GPT-6 Sol vs Astra vs Luna: Which Model to Use

The names are Latin: luna is moon, sol is sun, astra is stars. The order runs small to large, and so does the price. OpenAI's own positioning is clear about who each tier is for:

• GPT-6 Luna: summarising, extraction and straightforward questions. High volume, low cost. Also the model Free and Go users get in the ChatGPT desktop app.
• GPT-6 Sol: complex everyday work that developers and knowledge workers repeat. Coding tasks, analysis, drafting, routine agent workflows.
• GPT-6 Astra: the most complex, multi-faceted projects. Long agentic runs, hard research, computer use across many applications.

So "chat gpt sol meaning" has a practical answer: Sol is the middle tier, and the one most people should use most of the time.

This is where the confusion comes from. GPT-5.6 shipped on 9 July 2026 as three models: Sol, Terra and Luna (terra is Latin for earth). GPT-6 has reused two of those names and added a new top tier:

• GPT-5.6 (9 July 2026): Sol, Terra, Luna.
• GPT-6 (3-4 September 2026): Astra.
• GPT-6 (22 September 2026): Sol and Luna.

As of 23 September 2026, OpenAI has not announced a GPT-6 Terra. If you are asked to choose between "Sol, Terra or Luna", you are looking at the GPT-5.6 family, or at a guide written before GPT-6 Sol shipped. We do not compare GPT-5.6 Terra here, because OpenAI has not published GPT-6-era benchmarks for it.

The same name does not mean the same model. "Sol" in September is GPT-6 Sol, not GPT-5.6 Sol. Check the full model ID in the API (for example gpt-6-astra) rather than the short name.

Checked 23 September 2026. Source: OpenAI launch pricing as reported by VentureBeat, Vellum and DataCamp. Standard tier. Astra cache writes cost $12.50 per million; Astra fast mode costs 2x for up to 2.5x speed.

The ladder is clean. Sol costs exactly 20 times Luna on every rate. Astra costs 5 times Sol and 100 times Luna. In sterling, Astra output is approx £39.50 per million tokens against approx £7.90 for Sol and approx £0.40 for Luna.

GPT-6 is also a straight price cut on GPT-5.6. GPT-5.6 Sol is $4 input / $0.50 cached / $20 output on its current promotional rate (standard $5 / $30); GPT-6 Sol is $2 / $0.20 / $10, so input and output halve and cached input falls 60%. GPT-5.6 Luna was $0.20 / $0.02 / $1.20; GPT-6 Luna halves input and cached input and cuts output by about 58%. If you are still on GPT-5.6 Sol, moving to GPT-6 Sol is cheaper before you look at a single benchmark.

OpenAI has not published one table with all three models on every test. Only one benchmark, OSWorld 2.0 in OpenAI's offline harness, includes Astra, Sol and Luna together. The rest cover two of the three. Effort settings also vary, and that matters.

Source: OpenAI's GPT-6 Sol and Luna launch, 22 September 2026, via VentureBeat and Vellum. A dash means OpenAI did not report that model on that benchmark.

AutomationBench needs a health warning. OpenAI's chart shows Sol (33.2%) beating Astra (30.3%), but Astra was run at low effort and Sol at xhigh. That is not a like-for-like result. Anthropic's Opus 5.5 launch table lists Astra at 41.4% on AutomationBench, as reported by OpenAI, which suggests Astra at higher effort sits well above Sol. The fair reading: Sol at full effort beats Astra at minimum effort, at $0.27 per task.

Agents' Last Exam shows a 2.9-point gap between Astra and Sol (59.3% vs 56.4%). For a model that costs a fifth as much, that is a small gap. DeepSWE shows Luna only 2.2 points behind Sol (66.6% vs 68.8%) on coding, with both run at max effort.

On the independent Artificial Analysis Intelligence Index, Astra scores 53 and Sol (max) 48. Luna has no score in our sources. Artificial Analysis also measured Sol at 131.2 output tokens per second. For the wider picture, see our September 2026 frontier benchmarks roundup.

This is the chart that best explains the tiers. Luna, Sol and GPT-5.6 Sol sit within 3.5 points of each other. Astra is in a different place. If your work looks like computer use (navigating applications, filling forms, driving a desktop), Astra's premium buys a real improvement. Note that OpenAI compared against Claude Opus 5 here, not Opus 5.5, which launched 90 minutes before Sol.

Two further OpenAI claims favour Sol. It makes "about half as many mistakes" as GPT-5.6 Sol on factuality, "reaching Astra-level reliability at much lower cost". And on an adversarial coding deception test, Sol deceived 1.3% of the time (down from 10.4%) and Luna 2.8% (down from 9.5%).

Per-token prices hide how fast the gap grows. We modelled a bulk extraction job: 100 million input tokens and 10 million output tokens a month, no caching, at standard list prices. This is exactly the work OpenAI positions Luna for.

Luna does the job for $15 a month (approx £11.85). Sol costs $300 (approx £237). Astra costs $1,500 (approx £1,185). On a chatbot workload (10M input at 80% cache hits, 2M output) the figures are $1.28, $25.60 and $128. On a coding agent (50M input at 90% cache hits, 5M output) they are $3.45, $69 and $345.

One caveat. This is per-token list cost. Models use different numbers of tokens for the same task, especially at higher reasoning effort, so real cost per task can differ. Test on your own prompts before committing. Our AI API pricing comparison runs the same model across every major vendor.

• Free and Go: GPT-6 Luna, in the desktop app.
• Plus, Pro, Business and Enterprise: GPT-6 Astra is available.
• Pro, Business and Enterprise: also get GPT-6 Astra Pro.
• ChatGPT Work and Codex: Sol and Luna are both available, as they are in the API.

OpenAI also loaded a banked usage reset into Plus, Pro and Business accounts on 22 September. For developers, Astra is also on Amazon Bedrock. Context windows differ: Astra supports up to 1M tokens, and Artificial Analysis lists Sol at about 872K. We have no confirmed Luna figure.

Astra is the only GPT-6 model rated "Critical" for cybersecurity under OpenAI's Preparedness Framework. In practice it refuses to write proof-of-concept exploits, and expanded security access goes through OpenAI's "Daybreak" programme. The UK AI Security Institute also found Astra could evade monitoring under adversarial prompting. Our GPT-6 Astra review covers the detail.

For security teams, the most capable tier is therefore the most restricted. That is a reason to test Sol first on legitimate security work, not a reason to pick Astra.

1. Is the task simple and repeated at volume? Summaries, classification, extraction, tagging, short answers. Use Luna. At 20x cheaper than Sol, it is worth testing even if you expect to need Sol.
2. Is it real work a person would do, repeated often? Code changes, reports, analysis, routine agents. Use Sol. This covers most professional use.
3. Does it drive a computer or span many tools and steps? Test Sol first, then Astra. Astra's 12-point OSWorld lead is the clearest case for paying 5x.
4. Has Sol already failed on it? Long, multi-faceted research or engineering projects where errors compound. Use Astra.
5. Still on GPT-5.6 Sol? Move to GPT-6 Sol. It is half the price and scored higher on every shared test OpenAI published.

Do not use Astra for:

• High-volume, simple jobs. It costs 100x Luna. On our extraction model that is $1,500 against $15 a month.
• Everyday chat and drafting. Sol is five times cheaper, and OpenAI itself says Sol reaches Astra-level factual reliability.
• Offensive security research. Astra refuses proof-of-concept exploits under its Critical cyber rating.
• Unattended runs where monitoring matters. The UK AISI finding on monitoring evasion is a reason to keep a human in the loop.

Do not use Luna for: long agentic tasks. A 2-point gap per task grows over dozens of steps. It also lacks published scores on Agents' Last Exam and AutomationBench, so its ceiling on complex work is unproven.

Where Sol loses: computer use (12.1 points behind Astra on OSWorld 2.0 offline), and against Claude Opus 5.5 on AutomationBench, where Anthropic's table shows Opus 5.5 at 40.0% against Sol's 33.2%.

Anthropic's range maps loosely onto the same three tiers. Here are list prices only; benchmark comparisons are in our dedicated head-to-heads.

Checked 23 September 2026. Per million tokens, standard tier. Sources: OpenAI via VentureBeat, Vellum and DataCamp; Anthropic pricing documentation.

Three things stand out. Luna costs a tenth of Haiku 4.5. Sol and Claude Sonnet 5 have identical list prices, including $0.20 cached input. And Anthropic's flagship is cheaper than OpenAI's: Opus 5.5 costs 40% of Astra per token. See GPT-6 Sol vs Claude Opus 5.5 and Claude Opus 5.5 vs GPT-6 Astra for the benchmarks, or Claude API pricing for every Anthropic rate. Our GPT-6 Sol and Luna launch coverage has the full release detail.

• OpenAI: Introducing GPT-6 Sol and Luna
• OpenAI: GPT-6 Astra
• VentureBeat: OpenAI releases GPT-6 Sol and Luna
• TechCrunch: OpenAI launches GPT-6 Sol and Luna
• Vellum: GPT-6 Sol and Luna benchmarks explained
• DataCamp: GPT-6 Astra
• Artificial Analysis: GPT-6 Sol
• Anthropic: Claude Opus 5.5
• Anthropic: Claude pricing documentation