GPT-5.6 for Small Business: Sol vs Terra vs Luna

Q: Which GPT-5.6 model should a small business use?

Use Sol when the problem is hard or unclear, Terra when the plan is already made, and Luna for narrow repeatable work that is inexpensive to verify.

Q: When should a business use GPT-5.6 Luna?

Use Luna for narrow, repeated work such as classification, extraction, metadata, and structured transformations when mistakes are inexpensive to detect.

Q: Were GPT-5.6 juice values permanently reduced?

OpenAI product lead Tibo Sottiaux said internal reasoning-budget experiments were reverted. Circulating reduced hidden values are not a current public product specification.

Executive summary -- TL;DR / BLUF

Do not buy the strongest model for every task. Route the work to the least expensive model that can finish it reliably, show evidence, and stop at the owner's approval boundary.

Sol investigates. Start with Sol High when the problem is difficult, unclear, or spread across systems.
Terra executes. Use Terra Medium when the plan, boundaries, and acceptance checks are already defined.
Luna processes. Use it for narrow, repeated work with an inexpensive or automatic review path.
Max is an escalation. Use it after High fails because the model did not explore enough.
Ultra is a team. Use it only when the work can be separated into independent branches.

Our week of real use matched that split. Sol found important problems but needed a stopping rule. Terra was sharp against named gates. Luna reliably produced structured output when a validator was part of the workflow.

The business decision is uncertainty plus review cost

A model picker makes this look like a technical preference. For an owner, it is an operating decision. How much of the path is unknown? How expensive is a mistake? How quickly can someone check the result?

The cheapest token price is not always the cheapest workflow. A low-cost output becomes expensive when it creates retries, cleanup, or a customer-facing mistake.

The useful model is the least expensive one that can reach a verified result.

What one week of field use showed us

This was observational use across websites, evaluation work, research, and tooling--not a controlled head-to-head benchmark. The models handled different work, so these are operating notes rather than universal rankings.

Sol handled uncertainty best. It carried long investigations and found material errors in the first fact-check of this guide. It also continued after the useful answer had already appeared. Sol needs an evidence target and an exit condition.

Terra improved when the decision was already made. Against explicit acceptance criteria, it found concrete failures and returned clear approve-or-hold judgments. Terra needs a bounded deliverable and named gates.

Luna worked when the output shape was exact. It produced machine-checkable JSON and stayed within requested boundaries. Luna needs a schema, a known count, a sample audit, or another cheap validator.

Use Sol when part of the job is figuring out the job

Sol is the right starting point for difficult research, broken workflows, cross-system failures, unclear requirements, and consequential planning. Use it when the model must inspect evidence, compare paths, or decide whether the premise is correct before it can execute.

Start on Sol High. Give it room to investigate, but name what it may change, what it must not touch, what evidence counts as completion, and when it should stop.

Do not choose Sol only because a task matters. A known five-line change can matter and still belong to Terra.

Use Terra when the path and finish line are visible

Terra fits approved plans, bounded site updates, documented migrations, checklist reviews, test writing for known behavior, and other work where execution matters more than discovery.

Start on Terra Medium. OpenAI lists Terra's API token prices at half of Sol's: $2.50 per million input tokens and $15 per million output tokens, compared with Sol at $5 and $30. The real saving appears only when the same quality gate still passes.

Use Luna behind repeated work

Luna fits classification, extraction, metadata, titles, record cleanup, simple summaries, and structured transformations. Its API pricing is $1 per million input tokens and $6 per million output tokens.

The test is not whether the task sounds easy. The test is whether it is narrow, repeatable, and cheap to verify. If one unnoticed error can contaminate hundreds of records, add validation or use a stronger model.

Make Max earn its cost

Higher reasoning effort gives a model more room to explore and check. It does not guarantee that the improvement will justify the added time and usage.

DataCurve's raw July 9 DeepSWE v1.1 artifact covers 113 coding-agent tasks:

Sol effort	Score	Estimated cost per task
High	69.4%	$3.47
Extra High	70.7%	$4.70
Max	72.7%	$8.39

High to Max added about 3.3 observed points at roughly 2.4 times the estimated cost. The published confidence intervals overlap slightly. That is evidence from one harness, not proof that Max will improve your next business task.

If High failed because the brief was vague, the source material was wrong, permissions were missing, or the environment was broken, fix that first. More compute does not repair a bad premise.

Route the check with the work

Work state	Starting route	Verification
Hard or unclear	Sol High	Evidence record, reproduction, or tested plan
Decided and bounded	Terra Medium	Named checklist, acceptance gate, or reviewed deliverable
Narrow and repeated	Luna	Schema, deterministic check, or sample audit
High failed from shallow exploration	Sol Max	Compare new evidence against the failed attempt

A stronger model does not earn broader authority. Prices, refunds, customer promises, legal or health claims, sensitive data, and anything speaking in the owner's voice still need a visible approval step.

Check the product before copying someone else's picker

Standard ChatGPT: GPT-5.6 Sol powers Medium, High, and Extra High. Terra and Luna are not selectable there.
ChatGPT Work and Codex: eligible plans can expose Sol, Terra, and Luna. Max and Ultra vary by product and plan.
API: all three models support the published effort controls from none through max.

Preview screenshots age quickly. Current documentation should be the source of truth.

Max, Fast, and Ultra are different controls

Max gives one model more reasoning time. Ultra coordinates multiple agents and only helps when their assignments are genuinely independent. Fast is separate: current Codex Speed documentation lists GPT-5.5 and GPT-5.4, not GPT-5.6.

Parallel agents can reduce elapsed time across independent branches. They waste usage when every worker reads the same context or collides on the same decision.

Separate token price, benchmark cost, and Codex credits

API price is the published cost per million input and output tokens.
Benchmark cost is an estimate from one evaluation harness, not a fixed price for your next job.
Codex credits use a separate rate card. For most plans, one million input/output tokens maps to Sol 125/750 credits, Terra 62.5/375, and Luna 25/150, with lower cached-input rates.

Context length, tool output, retries, and parallel agents all change real consumption. Budget from the current Codex rate card, not a benchmark average.

P.S. The current five-hour limit and "juice value" update

As of July 12, 2026, the five-hour restriction for Codex and ChatGPT Work temporarily does not apply to Plus, Business, or Pro. OpenAI product lead Tibo Sottiaux published the update, and an accessible contemporaneous report reproduces it. Weekly limits still apply. This is not unlimited usage or a permanent plan promise.

Sottiaux also said OpenAI experimented with internal reasoning budgets--"juice values"--while investigating unexpectedly high usage and then reverted the experiments. The public screenshot thread is preserved here. Exact reduced hidden values circulating online are not current public specifications.

A five-step owner routing rule

Name the job and what finished means.
Classify the path as unclear, known, or repeatable.
Choose Sol, Terra, or Luna for that state.
Name the evidence, validator, and owner approval boundary.
Increase effort only after diagnosing why the first attempt failed.

For the coding-agent architecture, product-surface details, and implementation policy, read KyaniteLabs' technical GPT-5.6 routing guide.

FAQ

Which GPT-5.6 model should a small business use?

Use Sol when the problem is difficult or unclear, Terra when the plan is already made, and Luna for narrow repeated work that is inexpensive to verify.

Should GPT-5.6 Sol use High or Max reasoning?

Start difficult, uncertain work on Sol High. Escalate to Max only when High failed because it did not explore deeply enough.

When should a business use GPT-5.6 Luna?

Use Luna for classification, extraction, metadata, and structured transformations when the output is narrow and mistakes are inexpensive to detect.

Do Codex and ChatGPT Work currently have a five-hour usage window?

As of July 12, 2026, OpenAI says the five-hour restriction for Codex and ChatGPT Work temporarily does not apply to Plus, Business, or Pro. Weekly limits still apply.

Were GPT-5.6 juice values permanently reduced?

No current public specification says that. OpenAI product lead Tibo Sottiaux said the internal reasoning-budget experiments were reverted.

Sources

Fact-checked July 12, 2026. Model availability, prices, product controls, and usage rules can change.

GPT-5.6 for small business: how to choose Sol, Terra, or Luna