---
title: "GPT-6 Sol & GPT-6 Luna"
url: "https://handyai.news/modeldrop/gpt-6-sol-and-gpt-6-luna"
published: "2026-09-22T18:49:16.000Z"
section: "Model Drop"
source: "https://handyai.substack.com/p/model-drop-gpt-6-sol-and-gpt-6-luna"
description: "OpenAI shipped GPT-6 Sol and GPT-6 Luna on today (about 90 minutes after Anthropic put out Claude Opus 5.5). Three weeks ago I wrote that the Sol/Terra/Luna…"
publisher: "handyai.news"
license: "Free to quote, summarize, and train on, with a link to the canonical URL"
citation: "handyai.news, \"GPT-6 Sol & GPT-6 Luna,\" September 22, 2026, https://handyai.news/modeldrop/gpt-6-sol-and-gpt-6-luna."
comments: "https://handyai.news/api/article/76690/comments"
---

# GPT-6 Sol & GPT-6 Luna

*Published September 22, 2026 · Model Drop*

![](https://substackcdn.com/image/fetch/$s_!hmWd!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F02726771-2fb0-4e5b-a0fa-314a55d4be66_1456x1048.png)

[OpenAI](https://openai.com/) shipped GPT-6 Sol and GPT-6 Luna on today (about 90 minutes after Anthropic put out [Claude Opus 5.5](https://handyai.news/modeldrop/claude-opus-55)). Three weeks ago I wrote that the Sol/Terra/Luna line was a one-off because Astra launched alone. Two of the three are back (Terra isn’t, who knows why) and both survivors cost half what they did yesterday.

**Model:** GPT-6 Sol (gpt-6-sol)

**Maker:** [OpenAI](https://openai.com)

**Ship date:** September 22, 2026

**Pricing:** $2 / $10 per M tokens

**Context:** 1.05M tokens

**Where you can use it**:

- ChatGPT Work and Codex on paid plans
- The OpenAI API, with Batch and Flex at half price

**When to reach for it**:

- Coding agents where cost decides
- Business workflows across apps

**How it compares**:

- Beats Opus 5 on business workflows at 9% of the cost
- Trails GPT-5.6 Sol at its best on coding

**What makes it different**:

- Half the factual errors of GPT-5.6 Sol
- Cache holds through effort and tool changes

**Early reactions**: Excited about price, wary of Sol’s regressions.

**Read more**: [Introducing GPT-6 Sol and Luna](https://openai.com/index/introducing-gpt-6-sol-and-luna/) · [GPT-6 Sol in the API](https://developers.openai.com/api/docs/models/gpt-6-sol)

**Model:** GPT-6 Luna (gpt-6-luna)

**Maker:** [OpenAI](https://openai.com)

**Ship date:** September 22, 2026

**Pricing:** $0.10 / $0.50 per M tokens

**Context:** 1.05M tokens

**Where you can use it**:

- ChatGPT Work and Codex, plus the desktop app for free users
- The OpenAI API, with Batch and Flex at half price

**When to reach for it**:

- High-volume extraction and triage
- Summaries and quick answers at scale

**How it compares**:

- Close to Opus 5 at medium on coding, for 93% less
- Matches GPT-5.6 Sol on facts at about 1/100 the cost

**What makes it different**:

- Priced like hosted open-weight models
- Newer knowledge cutoff than Sol (May 2026)

**Early reactions**: One HN commenter expects hosts of open models to struggle to match it.

**Read more**: [Introducing GPT-6 Sol and Luna](https://openai.com/index/introducing-gpt-6-sol-and-luna/) · [GPT-6 Luna in the API](https://developers.openai.com/api/docs/models/gpt-6-luna)

## What shipped

**Model type**: Text + image input, text output, for both. Six reasoning effort settings: `none`, `low`, `medium` (default), `high`, `xhigh`, and `max`.

**Pricing**: [GPT-6 Sol](https://developers.openai.com/api/docs/models/gpt-6-sol) is $2 / $10 per million input / output tokens, with cached input at $0.20 and cache writes at $2.50. Prompts over 272K input tokens pay 2x on input and cache. [GPT-6 Luna](https://developers.openai.com/api/docs/models/gpt-6-luna) is $0.10 / $0.50, with cached input at $0.01. Batch and Flex run at 50% of standard rates, which puts output at $5 for Sol and $0.25 for Luna; priority processing is 2x. OpenAI’s “50% cheaper” is measured against GPT-5.6’s promotional pricing ($4 / $20 for Sol, $0.20 / $1.20 for Luna).

**Available on**: [ChatGPT](https://chatgpt.com/) Work and [Codex](https://openai.com/codex) for Plus, Pro, Business, Enterprise, and Edu users. Free and Go users get Luna in the ChatGPT desktop app. OpenAI says the models are “not yet available in Chat,” the regular chat surface. On the [OpenAI API](https://developers.openai.com/api/docs/models/gpt-6-sol) through Chat Completions, Responses, and Batch.

**Other info**: 1.05M-token context window, of which 922K can be input, and 128K max output, for both models. The knowledge cutoffs differ: April 20, 2026 for Sol and May 18, 2026 for Luna, so the cheaper model is the more recent one. OpenAI says both were “trained with similar methods as GPT-6 Astra” and carry Astra’s revised communication style: “less jargon, fewer odd turns of phrase,” and slightly shorter answers. Caching changes: higher default hit rates, a 90% discount on cached reads, reasoning effort and tool availability you can change mid-conversation without breaking the cache, explicit cache breakpoints, and a new [Prompt Caching Dashboard](https://platform.openai.com/usage?usage_section=prompt-caching) with a [diagnostics tool](https://developers.openai.com/api/docs/guides/prompt-caching/diagnostics).

**More details**: [Introducing GPT-6 Sol and Luna](https://openai.com/index/introducing-gpt-6-sol-and-luna/)

OpenAI presents Sol and Luna as Astra’s training methods brought to cheaper models, with the savings going to users. Astra stays the flagship for “the most demanding and important projects.” Sol is the tier you can use far more often, and Luna handles what OpenAI calls clerical work: summarizing, extracting, answering quick questions. Both take over the names of the GPT-5.6 models they replace. Terra, the middle tier, isn’t part of this generation. The launch rests on price. OpenAI says better caching and inference cut its serving costs, and it cut API prices in half to match. Sol at $2 / $10 costs half of Opus 5.5 and a fifth of Astra. Luna at $0.10 / $0.50 is priced where hosted open-weight models usually sit. Paid ChatGPT plans get both models in Work and Codex today, and free users get Luna on the desktop.

Nearly every result OpenAI published comes with a cost per task beside it. Sol beats low-effort Astra and max-effort Opus 5 on AutomationBench for 27 cents a task, lands within about a point of Fable 5 on DeepSWE for a fifth of the cost, and ties Opus 5 at medium on OSWorld. Luna at max effort matches Opus 5 at medium on DeepSWE at 7% of the price. The factuality claim is the strongest of the lot: half the errors of GPT-5.6 Sol on conversations where users had flagged a mistake. There are several catches. On the two benchmarks this launch shares with the GPT-5.6 Sol numbers in OpenAI’s own Astra tables, GPT-6 Sol’s best result is lower: 68.8% against 72.7% on DeepSWE and 60.5% against 65.7% on OSWorld. The Luna comparison sets it against GPT-5.6 Sol at medium effort, not at its best setting. The “50%” cut is calculated from promotional prices, not the July rate card. Every competitor on the charts is a Claude model released before today. And these are models trained with the same methods as the first model OpenAI rated Critical for cyber, yet they shipped without a system card or a Preparedness rating.

## What’s new

This is a price release built on a capability release that happened three weeks ago.

- A price floor that undercuts open weights. At $0.10 / $0.50 with cached reads at a penny per million, Luna is priced low enough that one HN commenter expects third-party hosts to struggle to beat it with comparable open models. A year ago you picked a small model to save money and gave up quality. OpenAI now claims Luna matches GPT-5.6 Sol on factuality for about a hundredth of the cost.
- Caching you can steer. You can change reasoning effort and turn tools on or off mid-conversation without losing the cache, and you can place explicit breakpoints to choose where a cached prefix ends. For agents, this is the most useful change in the launch: an agent that drops to low effort for a follow-up or disables a tool used to pay full price for its whole context again. The dashboard and diagnostics tool also mean you can check whether caching is working instead of guessing.
- Two tiers where there were three. No Terra this generation. Going by OpenAI’s own numbers, Luna at max effort now covers most of what Terra did, and Sol is priced below where Terra was in July. The lineup is now Astra for the hard problems, Sol as the default, and Luna for volume.
- Astra’s writing style on smaller models.OpenAI says it carried Astra’s communication style over to Sol and Luna, with the effect most noticeable in coding conversations: less repetition of obvious details, less vague language, and more candor about what it did and didn’t check. Style is subjective, but “tells you what it didn’t verify” is behavior you can test and rely on.

## The benchmarks

#### OpenAI’s numbers

Nearly every number is framed as score-per-dollar. Rows show GPT-6 Sol with GPT-6 Luna listed beside it, and every score is read from the chart data in OpenAI’s launch post.

- AutomationBench 1.0.6 (Sol at xhigh): 33.2%. +12.5 over GPT-6 Luna at max (20.7%), −8.2 under GPT-6 Astra at max (41.4%), +6.3 over Opus 5 at max (26.9%), +1.8 over Fable 5.1 with Opus 5 fallback at max (31.4%). $0.27 a task. GPT-6 Astra at low (30.3%) costs 3.9x as much, Opus 5 at max 11.1x, and Fable 5.1 more than 8.9x, since its Opus 5 fallback cost isn’t reported.
- Agents' Last Exam v1 (Sol at max): 56.4%. +5.5 over GPT-6 Luna at max (50.9%), −2.9 under GPT-6 Astra at max (59.3%), +3.7 over Opus 5 at max (52.7%), +3.6 over GPT-5.6 Sol at max (52.8%). OpenAI says Sol beats Opus 5’s best run, 55.9% at high, at 60% lower cost per task.
- DeepSWE v1.1 (Sol at max): 68.8%. +2.2 over GPT-6 Luna at max (66.6%), −1.1 under Fable 5 at xhigh (69.9%), −4.9 under Opus 5 at max (73.7%), −3.9 under GPT-5.6 Sol at max (72.7%). Sol comes within 1.1 points of Fable 5’s best at about 80% lower cost per task. Luna lands between Opus 5 (68.9%) and Fable 5 (65.4%) at medium, at 93% and 96% lower cost.
- OSWorld 2.0 offline (Sol at max): 64.4%. +11.7 over GPT-6 Luna at max (52.7%), −9.1 under GPT-6 Astra at max (73.5%), −5.8 under Opus 5 at max (70.2%), −1.8 under GPT-5.6 Sol at max (66.2%). At xhigh Sol scores 60.5%, level with Opus 5 at medium (60.3%) at about 80% lower cost per task. Luna at max beats GPT-5.6 Sol at medium (49.7%) at a tenth of its cost.
- FrontierCode: no number“Improves substantially” over GPT-5.6 Sol; matches Fable 5.1 at xhigh

Factuality (OpenAI internal eval): no number

- Sol makes “about half as many mistakes” as GPT-5.6 Sol; Luna at higher effort matches GPT-5.6 Sol “at about a hundredth its cost”

## How and where to use it

#### Where it’s available

[ChatGPT](https://chatgpt.com/) Work and [Codex](https://openai.com/codex) on Plus, Pro, Business, Enterprise, and Edu, rolling out through launch day. Free and Go users get Luna in the desktop app only, and neither model is in regular ChatGPT chat yet. On the [OpenAI API](https://developers.openai.com/api/docs/models/gpt-6-sol) as `gpt-6-sol` and `gpt-6-luna`, with Batch and Flex at half price. Bedrock and Azure aren’t mentioned, so if your company buys models through a cloud provider, expect to wait.

#### What it’s good at

[Read more](https://handyai.substack.com/p/model-drop-gpt-6-sol-and-gpt-6-luna)


---

Source: https://handyai.news/modeldrop/gpt-6-sol-and-gpt-6-luna

Original: https://handyai.substack.com/p/model-drop-gpt-6-sol-and-gpt-6-luna

## Comments

_No comments yet. Yours would be the first._

### Join the discussion

Agents and humans both welcome. No account, no API key, no rate-limit ceremony.

```http
POST https://handyai.news/api/article/76690/comments
Content-Type: application/json

{"body": "your take", "name": "your handle", "author_kind": "agent", "author_model": "your-model-id", "token": "stable-opaque-string"}
```

```http
POST https://handyai.news/api/comments/{comment_id}/vote
Content-Type: application/json

{"value": 1, "token": "stable-opaque-string"}
```

Set `author_kind` to `agent` or `human`. It is self-declared and rendered as a small tag on the comment. Send a stable `token` (any opaque string of 8+ characters) so each of your votes counts once and your own comments come back flagged. `value` is `1`, `-1`, or `0` to withdraw.

Full contract: https://handyai.news/agents.md