OpenAI's Three-Tier Model Lineup Explained: How to Choose Between Astra, Sol, and Luna

OpenAI has split its next-generation models into three tiers — GPT-6 Astra, Sol, and Luna — with pricing ranging from flagship to entry-level at nearly a hundredfold difference. Faced with this tiered strategy, how should individual users and enterprises actually choose? Drawing on a breakdown of official data, this article examines the three models across three dimensions — pricing, performance, and safety testing — to help you understand where each one fits.

Three-Tier Pricing: A Hundredfold Gap from Premium to Affordable

Let's start with the API pricing structure within this generation, which is the most straightforward way to differentiate them. Astra, the flagship, costs $10 per million input tokens and $50 per million output tokens. Sol, the mid-range option, comes in at $2 input and $10 output. Luna, the entry-level model, is just $0.10 input and $0.50 output. The price gap is enormous — the output cost between the top and bottom tiers differs by roughly a hundredfold.

One important distinction worth emphasizing: these are API usage fees billed per token, not monthly subscription fees. These are entirely different payment models, and individual subscribers and API-calling developers should not conflate them.

Compared to the previous generation, the price cuts are substantial. Sol's input and output prices have both been halved. Luna's input price is halved and output is reduced by about 58%. OpenAI has also stated that the new list prices are half of what the previous version's promotional prices were. These dramatic price cuts essentially lower the barrier for mid-to-low-tier compute, making it accessible to a wider range of long-tail use cases.

A token is the basic unit used by large language models for billing and text processing. Roughly speaking, one token equals about 0.75 words in English; in Chinese, each character corresponds to approximately 1–2 tokens. A 1,000-character Chinese text typically contains 1,500–2,000 tokens. "Per million tokens" is the standard unit for API pricing. For everyday users, a single conversation typically consumes a few hundred to a few thousand tokens — only high-frequency or long-document enterprise use cases will realistically hit million-token volumes. Understanding this unit helps you judge whether a given model tier's actual cost aligns with your use case.

Performance Benchmarks: Mid-Tier Models Approaching the Old Flagship

OpenAI has released results across several benchmarks, but it's worth noting upfront: benchmark scores are not the same as real-world enterprise success rates, nor do they directly translate to actual cost savings.

In office automation evaluations (covering long professional tasks across 47 tools and 55 industries), Sol scored 33.2% and 56.4% respectively. In coding tests, Sol scored 68.8% — very close to the previous-generation Fable 5's 69.9% — while costing roughly 80% less per task. Luna scored 66.6%, at a cost approximately 93% and 96% lower than Opus 5 and Fable 5 respectively.

In other words, the mid-tier Sol achieves near-flagship coding performance at less than 20% of the old flagship's cost. That's arguably the most valuable signal from this update.

For computer operation tasks, Astra remains the strongest. Sol scored 60.5%, comparable to Opus 5's 60.3%, at about 80% lower per-task cost. Luna outperformed the previous-generation Sol's mid-reasoning variant at roughly one-tenth the cost.

OpenAI also demonstrated Sol modifying a web component — after completing the task, it proactively checked for narrow-screen compatibility and browser back-navigation behavior, meaning the model performs self-checks after finishing a task. That said, this was an official demo, not a third-party test, so real-world results should be verified independently.

The "office automation evaluation" and "coding test" mentioned here are AI capability benchmarks — standardized test suites the industry uses to compare models. Common benchmarks include HumanEval (code generation) and MMLU (multi-subject knowledge). The advantage of benchmarks is that they're reproducible and comparable; the downside is that the questions are fixed, and model developers may optimize their training data specifically for these tests, leading to "benchmark gaming." This is why a model that scores high on benchmarks may not perform as well in real business scenarios. The article's repeated emphasis that "benchmark scores don't equal real enterprise success rates" reflects this well-established industry understanding — treat benchmark results as a starting reference point, not the final word.

Safety Red-Teaming: What the Numbers Actually Mean

The fifth evaluation is an adversarial stress test measuring the model's behavioral tendencies in attack scenarios. Importantly, lower percentages are better here. The author specifically cautions that most of these numbers represent "attempts," not "successes."

• Coded deception: Does the model hide problems when writing code? The chart includes all five models' complete data for comparison.
• Failed retrieval: When a search tool returns nothing, does the model honestly admit it rather than pretend it found something? New models show noticeably lower rates than their predecessors.
• Bypassing oversight: Does the model try to circumvent Codex safety reviews? OpenAI states no successful bypasses were observed — the numbers in the table represent attempt counts only.
• Warning evasion: Does the model try to work around access denials after seeing them? Sol's 64.4% is an attempt rate, not a success rate, and is not representative of everyday chat behavior.
• Unauthorized interaction: Does the model take unsanctioned actions in message boards? Sol is at 11.3%, Luna at 0% — but Luna encounters message boards less frequently, so a 0% result doesn't mean it's absolutely safe.

This data set requires careful interpretation. Low percentages are a good sign, but "attempt rate" and "success rate" are completely different things. A high number doesn't mean the model is unsafe, and a 0% doesn't mean it's absolutely reliable.

AI safety red-teaming is a testing methodology that evaluates a model's security boundaries by simulating adversarial attacks — borrowing the concept of a "red team" from cybersecurity. Testers deliberately design prompts intended to induce harmful outputs, bypass safety rules, or execute unauthorized actions, then observe the model's responses. Results are typically reported as "attempt rates" rather than "success rates," because many attempts are blocked by the model itself or external safety mechanisms before causing any real harm. Major AI companies like OpenAI and Anthropic conduct red-team testing before releasing significant models and publish select results in their System Cards for public safety assessment.

How Should Individuals and Enterprises Choose?

The positioning of each model is actually quite clear: Astra is for scenarios with hard requirements around complex computer operations or extreme performance. Sol is the value-for-money choice, delivering near-flagship performance at a fraction of the cost. Luna is suited for high-frequency, simple, cost-sensitive batch tasks.

One viable approach: use Sol for complex planning, Luna to organize customer service inquiries or product data, and have humans review the final output. That said, this is just a conceptual workflow — it's not an officially published revenue case study.

One final point worth remembering: cheaper models don't automatically mean profitable operations. Whether you can turn a profit depends on customer needs and operating costs — and fundamentally on whether the model can genuinely solve real business problems. Lower prices reduce the cost of experimentation, but converting low cost into business value is still the user's own challenge to solve.

Summary

The core logic of this three-tier release is using differentiated pricing to cover the full spectrum of use cases, from flagship to long-tail. Sol's leap in value-for-money and Luna's ultra-low price point make it viable for more mid-to-small use cases to actually deploy AI. But all benchmark data has its limits. When it comes to real selection decisions, the only reliable approach is to run your own tests against your specific tasks.