The useful comparison is not price per token
OpenAI’s July 30, 2026 GPT-5.6 update changed the arithmetic for teams that run AI workloads at scale. GPT-5.6 Luna is now priced at $0.20 per million input tokens and $1.20 per million output tokens. Terra is $2 per million input tokens and $12 per million output tokens. Standard Sol pricing remains $5 per million input tokens and $30 per million output tokens. OpenAI also says the lower Luna and Terra prices are reflected in how usage is counted in Codex and ChatGPT Work, while subscription prices and quota budgets remain unchanged.
Those are important numbers, but they do not answer the question a marketing team should ask: which model produces the lowest cost per accepted result?
An inexpensive output that a human must rewrite is not inexpensive. A more capable output that takes too long may also be poor value. The right comparison includes model cost, retries, review, escalation, latency and the value of the work that the result enables.
OpenAI’s official price-performance announcement describes Luna as its fastest and most affordable GPT-5.6 model, Terra as the balanced option for everyday work, and Sol as the premium model with a new Fast mode in the API. This guide turns that product update into a measurement framework for marketing operations.
GPT-5.6 Luna vs Terra vs Sol: the current API rates
The standard published rates are easiest to understand in a table.
| Model | Input per 1M tokens | Output per 1M tokens | Position in the family | First workload to test |
|---|---|---|---|---|
| GPT-5.6 Luna | $0.20 | $1.20 | Fastest, lowest-cost tier | High-volume, repeatable work |
| GPT-5.6 Terra | $2.00 | $12.00 | Balanced everyday tier | Evidence synthesis and normal implementation |
| GPT-5.6 Sol | $5.00 | $30.00 | Highest-intelligence tier | Ambiguous, high-impact or difficult work |
OpenAI reports that Luna fell from $1/$6 to $0.20/$1.20 per million input/output tokens, while Terra fell from $2.50/$15 to $2/$12. Sol’s standard rates did not change. The price reduction is therefore real, but it is not a guarantee that Luna can replace Sol for every job.
The GPT-5.6 Luna model documentation is the right place to recheck current availability, aliases and pricing before changing a production budget. Pricing and model behavior are operational data, not permanent editorial facts.
For background on the broader model family, see DMT’s GPT-5.6 prompting guide for Sol, Terra and Luna. This article focuses on a narrower decision: how to select a model using accepted work as the unit of value.
What “cost per accepted result” means
Define an accepted result before you compare models. An accepted result is an output that passes the minimum quality and policy checks for a specific job without unplanned repair work.
Examples include:
- a search query classified into the correct intent bucket;
- a meta description that meets the character and claim rules and needs no rewrite;
- an ad variation that passes brand and policy review;
- a structured extraction that validates against the expected schema;
- an article brief with working sources, a clear angle and no obvious cannibalization;
- a routine implementation step that passes its tests and review.
The measurement unit is:
Cost per accepted result = total model cost + attributable review cost + retry cost + escalation cost, divided by accepted results.
For a first API-only comparison, you can measure model cost as:
(input tokens × input rate + output tokens × output rate) ÷ 1,000,000.
Then track the other costs separately. This separation matters because a cheaper model can have a lower token bill but a higher rejection rate. It also prevents a high-quality model from appearing inefficient merely because it was given harder work.
A simple arithmetic comparison
Assume the same job uses 1 million input tokens and 100,000 output tokens, stays below any long-context multiplier, and does not include tool-specific charges.
| Model | Input cost | Output cost | Token cost for the job |
|---|---|---|---|
| Luna | $0.20 | $0.12 | $0.32 |
| Terra | $2.00 | $1.20 | $3.20 |
| Sol | $5.00 | $3.00 | $8.00 |
If each model produced one accepted result, Luna would win on token cost. But suppose the workflow accepts 70% of Luna outputs, 90% of Terra outputs and 97% of Sol outputs. The approximate token cost per accepted result becomes:
- Luna: $0.32 ÷ 0.70 = $0.46;
- Terra: $3.20 ÷ 0.90 = $3.56;
- Sol: $8.00 ÷ 0.97 = $8.25.
Luna still wins in this example, but the gap is not the same as comparing raw prices. If Luna requires a costly human rewrite on rejected cases, the result changes again. The point is not to produce a universal winner; it is to measure the workload you actually run.
The four metrics to record for each model
1. Acceptance rate
Use a labelled sample and define pass/fail rules before you run it. For query classification, calculate accuracy or macro-F1. For structured extraction, calculate field-level validity and completeness. For content, use a checklist for source coverage, factual correctness, intent fit, readability and editorial risk.
Do not let the model grade itself. Use a human reviewer, deterministic validator or an independent evaluation model for the quality decision. If a reviewer model is used, keep its rubric and model version in the log.
2. Cost per accepted result
Include retries. A run that asks Luna to repair its own answer three times is not equivalent to a single-pass result. Count the complete input and output tokens for every attempt, and mark whether the extra work was triggered by the model, the prompt, a tool failure or a human change.
3. Time to accepted result
Measure wall-clock time from request start to approval, not just first-token latency. A quick draft that waits for several repair rounds may be slower than a single Sol answer. Conversely, a background classification queue may have no meaningful human waiting cost, making raw speed less important.
4. Escalation rate
Record how often the workflow sends a case to Terra, Sol or a human. A healthy routing system does not force Luna to solve every ambiguous case. It uses a confidence threshold and preserves a safe path for uncertainty.
A model-routing matrix for marketing work
| Job | First model to test | Escalate when | Acceptance check |
|---|---|---|---|
| Search-query classification | Luna | Low confidence or mixed intent | Label accuracy on a held-out sample |
| URL, title and metadata extraction | Luna | Missing fields or malformed values | Schema validation and exception queue |
| Ad headline and description variants | Luna | Policy, legal or brand-sensitive wording | Human approval against a written checklist |
| Internal-link candidate labelling | Luna | Ambiguous page ownership or repeated anchors | URL resolution and anchor-ledger review |
| SEO brief synthesis | Terra | Conflicting sources or unclear intent | Source coverage, intent and duplicate review |
| Campaign diagnosis | Terra or Sol | Conflicting attribution signals or high spend | Reproducible calculations and assumption log |
| New strategy or migration decision | Sol | Never skip human approval | Evidence-backed recommendation and risk review |
| Routine implementation after a locked brief | Luna | Tests fail or scope changes | Automated tests plus independent review |
This is a starting hypothesis, not a promise about model performance. OpenAI itself recommends matching the level of intelligence to the outcome, urgency, scale and cost of error. Your evaluation should be stricter for work that changes budgets, publishes claims, sends external messages or modifies customer data.
How to run a fair seven-day evaluation
- Freeze the task sample. Select representative marketing jobs from the same period. Do not give Luna only easy work and Sol only ambiguous work.
- Write the acceptance rubric first. Define what counts as correct, incomplete, unsafe or requiring escalation.
- Run the same inputs. Keep system instructions, source set, output schema and tool permissions comparable.
- Log tokens and retries. Store input tokens, output tokens, cache behavior where available, elapsed time, model, reasoning setting and every retry.
- Review blindly. If possible, hide the model name from the quality reviewer so brand expectations do not affect the score.
- Calculate accepted-result economics. Report acceptance rate, cost per accepted result, time to accepted result and escalation rate together.
- Promote only bounded jobs. Move one workflow step into production at a time, with rollback and a sample-based drift check.
For teams using Codex, OpenAI’s lower subscription usage accounting for Luna and Terra should be tracked as a separate operational measure. It is not the same thing as converting API token prices into a fixed number of extra messages. DMT’s GPT-5.6 Codex usage-limit guide covers why context size, tools and workflow shape affect consumption.
What marketers should not conclude from an 80% cut
The price change does not mean that every marketing workflow should move to Luna. It does not prove that Luna has the same quality as Sol on complex strategic work. It does not guarantee five times more subscription messages, because API rates and product usage meters are different systems. It also does not remove the need for source verification, human approval or permission controls.
The useful conclusion is more practical: teams can afford to test more routing choices. A task that was too expensive to evaluate at scale may now support a labelled benchmark, a review pass or a broader exception sample. That extra measurement is more valuable than simply producing more generic content.
For related operating context, see DMT’s GPT-5.6 price-cut baseline, agent-harness context analysis, and governed Codex and ChatGPT Work guide. Those pages cover the underlying price update, orchestration risks and permission boundaries; this article owns the accepted-result measurement method.
FAQ
Which GPT-5.6 model is cheapest?
GPT-5.6 Luna has the lowest standard API price at $0.20 per million input tokens and $1.20 per million output tokens, according to OpenAI’s July 30 announcement.
Is Luna always the best model for marketing?
No. Luna is a strong first candidate for repeatable, high-volume, lower-risk work. Terra or Sol may be better when evidence conflicts, the cost of error is high or the output requires deeper judgment.
How do I compare models fairly?
Use the same labelled task sample, instructions, source set and acceptance rubric. Report acceptance rate, total cost, retries, time to accepted result and escalation rate together.
Did ChatGPT or Codex subscriptions become cheaper?
OpenAI says subscription prices and quota budgets remain unchanged. Luna and Terra usage now consumes fewer credits in Codex and ChatGPT Work, but OpenAI has not published a universal message multiplier.
Should I publish AI-generated marketing content without review?
No. Use lower costs to improve research, checking and experimentation. Keep factual review, brand review, approval and rollback controls for public content and consequential actions.
Bottom line
GPT-5.6 Luna is the obvious model to test first when a marketing task is repeatable, measurable and high volume. But the business decision should be based on cost per accepted result, not the headline token price. Terra can be the middle path for evidence synthesis and normal implementation. Sol remains the escalation tier for ambiguity and high-cost mistakes.
The most durable GPT-5.6 advantage is not generating five times as much output. It is routing each step to the least expensive model that can meet a tested quality standard—and spending the savings on better evaluation, review and measurement.
About the author
Tayeeb Khan writes Digital Marketer Tayeeb’s research-focused coverage of AI, SEO and digital marketing workflows. This guide separates OpenAI’s published product facts from the practitioner measurement framework proposed here.
