GPT-5.6: Sol, Terra and Luna — what each model can do and what it costs

On July 9, 2026, OpenAI introduced GPT-5.6, its new model family: Sol, Terra y Luna. Three weeks later, on July 30, it did something far less routine than a launch: it cut Luna's price by 80%. The cheapest model in the family went from USD 1 to USD 0.20 per million input tokens, and from USD 6 to USD 1.20 for output.

That second piece of news is the important one, and almost nobody covered it properly. A cut of that size, applied to the smallest model of the newest generation, changes the math for anyone running AI agents at volume. It is no longer true that "the good model is expensive and the cheap one is dumb": Luna scores 51 on the Artificial Analysis intelligence index — 8 points off the flagship — and costs one fifth of what it cost two weeks ago.

At AsisteClick the full 5.6 family has been available in the AI agent builder since this weekend's production release. You can pick Sol, Terra or Luna in the model selector of any agent, without touching code. This article explains what each one can do, where Luna sits on the cost/intelligence trade-off, and what it really costs to handle a thousand conversations with each model.

Table of contents

• What the GPT-5.6 family is
• Sol: the ceiling
• Terra: the awkward middle
• Luna: the surprise
• The map: intelligence against cost
• The benchmarks, side by side
• The price: what changed on July 30
• What it costs to handle 1,000 conversations
• Why the cut lands differently in LATAM
• Which one to choose for customer service
• How to measure it on your own traffic
• What the benchmarks don't tell you
• How to use the 5.6 family in AsisteClick
• The summary that matters
• Keep reading

What the GPT-5.6 family is

The names form an astronomical scale: Sol, Terra — earth, in Latin — and Luna. From most to least powerful, and from most to least expensive. It is OpenAI's first meaningful naming change in years — the mini y nano suffixes carried over from the 4.1 series are gone — and there is a logic to it: all three models share the same architecture and the same API capabilities, and what varies is the compute budget each one devotes to thinking.

All three share these traits:

• A 1.05 million token context window — up to 922K input and 128K output. Requests above 272K input tokens carry a surcharge (2× input and 1.5× output across the whole request), so the giant window is best used deliberately, not by default.
• Programmatic tool calling (programmatic tool calling): instead of requesting one tool per turn, the model writes JavaScript that runs in an isolated V8 runtime with no network access. OpenAI documents reductions of between 38% and 63.5% in tokens consumed for customers orchestrating many tools per conversation. For an agent that checks stock, validates a tax ID and books an appointment within the same conversational turn, that is real money.
• Prompt caching with a 90% discount off the input price, with explicit breakpoints and a 30-minute minimum cache life. On Luna that means USD 0.02 per million tokens read from cache, against USD 0.25 to write them. We come back to this in the cost section, because it is what moves the needle most in customer service.
• Text and image input, text output. Relevant for WhatsApp, where the customer sends a photo of the receipt, the broken product or the router label instead of describing it.
• Five reasoning effort levels — from light a ultra—, where each step adds roughly 50% to the cost. The same model can end up three times more expensive depending on how you configure it.

In other words: choosing the model is one of three decisions. The other two are the effort level and the caching strategy. A badly configured Sol can cost more and answer worse than a well configured Luna.

Sol: the ceiling

Sol is the flagship. It is the one to reach for when the task is genuinely hard: long reasoning chains, analysis of lengthy documents, intricate code generation, research work with many steps.

The figures published by Artificial Analysis put it at 58.9 points on the intelligence index (v4.1) and at the top of the Coding Agent Index with 80 points. Broken down by benchmark: 88.8% on Terminal-Bench 2.1, 72.7% on DeepSWE, 64.6% on SWE-Bench Pro, 90.4% on BrowseComp and 62.6% on OSWorld 2.0.

Two details more interesting than the ranking:

It is token efficient. Sol solves an index task with around 15,000 output tokens, against 16,000 for the previous generation. It sounds minor, but in a model whose output costs USD 30 per million, spending fewer tokens per task is a silent discount. Sam Altman claimed a 54% improvement in token efficiency for coding tasks.

It has a multi-agent mode. The ultra effort level runs four agents in parallel and lifts Terminal-Bench from 88.8% to 91.9%, and BrowseComp to 92.2%. It is the most expensive configuration in the family, and for customer service it is almost always overkill. It is designed for novel, multi-system problems with detailed constraints.

For our terrain — service and sales conversations — Sol is the exception, not the rule. It makes sense when the agent has to reason about a complex case with many sources: a complaint with a long history, a technical diagnosis with several variables, a quote that depends on chained business rules.

Terra: the awkward middle

On paper Terra is the balanced option: most of Sol's intelligence at half the price. It scores 55 on the index, 77.4 on the Coding Agent Index, 87.4% on Terminal-Bench, 69.6% on DeepSWE, 63.4% on SWE-Bench Pro and 87.5% on BrowseComp. On paper, a very solid model.

But here comes the most counterintuitive finding of the independent analysis, and it is worth saying plainly: Terra is dominated. Artificial Analysis says so explicitly — Sol and Luna are ahead of Terra at every point on the intelligence/cost curve. Translated: for any effort level, there is another option in the same family that is more intelligent at no extra cost, or equally intelligent for less.

The arithmetic after the July 30 cut makes it worse. Terra dropped 20% (from USD 2.50 to USD 2 on input, from USD 15 to USD 12 on output); Luna dropped 80%. The result is that today Terra costs ten times more than Luna in exchange for four points of intelligence index. Four points matter in some scenarios — we come back to that — but ten times the price is a very high price to pay for them.

There is one case where Terra does hold its ground, and it is visual: on OSWorld 2.0 — the computer-control benchmark — it scores 50.2% against Luna's 45.6%. If your agent operates graphical interfaces, the gap shows. For text conversation, it is hard to justify.

Luna: the surprise

Luna is the fastest and cheapest model in the family, and the reason GPT-5.6 is worth writing about. It scores 51.2 on the intelligence index, which places it — according to Cerebras' survey — 16th out of 576 models evaluated. This is not a small model: it is a frontier model with a smaller reasoning budget.

The revealing part is the pattern of the benchmarks, not the average. Look at where Luna comes close to Sol and where it doesn't:

• SWE-Bench Pro: 62.7% vs Sol's 64.6%. Gap: 1.9 points.
• Agents' Last Exam: 50.3% vs 52.7%. Gap: 2.4 points.
• Terminal-Bench 2.1: 84.7% vs 88.8%. Gap: 4.1 points.
• DeepSWE: 67.2% vs 72.7%. Gap: 5.5 points.
• BrowseComp: 83.3% vs 90.4%. Gap: 7.1 points.
• OSWorld 2.0: 45.6% vs 62.6%. Gap: 17 points.

One detail that has to be stated before drawing conclusions: all of those scores are with the effort level set to max, the highest of the five. And in that configuration Luna takes close to 127 seconds to the first token, with an output speed of 176 tokens per second. Two minutes of silence before it starts answering is acceptable for an agent processing a case in the background, and unacceptable in a live WhatsApp conversation.

In other words: the benchmarks measure each model's ceiling, not the configuration you will actually serve with. In conversational production you will run low or medium effort, where latency drops to seconds and the scores drop too — for all three models alike. The relative comparison between Sol, Terra and Luna holds; the absolute values do not. One more reason to measure on your own traffic before deciding.

The operational reading is straightforward. On text reasoning tasks and on agentic flows with tools, Luna is two or three points off the flagship. On visual interface control, it is seventeen points off. The gap is not uniform: it is narrow exactly where customer service lives and wide in a domain most conversational agents never enter.

That is what turns the price cut into product news and not just market news.

The map: intelligence against cost

The honest way to look at this is the crossing of two axes: how much a model knows (vertical axis) against how much it costs to solve a task with it (horizontal axis, on a logarithmic scale because the range runs from cents to dollars). The dashed line is the efficient frontier: the models sitting on it have no alternative that beats them on both axes at once. Those below and to the right are paying too much for what they deliver.

What the chart shows, in three readings:

Luna sits on the frontier. At 51 points and USD 0.21 per task, there is nothing in the family — nor in the comparison with the previous generation — that delivers that intelligence for less. It is the best-value point in the whole quadrant for conversational work.

Sol also sits on the frontier, at the expensive end. At 59 points and USD 1.04 per task, it lands one point below the most intelligent model measured — Anthropic's Fable 5, at 60 — but at around a third of its cost. If you need the reasoning ceiling, Sol is currently the most efficient way to buy it.

Terra sits inside the curve, and so do GPT-5.5 and Opus 4.8. GPT-5.5 at xhigh effort scores the same 55 points as Terra but costs USD 0.99 per task against USD 0.55. Opus 4.8 scores 56 points at USD 1.78. That is the concrete meaning of "the new generation moved the frontier": what was a reasonable purchase three weeks ago no longer is.

And the reference point at the cheap end: DeepSeek V4 Pro solves the task at USD 0.04 with 44 index points. Seven points less than Luna for a fifth of the price. For classification and routing — tasks where 44 points are enough — it is a real option worth keeping on the radar.

The benchmarks, side by side

Look at the last row of the benchmark block and the second-to-last column: on Agents' Last Exam, Terra scores 50.4% and Luna 50.3%. One tenth of a point apart, ten times the price.

The price: what changed on July 30

These are the rates per million tokens, before and after the cut:

Sol did not drop. The reason OpenAI gave is that GPT-5.6's improvements let Luna and Terra absorb workloads that previously required a premium model: instead of making the ceiling cheaper, they made the floor cheaper so that more work moves down a tier. It is a market play — there is strong competitive pressure in the low-cost segment — but the practical effect for anyone building product is the same.

The figure that almost never makes the write-ups is Luna's cached input: USD 0.02 per million. Two cents. In customer service, where every model call resends the same system prompt, the same knowledge base and a history that keeps growing, caching is not a minor optimization: it is most of the bill.

What it costs to handle 1,000 conversations

The benchmarks measure coding and web-browsing tasks. Our case is a different one, so let's do the real math, with explicit assumptions so you can redo it with yours.

Assumptions for a typical service conversation:

• 8 model calls per conversation (eight back-and-forth turns).
• 4,000 input tokens per call: system prompt, knowledge base fragments and accumulated history.
• Caching active: new tokens are written to cache (write rate, 1.25× input) and on following turns they are read with the 90% discount. The first call writes 4,000 tokens; the remaining seven write 400 new ones and read 3,600.
• 150 output tokens per reply.

That gives, per conversation: 6,800 input tokens written to cache, 25,200 read from cache and 1,200 of output. Multiplied by a thousand conversations and by each model's rates:

Three conclusions that come from this table and not from the benchmarks:

The difference between models is an order of magnitude, not a percentage. Choosing Sol by default for an agent that answers frequently asked questions is not "a bit more expensive": it is twenty-five times more expensive. In an account with real volume, that single decision determines whether the agent's unit economics work.

Caching weighs more than the model in the cost structure. Without caching, those same thousand dialogues with Luna would cost close to USD 7.84 instead of USD 3.64. The system prompt and the knowledge base are identical text on every call; paying full price for them eight times per conversation is the most common waste we see.

The model's cost is almost never the project's cost. USD 36 a month for ten thousand conversations is less than an hour of human work costs in most of the markets where we operate. What defines the return on an AI agent is not the token price: it is how many of those conversations it resolves without intervention. We wrote about that in the chatbot ROI formula, and it is still the calculation that matters.

Why the cut lands differently in LATAM

An 80% discount reads the same in any language, but it does not mean the same in every market. Three reasons why this move changes more things here than in the United States or Europe.

The model cost is paid in dollars; the saving is measured against local wages. A service operation in Buenos Aires, Bogotá or Mexico City compares AI spend against the cost of an hour of human agent time in local currency, while the OpenAI invoice arrives in hard currency. That gives every rate cut an amplified effect on the decision to automate: it does not just change the margin, it changes the threshold above which the project is justified. At USD 18.22 per thousand conversations, a pilot for a mid-volume SMB was a debate. At USD 3.64, it stops being one.

It lowers the entry floor, not the ceiling. Before the cut, a small operation that wanted frontier quality had two paths: pay for premium models at low volume — high cost per conversation — or use older models and settle for less quality. Luna at its current price breaks that trade-off: an account with two thousand monthly conversations reaches a top-20 model for less than USD 8 a month. Access to frontier quality is no longer correlated with company size.

WhatsApp volume punishes expensive models. In LATAM the service conversation lives on WhatsApp, and WhatsApp is asynchronous and conversational: many short messages, not one long, well-formed query. That multiplies the number of model calls per resolved conversation — hence the eight turns in the calculation above — and makes the unit price of the token weigh far more than in a form-like or email channel. An expensive model in a messaging channel gets paid for several times per conversation. It is exactly the scenario where Luna performs best and where Sol by default is hardest to justify.

One caveat: the model price is not the only cost of the conversation. On WhatsApp you also pay Meta per template conversation, and that cost did not drop. If your operation is heavy on outbound messages, the token is a smaller fraction of your bill — we cover that in the analysis of WhatsApp chatbot pricing for businesses.

Which one to choose for customer service

The recommendation from OpenAI and from the integrators who have already tested the family converges on an escalation strategy: start at the bottom and move up when the model gets stuck. Applied to our terrain:

The pattern: Luna as the default, Sol as a justified exception, Terra almost never. And one rule that matters more than the choice of model: do not move up a tier to paper over a prompt or knowledge base problem. A Luna with well-built context beats a badly fed Sol, and costs twenty-five times less. The three knowledge layers of an AI agent define the quality of the answer far more than the model you pick.

How to measure it on your own traffic

All the evidence in this article is third-party and about tasks that are not yours. The only measurement that matters for deciding is the one that runs on your real conversations. A four-step protocol you can execute in a week:

1. Build the test set from real conversations, not invented ones. Take between 80 and 150 closed conversations from the last 30 days, and deliberately include the hard cases: messages with typos, transcribed audio, ambiguous queries, customers who ask three things in one message and those who write a single word. A clean test set lies: it will tell you any model works.

2. Run the same set with Luna and with Sol, changing nothing else. Same prompt, same knowledge base, same effort level. If you change two variables at once you will not know which one explains the difference. Save both answers for each conversation, side by side.

3. Have someone on your team compare them blind. Without knowing which model produced each answer. Three binary criteria are enough: is it correct? is it complete? would you send it to the customer as is? The bias of knowing an answer came from "the expensive model" ruins the evaluation, and it is the most common mistake when evaluating internally.

4. Decide on the tie rate, not on the average. The number that matters is the percentage of cases where Luna ties or beats Sol. If it ties on 90% of your conversations, the answer is Luna with escalation for the remaining 10%, and the saving is 25× across most of the volume. If it ties on 60%, look at what the losing cases have in common: it is almost always missing context, not model capability, and it is cheaper to fix by improving the knowledge base than by moving up a tier.

The useful output of this exercise is not a winner: it is a routing criterion. Agents handling frequently asked questions go to Luna; those handling complex cases go to Sol; and the list of cases where Luna fails becomes the improvement list for your knowledge base.

What the benchmarks don't tell you

Four warnings, because an article that only celebrates numbers is no use to you for deciding.

OpenAI's latency and cost figures are simulations, not production measurements. It is stated in the launch documentation. The Artificial Analysis numbers are independently measured, but on their own task suite.

The v4.1 index was reweighted toward agentic work. Artificial Analysis upgraded three evaluations, removed one and reweighted the set toward agent tasks. That is good news for our case — agents with tools resemble what we do more than a knowledge exam does — but it means the scores are not one-to-one comparable with rankings from earlier versions.

None of these benchmarks measures customer service in Spanish. There is no public evaluation of conversational quality in Latin American Spanish, nor of handling regional idioms, nor of tolerance for messages with typos and transcribed audio — which is 100% of real WhatsApp traffic. The scores orient you; they do not replace measuring on your own traffic.

The published scores are from maximum effort, not the one you will use. This is the most overlooked warning: a model evaluated at max can take more than two minutes to start answering. When you read any benchmark table — this one included — ask yourself which effort level it was measured at, and compare against the level your case tolerates in latency.

Today's price is not the price forever. Luna dropped 80% twenty-one days after launch. Any architecture that depends on a single point-in-time price is fragile. The right conclusion is not "migrate everything to Luna": it is build so that changing model is a configuration decision, not a project.

How to use the 5.6 family in AsisteClick

All three models are already available in the AI agent builder. In the configuration of any agent based on generative AI, the model selector now includes GPT 5.6 Sol, GPT 5.6 Terra y GPT 5.6 Luna, with a description of each one alongside. Changing model means picking an option from a list and saving: it requires no migration, no rewriting prompts, no touching the knowledge base.

That matters because of what we said above. Since the model is a parameter of the agent and not an architectural decision, you can do what we recommend: set Luna as the default, measure, and move only the agents up to Sol where the measurement justifies it. If the price moves again — and it will — the answer is changing a selector.

The same setting is exposed in the Playbooks API, so if you administer many accounts you can set the model per agent programmatically instead of going in one by one.

If you are evaluating which model suits your operation and you don't want to run the test blind, our team takes care of it: we review your real conversations, compare Luna's and Sol's answers on the same traffic and tell you, with your own data, where each one fits.

The summary that matters

The GPT-5.6 family moved the frontier of what it costs to run AI agents at production quality. Sol is the ceiling, and it is currently the most efficient way to buy the maximum of reasoning. Terra ended up in an awkward position that the arithmetic of the price cut made worse. And Luna — 51 index points, USD 0.20 per million input tokens, two cents if they are cached — is the news: a frontier model at the price of a small one, with a gap of barely two or three points exactly in the tasks where customer service lives.

The practical decision is not which is the best model. It is which is the cheapest model that handles your conversation well, and how you make sure you can change it when the market moves again. Because at this rate, it will.

Frequently asked questions

What are Sol, Terra and Luna?

They are the three models in OpenAI's GPT-5.6 family, introduced on July 9, 2026. Sol is the flagship (maximum reasoning capability), Terra the mid tier and Luna the fastest and cheapest. All three share architecture, API capabilities and a 1.05 million token context window; what varies is the compute budget devoted to reasoning and, as a result, the price.

What does GPT-5.6 Luna cost after the price cut?

Since July 30, 2026, Luna costs USD 0.20 per million input tokens and USD 1.20 per million output, an 80% drop from the USD 1 and USD 6 at launch. Cached input costs USD 0.02 per million. Terra dropped 20% (USD 2 and USD 12) and Sol stayed at USD 5 and USD 30.

Is Luna enough for a customer service agent?

In most cases, yes. On text reasoning benchmarks and on agentic flows with tools, Luna lands between 1.9 and 2.4 points of Sol: on Agents' Last Exam it scores 50.3% against 52.7%, and on SWE-Bench Pro 62.7% against 64.6%. The big gap appears in graphical interface control (45.6% against 62.6% on OSWorld 2.0), which is not what a conversational agent does. The recommendation is to start with Luna and escalate to Sol only the cases where measurement on your own traffic justifies it.

Why choose Luna over Terra?

Because Terra is dominated on the intelligence/cost curve: for any effort level there is another option that is more intelligent at no extra cost, or equally intelligent for less. After the July 30 cut, Terra costs ten times more than Luna for four points of index, and on Agents' Last Exam the difference between them is one tenth of a point (50.4% against 50.3%). Terra is mainly justified on visual tasks, where it scores 50.2% against 45.6% on OSWorld.

What does it cost to handle 1,000 conversations with each model?

With assumptions of eight turns per conversation, 4,000 input tokens per turn with caching active and 150 output tokens per reply: Luna costs around USD 3.64, Terra USD 36.44 and Sol USD 91.10. That is a 25× difference between the cheapest and the most expensive in the family. Enabling prompt caching is what affects the bill most: without caching, those same thousand conversations with Luna would cost around USD 7.84.

Can I use GPT-5.6 in AsisteClick already?

Yes. All three models are available in the model selector of the AI agent builder and also via the Playbooks API. Changing an agent's model means picking an option and saving: it requires no migrating the agent, no rewriting prompts and no touching the knowledge base.

Keep reading

• GPT-5.4 vs Mini vs Nano: which model to use in 2026 — the previous generation's comparison, useful for seeing how far the frontier moved in a few months
• Chatbot NLP vs GPT vs hybrid: which technology to choose — to decide first which engine your agent needs
• The 3 knowledge layers of an AI agent — why context defines quality more than the model does
• The chatbot ROI formula — how to calculate the return beyond cost per token
• Why AI agents fail in customer service — the process causes no model fixes