What changes between the three tiers?

The three models belong to the same GPT‑5.6 generation and expose the same basic API shape. All three have a 1,050,000-token context window, a 128,000-token maximum output, text and image input, text output, function calling, structured outputs, streaming, and support for OpenAI Responses API tools including web search, file search, code interpreter, hosted shell, computer use, MCP, tool search, image generation, and apply patch.

Pricing: Sol is 5× Luna and 2× Terra

OpenAI’s standard API rates are straightforward. Each step down the family halves the cost from Sol to Terra, while Luna costs one-fifth as much as Sol. Cached input is discounted by 90%. Cache writes are billed at 1.25× the uncached input rate, and OpenAI documents a 30-minute minimum cache life for GPT‑5.6 and later models.

Benchmark overview: the family scales gradually, but not uniformly

The following figures are copied from OpenAI’s published GPT‑5.6 evaluation tables. They are not a composite ranking. Each benchmark has its own dataset, harness, tools, reasoning setting, and failure definition. The most useful pattern is the shape of the gaps: Terra is often close to Sol, while Luna is the strongest cost-saving option but gives up more on the hardest tasks.

Coding and agentic work

Sol leads the family on the independent Artificial Analysis Coding Agent Index, DeepSWE, and Terminal‑Bench 2.1. Terra remains close on Terminal‑Bench and SWE‑bench Pro, while Luna is still competitive on the published agentic suite at one-fifth of Sol’s token price.

Professional work and computer use

The family’s biggest operational separation appears in computer-use and browsing evaluations. Sol leads Terra and Luna by wider margins on OSWorld 2.0 and BrowseComp. This matters for workflows that must navigate interfaces, inspect rendered output, browse for evidence, or operate across many intermediate steps.

Science, math, multimodal, and tools

For difficult scientific and abstract-reasoning tasks, the tier gap widens. Sol leads FrontierMath Tier 4 by 14.5 points over Terra and 24.5 points over Luna. On lower-difficulty or broader tasks, the gaps are smaller. The multimodal and tool-use results also show that Luna is not uniformly below Terra: it edges Terra on Toolathlon and BenchCAD, so model selection should follow the workload rather than a simplistic tier assumption.

Cybersecurity benchmarks: capability evidence, not deployment advice

OpenAI reports substantial capability differences on its cybersecurity evaluations. These are dual-use benchmarks. We include them as published model-evaluation data, not as instructions for offensive activity. Production security work should remain authorized, sandboxed, monitored, and reviewed by qualified people.

Long context: same window, different recall quality

All three models advertise the same 1.05M-token context window, but the published long-context evaluations show that capacity and retrieval quality are different properties. Sol and Terra are close in the 256K–512K MRCR range; Luna’s score falls substantially on the same test. GraphWalks shows a similar pattern at 1M tokens, although Luna slightly beats Terra at 256K.

Independent Artificial Analysis view

Artificial Analysis independently evaluated the GPT‑5.6 family before launch. Its rounded Intelligence Index scores were 59 for Sol, 55 for Terra, and 51 for Luna. It reported estimated Intelligence Index task costs of approximately $1.04, $0.55, and $0.21, respectively, and Coding Agent Index scores of 80, 77, and 75.

These are estimated benchmark-task costs, not API invoices. They include the evaluation harness, reasoning effort, output usage, and task behavior. They are useful for efficiency comparisons, but your actual workload may have a different input/output ratio, tool pattern, cache hit rate, and retry rate.

Capability-versus-cost radar

Which model should you use?

Verdict

GPT‑5.6 Sol is the capability leader. It leads every row in the official coding, professional, science, computer-use, academic, and long-context tables included here. It is also the most expensive tier.

GPT‑5.6 Terra is the safest default recommendation. It retains much of Sol’s performance on everyday work, costs half as much, and shares the same 1.05M context and 128K maximum output. For many production applications, it is the tier that should be measured first.

GPT‑5.6 Luna is the economics play. At one-fifth of Sol’s token price, it is compelling for high-volume and retryable workloads. Its weaker results on the hardest reasoning and long-context evaluations mean it should not be selected blindly for complex autonomous work.

The best answer is therefore not one model. It is a measured routing policy: Luna for volume, Terra for the broad middle, and Sol for the tasks that earn the premium.

Sources and methodology

• OpenAI — GPT‑5.6: Frontier intelligence that scales with your ambition. Primary source for all benchmark tables, model positioning, availability, pricing, caching, and the Sol/Terra/Luna comparison.
• OpenAI API docs — GPT‑5.6 Sol, Terra, and Luna. Primary source for model IDs, context, output limits, modalities, features, and token prices.
• Artificial Analysis — GPT‑5.6 benchmarks across Intelligence, Speed and Cost. Independent evaluation context, rounded index scores, and estimated benchmark-task costs.
• Figures are reproduced as published and retain their original metric scales. A benchmark score can depend on reasoning effort, scaffolding, tools, model harness, dataset, and time limit. Radar and bar charts are visual aids, not composite rankings. “Cost per task” figures from Artificial Analysis are estimates and are not guaranteed API bills.
• Cybersecurity results are included for capability context only. Any real security testing should be authorized, isolated, monitored, and performed with appropriate safeguards.