Reviews

Analysis: GPT-5.6 Sol, Terra and Luna — OpenAI's three-tier family, what it costs and where each model fits

OpenAI has made its GPT-5.6 family generally available after a limited preview. We break down the three tiers, launch API pricing from $1 to $5 per million input tokens, the new developer features and who each model suits.

Abstract illustration of concentric orbital rings around a bright core, in orange and red.
Abstract illustration of concentric orbital rings around a bright core, in orange and red.

OpenAI made its GPT-5.6 model family broadly available on Thursday 9 July 2026, across ChatGPT, its Codex coding tool and its API. The release follows a limited preview that began on 26 June. The family has three models: GPT-5.6 Sol, the flagship; GPT-5.6 Terra, a middle tier; and GPT-5.6 Luna, a low-cost model for high-volume work.

This analysis is based on OpenAI’s developer documentation and reporting of the company’s announcements, including its published benchmark claims; we have not tested the models hands-on.

A new way of naming tiers

Earlier GPT-5 releases used suffixes such as “mini” and “nano” to mark smaller models. GPT-5.6 replaces them with names. OpenAI’s developer documentation describes Sol as a “flagship model for complex professional work” that roughly corresponds to the unsuffixed tier of earlier families, Terra as the rough equivalent of the mini tier, and Luna as the rough equivalent of nano. In the API, the plain gpt-5.6 alias routes to Sol.

The practical effect is that the version number now tells you the generation, and the name tells you the size. Whether that proves clearer for customers than the old suffixes will depend on how consistently OpenAI applies it in future releases.

Specs and launch pricing

All three models share the same context and output limits, which means the choice between them is about quality and cost rather than how much text they can handle.

GPT-5.6 Sol GPT-5.6 Terra GPT-5.6 Luna
API model ID gpt-5.6-sol gpt-5.6-terra gpt-5.6-luna
Input (per million tokens) $5.00 $2.50 $1.00
Output (per million tokens) $30.00 $15.00 $6.00
Context window 1.05M tokens 1.05M tokens 1.05M tokens
Maximum input 922K tokens 922K tokens 922K tokens
Maximum output 128K tokens 128K tokens 128K tokens
Knowledge cutoff 16 February 2026 16 February 2026 16 February 2026
Inputs and outputs Text and images in, text out Same Same

Prices are the launch rates as reported by TechCrunch. All three models support reasoning effort settings of none, low, medium (the default), high, xhigh and max.

For comparison, Anthropic lists its Claude Fable 5, released a month earlier, at $10 input and $50 output per million tokens, so Sol’s list price is half that on input and 40% lower on output.

What’s new for developers

OpenAI’s API changelog for 9 July lists several additions alongside the models:

  • Programmatic tool calling, a new way for the model to call tools.
  • Explicit prompt caching controls, letting developers direct how caching behaves.
  • Persisted reasoning, which keeps the model’s reasoning across requests.
  • A new “max” reasoning effort, above the previous highest setting.
  • Pro mode, offered alongside the family.
  • Multi-agent orchestration, in beta in the Responses API.
  • Original-resolution image input, via “original” or “auto” image detail settings.

The models are available through the Responses, Chat Completions and Batch endpoints.

What OpenAI claims

OpenAI’s headline claim is about efficiency rather than raw scores. Chief executive Sam Altman told CNBC that Sol is 54% more token-efficient on agentic coding tasks, and that it is “as good or better” than competing models. He framed the point in terms of cost: “Every enterprise now is thinking about spend and the value they’re getting in exchange for AI.”

CNBC reported that OpenAI’s blog post called Sol its “strongest model yet”, and TechCrunch reported that OpenAI described it as its “best coding model yet.” These are vendor-reported claims; OpenAI’s comparisons with rival models use its own choice of benchmarks and settings.

How the release unfolded

The models were first announced on 26 June but made available only to a “small group of trusted partners” at the request of the US government, according to both CNBC and TechCrunch. OpenAI said at the time: “We don’t believe this kind of government access process should become the long-term default.” The broad release on 9 July ended that restricted period.

Who each model is for

Sol is positioned by OpenAI for complex professional work. In practice that points to long agentic coding sessions, multi-step analysis and tasks where getting the answer right first time matters more than per-token price. OpenAI’s efficiency claim is the key argument here. If Sol completes a task using fewer tokens, its cost per finished job could be lower than the list price suggests.

Terra is pitched as the balance between intelligence and cost. At half Sol’s price on both input and output, it is the natural candidate for production workloads such as customer-facing assistants and document processing, where Sol’s extra capability may not be needed on every request.

Luna is described by OpenAI as being for efficient, high-volume workloads; typical examples would be classification, extraction and routing. At a fifth of Sol’s input price, it is the model to test first for pipelines that make millions of calls.

What it means

Our take: GPT-5.6 is a release about price-performance as much as capability. OpenAI is asking customers to judge models by the cost of completing a task, not the cost of a token, and its headline efficiency figure is built to support that argument.

That framing is reasonable, but it moves the burden of proof to the buyer. Token use varies with the task, the reasoning effort setting and the tooling around the model. Teams considering a switch should run their own workloads across all three tiers, at more than one effort level, before concluding which offers the best value.

Sources

See All

More on OpenAI

See All

Type to search. ↑ ↓ to navigate, ↵ to open.