GPT-5.6 Benchmarks Explained: Sol vs Terra vs Luna, Pricing, and Best Use Cases

OpenAI's GPT-5.6 family has three durable capability tiers: Sol for the hardest work, Terra for balanced production workloads, and Luna for fast, cost-sensitive traffic. The benchmark story is not simply that Sol wins. Terra and Luna often retain much of the capability while using lower-priced tokens, making model routing more important than choosing one model for every request.

OpenAI released the full GPT-5.6 family on July 9, 2026. This guide explains the published benchmark results, what they measure, and how to combine quality, price, latency, and retry rates when selecting a model.

Data and prices checked July 19, 2026. Benchmark results come from OpenAI's launch materials and use the settings described there. They are not direct guarantees for a different application or agent harness.

GPT-5.6 benchmark summary

The table below focuses on representative professional and coding evaluations.

Sol leads across these selected evaluations. Terra stays close on several coding tasks. Luna is not universally stronger than GPT-5.5, but it remains competitive while occupying the fastest and least expensive GPT-5.6 tier.

The table also illustrates why a single average score is insufficient. A support classifier, code-review agent, financial-research workflow, and browser operator do not value the same capabilities.

What GPT-5.6 benchmarks actually measure

Coding-agent benchmarks

Coding benchmarks vary substantially. Some test whether a model can produce a patch. Others include repository exploration, terminal commands, test execution, and iterative correction.

OpenAI reports that GPT-5.6 Sol reached 80 on the Artificial Analysis Coding Agent Index while using fewer tokens and less time than several frontier alternatives in its comparison. Terra scored 77.4 and Luna 74.6. For a production coding agent, the important follow-up questions are:

• Does it find the correct files?
• Does it preserve project conventions?
• Does it run and interpret tests?
• Does it recover after a failed approach?
• How many model turns are required before a verified result?

A model with a slightly higher patch score can still be less economical if it produces long reasoning traces, repeats tool calls, or needs expensive human review.

Long-horizon agent benchmarks

Agents' Last Exam evaluates longer professional workflows across many domains. Sol scored 52.7%, while Terra and Luna scored 50.4% and 50.3%. The narrow spread is notable: it suggests that many professional tasks may not require the flagship tier.

Long-horizon evaluations are especially sensitive to the surrounding system. Tool design, memory, retry policies, permissions, and verification steps can change results as much as the base model.

Browsing and computer use

OpenAI reports 92.2% for GPT-5.6 Sol on BrowseComp and 62.6% on OSWorld 2.0. These results support research agents, browser workflows, and applications that must inspect their own output.

Computer-use benchmarks should not be interpreted as a promise of unattended reliability. Production systems still need scoped permissions, confirmation for important actions, timeouts, and recovery behavior.

Knowledge work and artifacts

GPT-5.6 is also positioned for documents, spreadsheets, presentations, research briefs, and frontend work. These outputs are difficult to capture with one benchmark because correctness, source fidelity, visual quality, and editability all matter.

For these workloads, build an internal rubric. Score factual accuracy, structure, adherence to a reference template, number of manual fixes, and time to an acceptable deliverable.

GPT-5.6 Sol vs Terra vs Luna

Choose GPT-5.6 Sol for high-value hard problems

Sol is the flagship tier. It is the best starting point for:

• repository-scale coding and difficult debugging;
• long-running agents with multiple tools;
• cybersecurity and scientific analysis within authorized use;
• complex research and professional artifacts;
• tasks where a failed result costs far more than the token bill.

Sol is not automatically the best default for high-volume traffic. Its official token rates are twice Terra's and five times Luna's.

Choose GPT-5.6 Terra as the balanced default

Terra is designed for everyday production work. It fits:

• coding copilots and code review;
• document analysis and research support;
• internal assistants;
• tool-using workflows that need more reliability than a lightweight model;
• products that want GPT-5.6 capability without Sol pricing.

Terra's benchmark scores stay close to Sol on several coding evaluations, making it a sensible first candidate for production testing.

Choose GPT-5.6 Luna for throughput and predictable tasks

Luna is the fastest and most affordable tier. It is suited to:

• extraction and classification;
• structured content processing;
• lightweight support;
• high-volume automation;
• a first routing stage that escalates difficult requests.

Luna can also handle some coding and professional tasks, but benchmark averages should not decide that boundary. Test where its failure or retry rate starts to erase the token savings.

GPT-5.6 API pricing

OpenAI prices GPT-5.6 per 1 million tokens. Poyo.ai currently provides all three tiers through the Responses API at the rates below.

OpenAI also documents explicit cache breakpoints, a 30-minute minimum cache life, cache writes at 1.25 times the uncached input rate, and cache reads at a 90% discount from uncached input.

Check the GPT-5.6 API model page for current Poyo pricing, supported model IDs, and API access.

Real workload cost comparison

Consider a monthly workload with 10 million uncached input tokens and 2 million output tokens.

This calculation is useful but incomplete. It assumes every model completes the same number of tasks with the same number of attempts.

Benchmark score versus cost per successful task

The more useful production metric is:

effective task cost =
  average API cost per attempt
  × average attempts per successful task
  + human correction cost

Suppose Luna costs one-fifth as much as Sol per token, but a difficult workflow succeeds on only one of three attempts. Terra might complete the same task once, with fewer tool calls and less human review. Terra would then be the cheaper production choice even though its token rate is higher.

Measure at least:

• pass rate;
• attempts per pass;
• input and output tokens;
• elapsed time;
• tool calls;
• human correction minutes;
• severe-error rate.

The last metric matters for workflows where an incorrect answer is more expensive than a failed request.

A practical GPT-5.6 routing strategy

A three-tier family is most valuable when the application routes work deliberately.

Is the task repetitive and easy to verify?
├── Yes: start with Luna
└── No
    ├── Is it normal production coding, analysis, or tool use?
    │   └── Start with Terra
    └── Is failure expensive or the task unusually difficult?
        └── Use Sol

A production router can use task type, request value, previous failures, confidence signals, or evaluator output. A common pattern is:

1. Send structured and low-risk requests to Luna.
2. Use Terra as the default for general production reasoning.
3. Escalate failed or high-value cases to Sol.
4. Compare the final cost per successful task by route.

Do not route solely on prompt length. A short security problem can be harder than a long extraction request.

How to evaluate GPT-5.6 in your application

Build an evaluation set from real traffic rather than generic prompts.

1. Select 30 to 100 representative tasks.
2. Label the expected outcome and unacceptable failure modes.
3. Run the same task on Luna, Terra, and Sol with comparable settings.
4. Record quality, latency, tokens, tools, retries, and human corrections.
5. Segment results by task type.
6. Choose routing rules from the segment results.
7. Repeat after prompt, tool, or model updates.

For agents, include difficult cases: broken tool responses, ambiguous data, missing files, and tasks that require self-correction. Easy demos tend to hide the difference between model tiers.

How to use GPT-5.6 on Poyo.ai

Poyo.ai exposes the model IDs gpt-5-6-luna, gpt-5-6-terra, and gpt-5-6-sol through /v1/responses.

curl https://api.poyo.ai/v1/responses \
  -H "Authorization: Bearer $POYO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5-6-terra",
    "input": "Review this migration plan and identify the highest-risk assumptions.",
    "reasoning": {"effort": "medium"}
  }'

Create a key in the Poyo dashboard and review the GPT-5.6 model page before production use.

Which GPT-5.6 model should you choose?

Start with Luna when throughput and easy verification matter. Start with Terra for general coding, analysis, and production agents. Use Sol when the problem is unusually hard or failure has a high business cost.

The best answer is rarely a permanent single-model choice. GPT-5.6 is structured as a family, and the strongest economic design uses the tiers together.

Frequently asked questions

Which GPT-5.6 model is best?

Sol is the most capable tier, but Terra may offer a better production balance and Luna is substantially cheaper for simple, high-volume tasks.

Is GPT-5.6 Luna good for coding?

Luna is capable of coding work, but published results trail Terra and Sol on the selected coding evaluations. Test it on your actual repository and escalate tasks that fail verification.

How much does the GPT-5.6 API cost?

Official OpenAI rates range from $1 input and $6 output per 1 million tokens for Luna to $5 input and $30 output for Sol. Poyo.ai currently lists lower pay-as-you-go rates for all three tiers.

Does GPT-5.6 use the Responses API?

Yes. Poyo.ai exposes the three GPT-5.6 model IDs through /v1/responses.

Should I replace GPT-5.5 with GPT-5.6 Terra?

Terra is a strong migration candidate, but compare task success, latency, tokens, and tool calls on a representative evaluation set before changing all traffic.

Sources

• OpenAI GPT-5.6 launch, benchmarks, availability, and pricing
• OpenAI API pricing
• Poyo.ai GPT-5.6 API