GPT-5.6 Sol vs Terra vs Luna: A Practical API Model Selection Guide

OpenAI’s GPT-5.6 API family presents a simple headline choice: Sol for maximum capability, Terra for balance, and Luna for economical volume. The concrete numbers make the trade-off more useful. Sol costs five times as much as Luna at the published input and output rates, yet some launch benchmarks show only a modest gap between the three tiers. Other evaluations—especially very long-context recall—show a much larger separation.

Prices, specifications, and benchmark tables in this guide were rechecked on July 29, 2026. The benchmark values below are not TechNest tests. They come from OpenAI’s launch disclosure and an independently operated Artificial Analysis evaluation, each run under its own documented harness and reasoning settings. Treat them as routing clues, not promises about a production workload.

Published specifications and prices

OpenAI’s model catalog lists the same 1.05-million-token context window, 128,000-token maximum output, and February 16, 2026 knowledge cutoff for Sol, Terra, and Luna. Their price and intended position differ.

Prompts above 272,000 input tokens receive higher rates for the full request: OpenAI documents two-times input pricing and one-and-a-half-times output pricing. Cache writes cost 1.25 times the uncached input rate, while cached reads receive the lower prices shown above. Tool calls may add separate fees.

What the coding benchmark shows

OpenAI’s launch table reports the Artificial Analysis Coding Agent Index v1.1 at max reasoning as 80.0 for Sol, 77.4 for Terra, and 74.6 for Luna. Artificial Analysis describes the same family at rounded scores of 80, 77, and 75 in its own launch analysis. Higher is better.

That chart argues against assuming that the lowest-price tier is automatically unsuitable for coding. Luna’s published score is 5.4 points below Sol while its token prices are 80% lower. Terra sits only 2.6 points below Sol at half the listed token price. Those ratios do not mean Luna will complete a particular repository task at one-fifth the cost: reasoning-token use, retries, tool calls, latency, and human repair can change the total.

The safe conclusion is narrower. Start with the least expensive tier that plausibly fits the risk, then measure accepted outcomes. For a mechanical refactor guarded by comprehensive tests, Luna may deserve the first trial. For ordinary implementation with incomplete tests, Terra is a reasonable baseline. For a stubborn cross-system failure or an architecture decision with expensive consequences, Sol may justify the premium.

Long context is where the tiers separate

A shared context-window specification does not imply shared recall quality. In OpenAI’s MRCR v2 eight-needle evaluation at 512,000 to one million input tokens, the launch table reports 73.8% for Sol, 72.5% for Terra, and 41.3% for Luna.

This is a much larger gap: Luna trails Sol by 32.5 percentage points, while Terra trails by 1.3. It supports a practical warning for workloads that place very large document collections directly into one prompt. Luna’s context capacity may accept the request, but that does not establish reliable retrieval across the full input.

Do not solve that problem only by paying for Sol. First remove irrelevant material, retrieve focused passages, preserve source identifiers, and require answers to cite the supplied evidence. Then test the resulting pipeline on documents shaped like the real workload. A stronger tier can help, but good retrieval reduces both cost and the amount of material a model can misread.

Independent intelligence and cost results

Artificial Analysis reports max-reasoning Intelligence Index scores of 59 for Sol, 55 for Terra, and 51 for Luna. Its estimated cost per index task was $1.04, $0.55, and $0.21 respectively. The evaluator also says Luna and Sol, rather than Terra, occupied its intelligence-versus-cost frontier across reasoning settings at the time of testing.

That result complicates the easy “Terra is always the default” story. Terra remains operationally convenient because it is the middle tier and costs exactly half of Sol at the published token rates. But a workload-specific sweep should compare more than the three models at one common effort setting. A higher-effort Luna run may beat a lower-effort Terra run on cost, while Sol at a lower effort may offer a better latency-quality balance than expected.

The independent index combines several evaluations. Its cost per task depends on the evaluator’s prompts, reasoning effort, provider behavior, and token use; it is not a quote for an application. Use it to choose evaluation candidates, not to calculate a production budget.

Compare the cost of an accepted result

Consider a request with 20,000 input tokens and 4,000 output tokens, before caching, long-context multipliers, or tool fees. At list price, one Luna call costs about $0.044, Terra about $0.11, and Sol about $0.22. Across 10,000 calls, that is approximately $440, $1,100, and $2,200.

Those figures are only call cost. If Luna needs two attempts while Terra generally succeeds once, Luna still uses fewer token dollars in this example, but the extra latency and review may erase the saving. If an inadequate answer adds ten minutes of correction, the human cost may dominate. Track first-pass acceptance, accepted result after retries, and correction time alongside tokens.

Caching matters most for stable, input-heavy workloads. Put reusable instructions and reference material in a consistent prefix, then measure actual cache hits. Do not count on the 90% cached-input discount if every request changes the supposedly reusable portion.

A tier-by-tier starting policy

Choose Luna for classification, extraction into a fixed schema, routing, short transformation, and other high-volume work with objective checks. A short prompt is not automatically an easy task, so promote tasks that repeatedly fail validation.

Choose Terra for the broad middle: drafting from supplied evidence, routine coding assistance, support-response preparation, and agent steps that need dependable instruction following without flagship pricing. It is a useful baseline, but the independent cost frontier is a reason to test neighboring Luna and Sol effort settings rather than stop at the middle name.

Choose Sol for difficult debugging, ambiguous architecture, long-horizon agent work, dense long-context synthesis, or final review where an error would trigger substantial rework. Sol should earn its premium on accepted-result data. Sending every request to it can hide weak prompts and routing rules behind a larger bill.

Run a fair internal evaluation

Build a frozen set of 50 to 200 representative tasks, including easy cases, normal cases, and known failures. Remove secrets and personal data. Define acceptance before the run: schema validity, factual support, required fields, test success, reviewer rubric, latency ceiling, and maximum cost are possible checks.

Run all three tiers with the same prompt, tools, reasoning effort, retry policy, and model snapshot. Then repeat promising model-and-effort combinations rather than comparing names alone. Record input, cached input, output and reasoning tokens, end-to-end latency, tool calls, retries, and human correction time. Keep benchmark values in a separate column so vendor or third-party scores do not get confused with your measurements.

Route by task risk instead of department. A support workflow might use Luna to label tickets, Terra to draft ordinary replies, and Sol to analyze an unfamiliar technical escalation. A development workflow might use Luna for mechanical transformations, Terra for ordinary implementation, and Sol for architectural review. Escalate when a cheaper tier returns invalid structured output, fails a test, lacks required evidence, or reaches a retry limit.

The practical choice

The new graphs point to two different lessons. On the published coding-agent index, the three tiers are relatively close; token price and automated verification may dominate. On the very-long-context recall evaluation, Luna falls much farther behind while Terra stays close to Sol. No single ranking captures both patterns.

Terra remains a defensible first baseline for an unevaluated mixed workload. Luna is the optimization target for bounded, checkable volume. Sol is the escalation tier for hard, long-context, or costly work. Promote or demote each task class only when cost per accepted result, latency, retries, and correction effort support the change.