GPT-5.6 Sol vs Terra vs Luna: Which Model Tier Should You Use?

Short answer: choose GPT-5.6 Sol when the task is difficult enough that better reasoning, long-context quality, or tool coordination can justify the highest token rate. Choose GPT-5.6 Terra as the balanced default for most production work. Choose GPT-5.6 Luna when volume and unit cost dominate, the work is bounded, and you can detect or cheaply recover from weaker outputs.

That rule is more useful than treating the names as three separate products. Sol, Terra, and Luna are durable capability tiers inside one GPT-5.6 family. Model tier should answer the quality-versus-cost question; reasoning effort, caching, and processing tier should answer separate questions about how much compute, reuse, latency, or throughput each request needs.

One family, three operating points

OpenAI positions Sol as the flagship frontier tier for complex professional work, Terra as the intelligence-and-cost balance, and Luna as the cost-sensitive high-volume tier. The gpt-5.6 alias routes API requests to Sol, so the unsuffixed name is not a neutral automatic router across all three.

These best-fit examples are editorial routing guidance derived from the official positioning, not a claim that every workload in a row will perform the same way. The defensible choice is the cheapest tier that passes a representative evaluation, including final-answer quality, tool success, total tokens, latency, and recovery cost.

Same interface, different capability headroom

All three model cards publish the same headline capacity and integration surface. Each has a 1,050,000-token context window, a 128,000-token maximum output, text and image input, and text output. Streaming, function calling, and structured outputs are supported; fine-tuning is not. The Responses API supports the same listed tools across the family: web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP, and tool search.

Equal capacity does not mean equal quality at the edge of that capacity. In OpenAI's own vendor-reported MRCR evaluation over 512K to 1M tokens, Sol scored 73.8%, Terra 72.5%, and Luna 41.3%. That is one benchmark rather than a guarantee, but it supports an important buying boundary: Luna can accept the same large prompt without being the safest choice for work that depends on finding and connecting details across it. OpenAI also cautions that its published latency and cost comparisons are estimates and real-world results can vary.

Tool availability therefore usually should not decide between these tiers. Capability headroom, task difficulty, and the cost of a miss should. If a workflow needs unsupported audio or video input, or fine-tuning, switching among Sol, Terra, and Luna will not remove that product constraint.

Standard token prices

The standard short-context rates below are exact USD prices per 1 million tokens. Cached input means a cache read; cache writes have their own higher rate.

Terra is exactly half Sol's standard rate in every column. Luna is one-fifth of Sol and 40% of Terra. Those ratios make Terra a sensible first benchmark: it leaves room to promote difficult cases to Sol or demote proven bulk cases to Luna without changing the API contract.

Per-token price is not the same as per-task cost. A cheaper tier can require more retries, longer prompts, more review, or an escalation call. A stronger tier can also spend more reasoning or output tokens. Compare completed-task cost rather than multiplying the input rate alone, and keep tool-call charges separate from model-token charges.

Output price matters especially for reasoning-heavy agents, long reports, code generation, and workflows that accumulate large intermediate responses. If output dominates the bill, the Sol-to-Terra and Terra-to-Luna gaps remain material even when much of the input is cached.

Long-context surcharge

A prompt with more than 272K input tokens moves the full request to long-context pricing. OpenAI does not charge the higher rate only on tokens beyond the threshold: input, cached input, and cache writes double for the whole request, while output rises to 1.5 times the short-context rate.

This threshold arrives far before the 1.05M-token capacity limit. A standard request with 300K uncached input tokens and 20K output tokens costs about $3.90 on Sol, $1.95 on Terra, or $0.78 on Luna before any separate tool charges. The example preserves the family price ratio, but it does not preserve quality: the vendor long-context result above is a reason to test Luna carefully rather than selecting it only because a corpus is large.

Long context should trigger an architecture review as well as a model choice. Retrieval, chunking, staged reduction, or removing stale conversation history may keep a request below the surcharge and improve focus. If the task genuinely requires global reasoning over the full corpus, Terra can be the stronger cost-quality compromise; Sol is the safer escalation when Terra misses cross-document evidence or dependencies.

Cache economics

GPT-5.6 changes prompt caching from a free write into an explicit economic tradeoff. A cache write costs 1.25 times the model's uncached input rate, while a successful read costs 0.1 times that rate. Ignoring the changing suffix and output, one write plus one later read costs 1.35 uncached-input units instead of 2.0 units for two uncached passes. That is a derived 32.5% saving on the reusable prefix after one successful reuse. A write that is never reused costs 25% more than an uncached pass.

Caching applies to eligible prompts of at least 1,024 tokens and depends on an exact reusable prefix. Stable instructions, examples, files, images, and tool definitions belong before variable request content. GPT-5.6 supports explicit cache breakpoints, and OpenAI recommends a stable prompt_cache_key for its more reliable matching behavior. The only supported minimum lifetime setting is 30 minutes; the prefix remains eligible for at least that long and may be retained longer.

The practical rule is to cache stable prefixes that will actually recur. Monitor cache_write_tokens and cached_tokens instead of assuming reuse. Explicit-only mode can avoid an unnecessary write when no explicit breakpoint is present. For high-volume workloads with a large shared prefix, caching can make Terra or Sol economically viable when their quality gain matters.

Caching lowers absolute cost but does not reverse the tier ladder because the write and read multipliers are the same for all three models. Sol remains five times Luna's token rate, and Terra remains 2.5 times Luna's. Long-context cached reads and writes also receive the long-context multiplier, so caching helps without making the >272K boundary disappear.

Latency, throughput, and service tier

OpenAI labels all three model cards as Fast and calls Luna its fastest tier, but it does not publish a universal standard-API tokens-per-second or end-to-end latency number for comparing Sol, Terra, and Luna. Reasoning effort, output length, tool calls, safety checks, and service tier can change the observed result. A special throughput claim for a limited provider route should not be treated as normal API speed.

The family supports none, low, medium, high, xhigh, and max reasoning effort. OpenAI recommends low as a latency-oriented starting point, medium as a balanced start, and max only for the hardest quality-first work. Pro mode keeps the selected model slug and spends more model work, increasing latency and token use. Ultra is not a fourth GPT-5.6 tier: it is a multi-agent product mode, while the API analogue is the Responses API multi-agent beta.

Processing choice can matter as much as model choice:

Do not buy Sol merely to solve a service-level problem. A latency-sensitive Terra request may need lower reasoning effort or Priority processing; a high-volume Luna job may fit Batch; an inexpensive experiment may fit Flex. Keep tier, reasoning effort, and processing route as separate controls so each can be evaluated and changed independently.

A practical routing policy

Start with Terra at the same reasoning effort you plan to use in production. Build a representative set of easy, typical, and failure-prone tasks, then measure correctness, evidence quality, tool completion, retries, output tokens, latency, and human review time.

Promote a workload to Sol when Terra misses dependencies, loses important evidence in long context, fails multi-step tool coordination, or creates enough review risk that the token savings are false economy. Sol best fits complex code changes, demanding research, professional artifacts, and long-horizon agents where an incremental quality gain has meaningful value.

Demote a workload to Luna when the task is narrow, repeated at scale, and objectively checkable. Good candidates include classification, extraction, normalization, routing, bounded summarization, bulk enrichment, and first-pass drafts with validation or escalation. Keep ambiguous inputs and failed checks on a Terra or Sol fallback path.

Re-evaluate the route when prompts cross 272K input tokens, cache reuse changes, output volume grows, or the workload moves between interactive and offline processing. Those changes can alter cost more than a small prompt tweak. They can also expose quality differences that were invisible on short, clean examples.

The final rule is simple: use Terra until evidence says otherwise. Move up to Sol for measured quality headroom and down to Luna for measured volume economics. Keep the same acceptance test across tiers, label OpenAI benchmark results as vendor claims, and never substitute a published context-window size for a workload-specific evaluation.