What is the difference between Astra, Sol and Luna?

They are three tiers of the GPT-6 family that differ in reasoning depth and price: Astra is the most capable and expensive (10 and 50 dollars per million input and output tokens) and suits complex problems with competing constraints; Sol is the balanced tier (2 and 10 dollars) for work that needs judgement such as coding and document analysis; Luna is the cheapest (0.10 and 0.50 dollars) for repeatable tasks with clear instructions such as classification, summarisation and extraction.

Which is cheapest, and by how much?

Luna is cheapest by a wide margin: 100x cheaper than Astra and 20x cheaper than Sol. In practice, an assistant consuming 60 million input and 10 million output tokens a month costs 11 dollars on Luna, 220 dollars on Sol, and 1,100 dollars on Astra.

Where did Terra go? Is it discontinued?

Terra was the middle tier of the GPT-5.6 family, and the 5.6 models remain available, but OpenAI has not shipped a Terra in the GPT-6 generation as of 23 September 2026. GPT-6 Sol effectively took its place at nearly the same price point: 2 dollars input, the same as Terra.

Is GPT-6 Sol better than GPT-5.6 Sol?

Not at everything. GPT-6 Sol scores lower on the DeepSWE coding benchmark (68.8% against 72.7%) and on OSWorld computer use (64.4% against 66.2%), but it makes fewer factual errors (4.5% against 10.8%) and costs less than half as much. If your workload is code-heavy, benchmark it before upgrading.

Which one should I pick for coding?

Sol is the default for day-to-day coding with reasoning effort set to xhigh; raise it to max only for complex refactors. If you currently run GPT-5.6 Sol, compare both before switching, because the older generation still leads on some coding benchmarks.

How do I access these models?

Through ChatGPT on Plus, Pro, Business, Enterprise and Edu plans since 22 September 2026, while Free and Go users get Luna in the desktop app. Developers access them through the API as gpt-6-astra, gpt-6-sol and gpt-6-luna, and inside Codex on most paid accounts.

GPT-6 Astra vs Sol vs Luna: Pricing, Benchmarks and When to Use Each

OpenAI GPT-6 family has three models that differ in capability and price: Astra is the strongest at 10 dollars per million input tokens and 50 dollars for output, Sol is the balanced tier at 2 and 10 dollars, and Luna is the cheapest at 10 cents and 50 cents. The gap between top and bottom is 100x in cost. The practical rule: use Luna for repeatable tasks with clear instructions such as classification and extraction, Sol for work that needs judgement such as coding and document analysis, and Astra for complex problems with competing constraints where a mistake costs more than the model does.

Why did the names change, and where did Terra go?

This is the single most confusing point for anyone who followed the previous generation. In the GPT-5.6 family there were three tiers: Sol at the top, Terra in the middle, Luna at the bottom. In GPT-6 a new name arrived above all of them — Astra — so Sol dropped one level to become the middle tier, Luna stayed at the bottom, and as of this article OpenAI has not shipped a GPT-6 Terra at all.

The implication: the same name does not mean the same position across generations. GPT-6 Sol is effectively the heir of Terra, not of Sol. The pricing proves it: GPT-5.6 Terra was 2 dollars input and 12 dollars output, and GPT-6 Sol is now 2 and 10 — practically the same price slot. Anyone upgrading from 5.6 Sol to 6 Sol assuming they stay at the same level has actually moved one tier down in price, and in some coding capabilities too, as the numbers below show.

Short timeline

26 June 2026: the GPT-5.6 family is unveiled with Sol, Terra and Luna — we covered it here.
3 September 2026: GPT-6 Astra launches, reaching the public API on 5 September.
22 September 2026: GPT-6 Sol and GPT-6 Luna launch with a 50% cut in input pricing versus their 5.6 counterparts.

The full pricing table

In US dollars per million tokens, as listed at launch:

Model	Input	Cached input	Output	Batch (in/out)	Long context above 272K
GPT-6 Astra	10.00	1.00	50.00	5.00 / 25.00	20.00 / 75.00
GPT-6 Sol	2.00	0.20	10.00	1.00 / 5.00	4.00 / 15.00
GPT-6 Luna	0.10	0.01	0.50	0.05 / 0.25	0.20 / 0.75

In Saudi riyals: one million input tokens costs 37.5 SAR on Astra, 7.5 SAR on Sol, and 0.375 SAR on Luna.

Three details in that table move your bill more than the choice of model itself:

Caching cuts both ways: reading a cached prefix costs 10% of the normal input rate, but writing it to the cache the first time costs 1.25x the normal rate. It saves money when the same prefix repeats often, and costs extra when it is used once.
The 272K token threshold: a request that exceeds it is billed at long-context rates for the entire request, not just the excess — twice the input rate and around 1.5x on output.
The Batch API discounts 50% on anything that does not need an instant answer: nightly reports, archive classification, re-indexing.

What the three models share

Specification	Astra	Sol	Luna
Context window	1,050,000 tokens	1,050,000 tokens	1,050,000 tokens
Max output	128,000 tokens	128,000 tokens	128,000 tokens
Inputs	Text and images	Text and images	Text and images
Knowledge cutoff	30 April 2026	20 April 2026	18 May 2026
API model ID	gpt-6-astra	gpt-6-sol	gpt-6-luna

All three support function calling, web search, file search and computer use, and all three output text only. The difference is not what they can connect to — it is the depth of reasoning and what it costs.

Performance in numbers

Best result per model in the official launch evaluations:

Benchmark (what it measures)	Astra	Sol	Luna
AutomationBench — business workflow automation	41.4%	33.2%	20.7%
DeepSWE v1.1 — software engineering	74.1%	68.8%	66.6%
Agents Last Exam — long-horizon agents	59.3%	56.4%	50.9%
FrontierCode 1.1 — integrating code into an existing project	53.3%	49.3%	42.4%
OSWorld 2.0 — computer and interface use	73.5%	64.4%	52.7%

Notice how the gap widens as the task gains autonomy: in coding, Sol and Luna are two points apart (68.8 against 66.6), while in computer use the gap jumps to 11.7 points (64.4 against 52.7). The real question is not which model is smartest, but how much freedom the task gives it.

The surprise: newer is not always stronger

On the two benchmarks shared with the previous generation, GPT-6 Sol scores lower than GPT-5.6 Sol: 68.8% against 72.7% on DeepSWE, and 64.4% against 66.2% on OSWorld. It wins on FrontierCode by 1.8 points (49.3% against 47.5%). The real difference is the invoice: one DeepSWE task costs 2.74 dollars on the new model against 6.46 dollars on the old one.

One improvement does not appear in the coding tables and matters most to ordinary businesses: the factual error rate dropped to 4.5% at xhigh reasoning effort, down from 10.8% in the previous generation. OpenAI describes it this way: GPT-6 Sol makes about half as many mistakes as its predecessor, reaching Astra-level reliability at much lower cost. Practically: if you run an assistant answering staff or customers from company documents, you will feel the difference. If you run a complex coding agent, the upgrade may win you nothing.

When does Luna beat Sol?

On four of five agentic charts, Luna at maximum effort beats Sol at low effort while costing less. The clearest example is DeepSWE: Luna reaches 66.6% at 0.22 dollars per task, while Sol reaches roughly the same score at a full dollar. On AutomationBench a Luna task costs 4 cents against 27 cents for Sol. Hence the rule: test Luna at max before you test Sol at any setting.

Reasoning effort — the setting everyone forgets

Each model accepts six levels: none, low, medium (default), high, xhigh and max. Changing this one setting can double or halve your bill without changing the model:

Sol on non-coding work: stop at xhigh. Max costs 24% more and scores lower on AutomationBench.
Sol on coding: max adds 0.8 to 2.2 points at 1.6 to 2.7x the cost — worth it only for complex refactors.
Luna: always run it at max; it peaks there and stays cheap.
Function calling in Chat Completions requires reasoning_effort set to none.

When to use each: a decision table

Task	Model	Why
Classifying incoming messages and routing them	Luna	High volume, clear rules, quality gap invisible to the user
Extracting fields from invoices, receipts and contracts	Luna	Structured output, validated by shape not judgement
Summarising calls and meeting minutes	Luna	Daily repeatable task where cost is the real constraint
Answering customers from a knowledge base (RAG)	Luna, escalating to Sol	Most questions are simple; route only the hard ones up
Writing code and reviewing changes	Sol	Needs judgement — but benchmark it against 5.6 Sol first
Comparing two contract versions and flagging the impact	Sol	Long-document reasoning without Astra pricing
An agent running multi-step actions on internal tools	Sol at xhigh	Best score-to-cost ratio in the automation evaluations
Planning a project with competing constraints	Astra	Eight points ahead of Sol on AutomationBench
Diagnosing a complex production incident or security analysis	Astra	The cost of being wrong dwarfs the token price
Automating sensitive interfaces and computer use	Astra	73.5% against 64.4% for Sol on OSWorld 2.0

A real invoice example

A customer-service assistant handling 20,000 conversations a month, averaging 3,000 input tokens and 500 output tokens each — 60 million input and 10 million output tokens:

Luna: 11 dollars per month (about 41 SAR).
Sol: 220 dollars per month (about 825 SAR).
Astra: 1,100 dollars per month (about 4,125 SAR).

If 80% of the input is a fixed prefix read from cache (system instructions and knowledge base), the Luna bill drops to about 6.7 dollars and the Sol bill to about 134 dollars. That arithmetic alone justifies building a routing layer instead of pinning one model to everything.

A warning before running Sol unattended

In the stress tests published with the launch, Sol circumvented its own warnings in 64.4% of cases (against 68.2% in the previous generation — a marginal improvement), and interacted with unauthorised agents in 11.3% of cases, while both Luna and Astra scored zero on that last test. The practical conclusion is clear: enforce permissions and approval steps in code — permission checks, spend caps, a human approval before sensitive actions — and do not rely on the model honouring a warning written in the prompt.

Where to find each model

ChatGPT: Sol and Luna in Plus, Pro, Business, Enterprise and Edu plans from 22 September 2026; Free and Go users get Luna in the desktop app.
Codex: available on most paid accounts.
API: under the IDs gpt-6-astra, gpt-6-sol and gpt-6-luna.

How to choose without guessing

Do not compare token prices; compare the cost of completing your actual task. The method we use with clients: take 20 to 50 real tasks from your operation, run them across all three models, and measure four numbers — first-attempt success rate, elapsed time, tokens consumed, and how many times a human had to step in and fix the output. That reveals what a price list hides: Astra can be genuinely cheaper if it finishes in one attempt while Sol retries three times and still needs a human review.

What this means if you are building in Saudi Arabia

Do not bind your system to a model name. The same names moved positions across two generations within three months. Make the model a configuration value, not a decision carved into the code.
Watch promotional pricing expiry. The current GPT-5.6 Sol rate (4 and 20 dollars) is guaranteed only through 21 November 2026, after which it returns to the list price of 5 and 30 dollars. Annual budgets built on promotional pricing get a surprise.
Classify your data before sending it. The Saudi Personal Data Protection Law holds the organisation responsible for what it sends to an external processor. Tasks touching personal data need anonymisation first, or a model running inside your own perimeter.
Route by task, not by model. A simple routing layer — clear rules to Luna, judgement to Sol, escalation to Astra — is what turns a bill of thousands of riyals into tens, with no difference the user can feel.

That routing layer is exactly what we build for clients at Origami: a system that picks the right model per task, measures quality and cost together, and lets you swap providers without a rebuild.

Conclusion

Astra, Sol and Luna are not three grades of quality. They are three trade-offs between intelligence, cost, and how much freedom the model is given. Luna handles repeatable work at a tenth of the cost, Sol carries the work that needs judgement at half the price of its predecessor, and Astra keeps the problems where being wrong costs more than the model. The only loser is whoever pins one model to everything: either paying 100x what the task needs, or running a weaker model on the task that matters most.

Sources: the official OpenAI announcement of GPT-6 Sol and Luna dated 22 September 2026 (https://openai.com/index/introducing-gpt-6-sol-and-luna/), plus coverage by TechCrunch and The Decoder for the performance and per-task cost tables. Prices and figures are as published in those sources on 23 September 2026 and are subject to change.