GPT-5.6 Sol, Terra or Luna: choose without blowing the budget

OpenAI no longer presents its newest model as one identifier with several speed settings. The GPT‑5.6 family contains Sol, Terra and Luna: three levels of capability, latency and price, supplemented by reasoning-effort controls and agentic modes.

This structure makes the catalogue more flexible but moves complexity onto product teams. Choosing Sol everywhere maximises the bill without guaranteeing the best product. Choosing Luna everywhere can make difficult workflows brittle. The right unit of selection is no longer the whole application: it is the task, its risk and the cost of failure.

The three models at a glance

These rates do not calculate a workflow's full cost. A model that costs half as much per token but produces longer answers, retries a task or makes excessive tool calls may cost more per useful result. Sol can conversely be economical when it completes in one pass a job that requires several attempts elsewhere.

Sol is not ChatGPT's new default mode

In standard ChatGPT conversations, GPT‑5.5 Instant remains the fast default. GPT‑5.6 Sol powers Medium, High and Extra High reasoning levels depending on the plan. Sol Pro targets difficult, long-running work.

Terra and Luna cannot be selected in a standard ChatGPT conversation. OpenAI makes them available through other surfaces, including Codex, ChatGPT Work and the API, with access depending on the plan. That distinction prevents a common false expectation: three names in a launch post do not mean every user will find all three in the same picker.

Rollout is gradual. An eligible account may not display Sol immediately, and Business or Enterprise administrators can control model access in a managed workspace.

Start with the cost of an error

A low-risk task with an easy validator should start on Luna. Ticket classification, language detection, five-field extraction or headline rewriting has an expected answer and a simple check. Throughput and price dominate.

Terra fits workflows whose difficulty varies: answering from a document base, preparing a limited code change, reviewing a contract against a rubric or coordinating a few tools. It is the natural candidate for an initial production test, provided failures are measured by category.

Sol is justified when a task combines multiple sources, long planning, tools, ambiguity and a high cost of error. A complex migration, technical investigation, scientific analysis or final agent synthesis may benefit from its additional capacity.

The useful rule is straightforward: use the least expensive model that reaches the required quality and reliability, then escalate when observable signals show that it is insufficient.

Routing is better than one universal choice

An efficient architecture can divide the work:

1. Luna classifies the request, extracts constraints and rejects incomplete cases;
2. Terra handles the normal flow with tools and a validated output format;
3. Sol takes over uncertain, long or sensitive cases;
4. a deterministic or human control validates the result before an irreversible action.

Routing should not rely solely on a model's intuition. Use observable rules such as context length, tool count, data category, uncertainty score, validator failure, retry count or the financial value of an action.

This approach also protects capacity. Ordinary requests do not occupy the slowest model, while difficult cases receive an explicit escalation path.

Benchmarks are a vendor starting point

OpenAI reports gains across coding, knowledge work, browsing, computer use, cybersecurity and science. The company says Sol obtains stronger results with fewer tokens in several evaluations and positions Terra and Luna as competitive at lower costs.

Those figures come from the provider and combine public benchmarks, internal evaluations and partner reports. They cannot directly predict success on a particular repository, document collection or user population.

A local evaluation should preserve the same inputs, tools, limits, validators and criteria for every model. At minimum, measure:

• complete task success;
• factual and structured-output errors;
• input and output tokens;
• median and 95th-percentile latency;
• tool calls and unnecessary loops;
• human interventions;
• cost per accepted result rather than cost per request.

Twenty representative cases provide an initial signal. A gradual rollout on real traffic can then reveal requests the test set did not anticipate.

Caching changes the cost of long contexts

GPT‑5.6 introduces explicit cache breakpoints and an announced minimum cache life of thirty minutes. Cache reads receive a 90% input discount, while writes are billed at 1.25 times the uncached input rate.

Caching is valuable when many requests share a long stable base: instructions, tool schemas, documentation or reference material. It offers less value when the prefix changes on every call or the data is never reused during the available window.

Place stable material before variable content, monitor the cache-hit rate actually achieved and avoid inflating context simply because it is discounted. A token read is still processed and may affect latency and model attention.

Programmatic Tool Calling and multi-agent

In the Responses API, OpenAI offers Programmatic Tool Calling: the model can write and run an in-memory program that coordinates tools and processes intermediate results. The aim is to reduce round trips in which every small tool result returns to the model context.

Multi-agent, initially in beta, can run subagents in parallel and synthesise their work. Parallelism may accelerate research made of independent branches, but it also multiplies calls, permissions and failure points.

Do not use several agents for a straightforward sequential task. Reserve them for genuinely decomposable work, impose a global budget, limit each role's tools and record the evidence used in the synthesis.

Cyber capability requires safeguards

OpenAI publishes a system card and says GPT‑5.6 substantially improves vulnerability discovery and remediation. The company also reports stronger controls for higher-risk biological and cybersecurity requests.

For an engineering team, a more capable model is not a reason to grant direct access to production secrets, systems or internal networks. Analysis should run in an isolated environment with filtered repositories, temporary identities, allowlisted commands and review of proposed fixes.

Additional refusals and checks may also affect legitimate work. A security evaluation should therefore measure both unsafe responses and unjustified blocks without trying to bypass protections.

A pragmatic migration strategy

Do not replace one model name throughout an application in a single change. First record the model, effort, token count, latency, tools and validator outcome for every workflow.

Then test Luna on high-volume deterministic tasks, Terra on the main flow and Sol on a sample of complex cases. Compare cost per success and define escalation rules. Retain a rollback path until error distributions are stable.

GPT‑5.6 provides more choice, not one universal answer. The team that benefits most will not be the one selecting the strongest model everywhere, but the one matching each capability level with a measured task, limited permission and visible budget.