GPT-5.6 Sol vs Terra vs Luna: API Routing Guide for Production Teams
OpenAI's July 2026 GPT-5.6 launch changed the routing problem for production AI teams. Sol, Terra, and Luna are not three random names for the same model. They are three cost-capability tiers that should map to different workloads, evaluation thresholds, and fallback rules.
Quick Answer
| Production situation | Start with | Escalate or fallback to | Why this is the practical route |
|---|---|---|---|
| Complex coding agents, code review, security analysis, research agents | Sol | Terra for lower-cost repeat runs | The value of a correct answer can outweigh the higher token price. |
| Product assistants, internal copilots, mixed chat and tool workflows | Terra | Sol for hard cases, Luna for low-risk volume | Terra is the best first default when the traffic mix is not yet segmented. |
| Summaries, classification, extraction, rewriting, batch back-office jobs | Luna | Terra when acceptance drops | Luna is the cost-control tier, but only after quality is proven on your prompts. |
| Unknown new workflow | Terra | Test Sol and Luna in parallel during evaluation | It gives a balanced baseline before you optimize. |
| Customer-facing rollout | Terra with controlled Sol/Luna routing | GPT-5.5 or another verified model as rollback | Production needs observability and rollback, not a one-line model swap. |
Confirmed Facts to Anchor the Decision
| Fact | What OpenAI states | Production implication on EvoLink |
|---|---|---|
| Release status | GPT-5.6 is a generally available model family announced on July 9, 2026. | Treat it as a current production candidate, not a rumor or preview-only item. |
| Tiers | Sol is the flagship tier, Terra is balanced, and Luna is the most cost-efficient tier. | Route by workload value instead of using one tier for every request. |
| Model IDs | Use gpt-5.6-sol, gpt-5.6-terra, or gpt-5.6-luna. | Keep model IDs in configuration, not hard-coded in business logic. |
| Alias behavior | On OpenAI's direct API, gpt-5.6 routes to gpt-5.6-sol. | EvoLink currently requires the exact Sol, Terra, or Luna model ID. |
| Standard short-context pricing | Sol: $5 input / $30 output per 1M tokens. Terra: $2.50 / $15. Luna: $1 / $6. | Use list price as a planning anchor, then verify EvoLink route pricing and account terms. |
| Standard long-context pricing | Above 272K prompt tokens, the full request uses 2x input and cache rates and 1.5x output rates: Sol $10 / $45, Terra $5 / $22.50, Luna $2 / $9 per 1M tokens. | Long-context workflows need a separate cost estimate. |
| Reasoning controls | GPT-5.6 supports reasoning effort up to max, pro mode, persisted reasoning, and Programmatic Tool Calling in the Responses API. | Evaluate tier, reasoning effort, caching, and tool behavior together. |
These facts make the article a routing guide, not a leaderboard. A leaderboard asks which model wins. A routing guide asks which model should handle which traffic safely and profitably.
Who This Guide Is For
This guide is for teams that already run AI features in production or are close to doing so: coding agents, customer support agents, research assistants, internal copilots, document workflows, and high-volume text automation.
The Routing Decision Tree
Use this decision tree before choosing a model ID.
- If a wrong answer can create expensive engineering, legal, financial, or customer harm, start with Sol.
- If the workflow mixes reasoning, chat, tools, and normal user requests, start with Terra.
- If the task is repetitive, schema-bound, easy to grade, and high-volume, test Luna.
- If you cannot describe the acceptance test, do not optimize for the cheapest tier yet.
- If a workflow already runs on GPT-5.5 or GPT-5.4, keep the old route as the baseline until GPT-5.6 wins on real traffic.
The mistake is treating Sol, Terra, and Luna as a preference setting. They should be a routing policy.
Sol vs Terra vs Luna at a Glance
| Tier | Best first use | Do not use it when | Acceptance signal | Cost posture |
|---|---|---|---|---|
| Sol | High-value reasoning, coding agents, hard research, multi-step tool work | The task is routine, easy to grade, and repeated at scale | It reduces failed tasks, retries, or human rework enough to justify spend | Quality-first |
| Terra | Default production candidate for mixed workloads | The workflow is clearly either frontier-hard or cheap-volume | It matches Sol on most user-visible cases while lowering cost | Balanced |
| Luna | High-volume summaries, extraction, classification, transformations | Failure is hard to detect or downstream harm is high | It passes the same acceptance suite as Terra for that workflow | Cost-first |
A production team may use all three in the same product. The point is to stop one model from carrying every traffic shape.
Workload Routing Matrix
| Workflow | Recommended route | Evaluation metric | Rollout note |
|---|---|---|---|
| Code generation agent | Sol first, Terra comparison | Test pass rate, edit distance, reviewer acceptance, tool-call success | Keep Sol for high-risk repo changes and route simple edits down only after data proves it. |
| Code review or security triage | Sol or Terra depending on severity | True positive rate, false positive rate, remediation usefulness | Use severity, affected surface, and confidence to decide escalation. |
| Customer support assistant | Terra default | Resolution rate, escalation rate, hallucination rate, latency | Use Sol for complex account or policy cases; use Luna only for internal drafts or summaries. |
| RAG over long documents | Terra first, Sol for hard synthesis | Citation accuracy, answer completeness, context cost | Long-context pricing and cache behavior matter as much as base model tier. |
| Structured extraction | Luna candidate, Terra fallback | Field accuracy, schema validity, downstream correction rate | Luna is attractive when the schema is stable and errors are cheap to catch. |
| Batch summarization | Luna candidate | Summary usefulness, compression ratio, retry rate | Batch and Flex pricing may change the economics, but verify latency expectations. |
| Strategic research or planning | Sol first | Decision usefulness, evidence handling, evidence traceability | Do not optimize for cheapest output when the work directs product or capital decisions. |
| Internal copilots | Terra default | Task completion, user correction rate, tool success | Use user role and action risk to decide when to escalate. |
Cost Planning: Do Not Stop at Token Price
OpenAI list pricing is useful, but production cost should be measured at the successful-task level.
| Cost layer | What to measure | Why it changes the Sol/Terra/Luna decision |
|---|---|---|
| Input tokens | Prompt, retrieved context, conversation history | Long context can move a workflow into a different cost band. |
| Output tokens | Final answer and visible intermediate outputs | A cheaper model that rambles may not be cheaper in practice. |
| Cache reads and writes | Reused prompt/context segments and cache creation | GPT-5.6 pricing includes cached input and cache-write lines that matter for repeated workflows. |
| Retries | Automatic retries, repair calls, validation loops | Lower-quality first attempts can erase token-price savings. |
| Fallback calls | Calls to Sol, GPT-5.5, or another route after failure | Routing must include the expected fallback rate. |
| Human review | Manual QA, support escalation, engineering cleanup | The model is only cheaper if it lowers total operational work. |
A practical formula is:
successful_task_cost = first_call_cost + cache_cost + retry_cost + fallback_cost + review_costUse Luna when this total goes down without hurting acceptance. Use Sol when it reduces failures enough to lower total cost or protect revenue. Use Terra when you need the strongest default before segmentation is mature.
Reasoning Effort Is a Separate Control
Do not confuse model tier with reasoning effort. OpenAI's GPT-5.6 guidance says teams can choose Sol, Terra, or Luna and also set reasoning effort intentionally. That means you should test combinations, not only model names.
| Control | Production question | Practical starting point |
|---|---|---|
| Model tier | How much base capability does the workload need? | Terra default; Sol for hard/high-value; Luna for accepted volume. |
| Reasoning effort | How much exploration and verification should the model spend? | Start medium, lower for latency-sensitive tasks, raise only when quality improves. |
| Pro mode | Does the workflow need a higher-effort mode without changing model slug? | Reserve for quality-first cases and test billing impact. |
| Persisted reasoning | Can multi-turn quality and cache efficiency improve by preserving reasoning items? | Test on agents and long workflows, not on one-off transformations. |
| Programmatic Tool Calling | Can tool-heavy workflows avoid passing every intermediate result through the model? | Useful for agents that filter, inspect, and adapt across tool calls. |
This is where many thin comparison posts fail. The real production choice is not just Sol versus Terra versus Luna. It is Sol/Terra/Luna plus reasoning effort, cache behavior, tool use, fallback policy, and route observability.
Evaluation Plan Before Migration
A GPT-5.6 migration should start with an evaluation pack, not a pricing spreadsheet.
| Step | What to do | Minimum output |
|---|---|---|
| Build the prompt set | Select 30-100 real prompts per high-value workflow. | A versioned eval set with easy, normal, and hard cases. |
| Define acceptance | Write pass/fail rubrics before testing. | Quality, safety, latency, cost, and repair thresholds. |
| Run side-by-side | Compare Sol, Terra, Luna, and current baseline. | Same inputs, same tools, same grading path. |
| Segment traffic | Mark requests as premium, default, volume, or fallback. | A route map that engineering can implement. |
| Canary in production | Start with a limited percentage or internal cohort. | Live latency, retry, cost, and complaint data. |
| Expand or rollback | Increase traffic only after metrics stabilize. | A documented decision with owner and rollback trigger. |
For subjective outputs, add a human review sample. For structured tasks, add automated graders. For agent workflows, include tool success, not only final answer quality.
Migration From GPT-5.5 or GPT-5.4
Do not replace GPT-5.5 just because GPT-5.6 exists. Use the previous model as a baseline until the new route wins on the metrics that matter.
A safe migration pattern:
- Keep the existing GPT-5.5 or GPT-5.4 route unchanged.
- Add a GPT-5.6 evaluation route group.
- Compare Terra against the existing default.
- Compare Sol on high-value or hard-failure prompts.
- Compare Luna on high-volume, easy-to-grade tasks.
- Move traffic by workflow, not by global model replacement.
- Keep rollback active until support, cost, and latency stabilize.
This is especially important if your prompts were tuned around a previous reasoning style. A better model can still need prompt and scoring adjustments.
EvoLink Implementation Pattern
On EvoLink, keep GPT-5.6 routing as configuration. The shape below is a planning example, not an API reference:
route_group: gpt-5-6-production
default: gpt-5.6-terra
premium_reasoning: gpt-5.6-sol
high_volume: gpt-5.6-luna
fallback: gpt-5.5
rollback_owner: ai-platformThis gives product and platform teams a clean control surface:
- Product can decide which workflows deserve Sol.
- Engineering can keep one integration path instead of three separate SDK paths.
- Finance can compare successful-task cost instead of raw token price.
- Support can trace which route produced an answer when a customer reports a problem.
Common Mistakes
| Mistake | Why it is risky | Better approach |
|---|---|---|
Hard-coding gpt-5.6 everywhere | OpenAI's alias routes to Sol, while EvoLink requires exact tier IDs. | Use explicit tier IDs and config-driven routing. |
| Choosing Luna only because it is cheaper | Cheap failed tasks can cost more after retries and repair. | Require workflow-level acceptance before shifting volume. |
| Treating Terra as a weak compromise | Terra may be the best production default for mixed traffic. | Use Sol only where measured value is higher. |
| Comparing one demo prompt | Demos hide edge cases, latency, and retry behavior. | Use representative prompt sets and canary traffic. |
| Ignoring long-context pricing | Requests above 272K prompt tokens move the full request into the long-context price band. | Separate short-context and long-context estimates. |
| Removing fallback too early | New model rollouts can expose prompt, latency, or product issues. | Keep GPT-5.5 or another verified fallback until metrics stabilize. |
Recommended Policy for Production Teams
Use this policy as a starting point:
| Policy area | Recommendation |
|---|---|
| Default route | gpt-5.6-terra for mixed production workloads after evaluation. |
| Premium route | gpt-5.6-sol for complex coding, hard research, high-risk analysis, and escalations. |
| Volume route | gpt-5.6-luna for validated summaries, extraction, classification, and batch transformations. |
| Alias use | Use an exact tier ID on EvoLink; OpenAI's direct gpt-5.6 alias resolves to Sol. |
| Fallback | Keep GPT-5.5 or another verified model during the migration window. |
| Review cadence | Re-check route quality after prompt changes, product launches, pricing changes, or traffic mix shifts. |
Sources
FAQ
Is GPT-5.6 Sol better than Terra and Luna?
Sol is the flagship tier, but better depends on the task. Use Sol when the cost of a wrong or weak answer is high enough to justify the premium route.
Should Terra be the default GPT-5.6 model?
For many production teams, yes. Terra is the strongest default candidate for mixed workloads because it balances quality, latency, and cost before you have precise segmentation.
When should I use Luna?
Use Luna for high-volume tasks that are easy to grade and have passed acceptance tests: summaries, extraction, classification, transformations, and low-risk drafts.
Is gpt-5.6 the same as gpt-5.6-sol?
gpt-5.6 alias routes to gpt-5.6-sol. EvoLink currently requires the exact gpt-5.6-sol, gpt-5.6-terra, or gpt-5.6-luna model ID.What prices should I use for planning?
Use OpenAI standard short-context list prices as the public anchor: Sol $5/$30, Terra $2.50/$15, Luna $1/$6 per 1M input/output tokens. Verify current EvoLink route pricing before customer commitments.
Does long context change the decision?
Yes. Above 272K prompt tokens, OpenAI applies long-context rates to the full request: 2x for input and cache roles and 1.5x for output. Document-heavy workflows therefore need their own estimate.
How should I evaluate GPT-5.6 against GPT-5.5?
Run side-by-side tests on representative prompts, then compare quality, latency, retry rate, tool success, and successful-task cost. Keep GPT-5.5 as fallback until GPT-5.6 wins in production metrics.
Should I use pro mode or max reasoning by default?
No. Treat them as quality-first controls for difficult workflows. Test billing, latency, and result improvement before turning them on broadly.
Do I need separate SDK integrations for Sol, Terra, and Luna?
Usually no. A single EvoLink integration with route configuration is easier to operate than three separate code paths.
What is the safest rollout pattern?
Start with evaluation, then canary traffic, then workflow-by-workflow expansion. Keep explicit model IDs, observability, and rollback active until metrics stabilize.
