invoice number, date and total as JSON: ..."}],
)
print(summary.output_text)

And in JavaScript:

import OpenAI from "openai";
const openai = new OpenAI();

const response = await openai.responses.create({
  model: "gpt-6-sol",
  reasoning: { effort: "medium" },
  input: [{ role: "user", content: "Write a migration plan for moving our queue to SQS." }],
});
console.log(response.output_text);

Migration gotchas from GPT-5.6

OpenAI's latest-model guide lists the changes that break existing code:

• Function calling in Chat Completions only works with reasoning_effort: "none" on Sol and Luna. With any other effort level, move tool-calling code to the Responses API.
• When reasoning is on, remove sampling parameters. Drop temperature, top_p and top_logprobs, and logprobs as well on Chat Completions. On Responses, remove message.output_text.logprobs from include.
• If you used minimal effort on older models, start with low and benchmark it.
• Prompt caching config changed. Replace prompt_cache_retention with prompt_cache_options: {"ttl": "30m"} (30m is the only value and the default). Caching starts at 1,024 visible input tokens. Use prompt_cache_key to keep per-customer cache accounting separate.

Because cached reads are 10% of the input price and cache writes are 125%, a stable system prompt and tool list at the start of every request pays for itself from the second call onwards.

GPT-6 Sol vs GPT-6 Astra: which should you use?

Astra costs five times as much as Sol per token, on both input and output. Here is what the extra money buys:

Computer use is where the gap is largest. On OpenAI's own OSWorld 2.0 numbers, Astra is about 12 points ahead, and 9to5Mac notes that OpenAI still recommends Astra for computer-use tasks. For code generation, refactoring and most agent loops that work through tools rather than a screen, Sol at high or xhigh effort covers most of the ground for a fifth of the price. A sensible pattern is to route to Sol by default and escalate to Astra when a task fails or is flagged as hard. For a deeper look at the flagship, see GPT-6 Astra vs GPT-5.6 Sol.

Which should you choose: Sol, Luna or Astra?

• Choose GPT-6 Luna for classification, extraction, summarisation, routing, RAG answer synthesis and any pipeline that makes millions of calls. At $0.10/$0.50 with none effort available, it competes with open-weight models on price. For a comparison with cheap Chinese alternatives, see our DeepSeek V4.1 Flash guide. Avoid it for hard multi-file coding, where its Coding Agent Index slipped.
• Choose GPT-6 Sol as your default for coding assistants, agent loops, analysis and customer-facing reasoning. If you are on GPT-5.6 Sol today, switching roughly halves your bill for similar capability and far fewer made-up answers. Run your own evals first if professional-document quality (the GDPval regression) matters to you.
• Choose GPT-6 Astra for computer use, browsing agents, long unattended runs where a failure costs more than tokens, and the hardest reasoning problems.
• Stay on GPT-5.6 Sol for now only if your evals show a real regression on your workload. Remember that its $4/$20 price is a promotion with an end date.

If you are weighing OpenAI against Anthropic's release from the same day, our Claude Opus 5.5 vs GPT-6 Astra vs Grok 4.7 comparison covers the flagship tier, and the Claude Opus 5.5 guide covers Anthropic's pricing and specs.

What is still unconfirmed?

• Whether GPT-6 Terra will ever ship. OpenAI has announced none, and the GPT-6 family currently has three tiers.
• API deprecation dates for GPT-5.6 Sol and Luna. None are published.
• Availability on Amazon Bedrock and Microsoft Foundry. It was not confirmed on launch day.
• Output speed in tokens per second. Artificial Analysis had not published a speed measurement for GPT-6 Sol at the time of writing.

FAQ

Is GPT-6 Sol the same as GPT-5.6 Sol?

No. GPT-5.6 Sol (gpt-5.6-sol) is the older flagship of the GPT-5.6 family at $4/$20 per million tokens. GPT-6 Sol (gpt-6-sol) is the new mid-tier model below GPT-6 Astra, released 22 September 2026 at $2/$10, with a newer knowledge cutoff (20 April 2026).

How much does GPT-6 Sol cost per million tokens?

$2 for input, $0.20 for cached input and $10 for output. Prompts over 272K input tokens cost 2× on input and 1.5× on output for the whole request. Batch and Flex are 50% off.

How much does GPT-6 Luna cost?

$0.10 per million input tokens, $0.01 cached and $0.50 output. That is half of GPT-5.6 Luna's input price and 58% less on output.

Is GPT-6 Sol better than GPT-6 Astra?

No. Astra is the more capable model, especially for computer use: it scores 72.6% against Sol's 60.5% on OSWorld 2.0 in OpenAI's numbers. Sol costs one fifth as much and handles most coding and agent work, so it is the better default for cost-sensitive teams.

Can free ChatGPT users use GPT-6 Sol or Luna?

Free and Go users get GPT-6 Luna in the ChatGPT desktop app. GPT-6 Sol is limited to paid plans (Plus, Pro, Business, Enterprise and Edu), and Codex access covers most paid accounts.

What is the context window of GPT-6 Sol and Luna?

Both have a 1,050,000-token context window with a 922,000-token input cap and up to 128,000 output tokens. The figures are the same as GPT-6 Astra and the GPT-5.6 family.

Sources

• OpenAI: GPT-6 Sol model reference, GPT-6 Luna model reference, API changelog, latest-model guide, prompt caching guide, launch post
• VentureBeat, TechCrunch, 9to5Mac
• Independent: Artificial Analysis, via OfficeChai

If your team is building products on GPT-6 Sol or Luna and needs more engineering capacity, Codersera can help you hire vetted remote developers who already ship with these models.