MarkdownModels

GPT-5.6 Sol, Terra and Luna: portfolio routing economics

Reviewed: 18/09/2026 · Status: Current public API information · Audience: Platform engineering, product owners and FinOps

OpenAI's GPT-5.6 family creates a clean three-tier routing ladder. Sol targets complex professional work, Terra balances intelligence and cost, and Luna targets high-volume workloads. All three list a 1.05M context window, 128K maximum output and configurable reasoning effort. The TokenOps opportunity is to treat them as a portfolio—not three isolated model choices.

Current rate card

Model	Intended tier	Input / 1M	Cached input / 1M	Output / 1M
GPT-5.6 Sol	Complex professional work	$4.00	$0.40	$20.00
GPT-5.6 Terra	Balanced intelligence and cost	$2.00	$0.20	$12.00
GPT-5.6 Luna	Cost-sensitive volume	$0.20	$0.02	$1.20

Prompts above 272K input tokens are listed at 2× input and 1.5× output for the full request. Cache writes are 1.25× uncached input. Sol's published rate is promotional through at least 21/11/2026, so forecasts need a review event rather than treating it as permanent.

Why output control dominates

For a 10K-input, 2K-output request without caching:

Model	Input cost	Output cost	Total
Sol	$0.04	$0.04	$0.08
Terra	$0.02	$0.024	$0.044
Luna	$0.002	$0.0024	$0.0044

Luna is about 18× cheaper than Sol for this shape. But routing decisions must use accepted-outcome economics. A cheap model that creates retries or human corrections can erase its unit-price advantage.

A production cascade

1. Luna: classify, extract, normalize, route, draft
2. Validate: schema, confidence, business rules, grounded citations
3. Terra: retry ambiguous or medium-complexity cases
4. Validate again
5. Sol: escalate only high-value failures or pre-approved task classes

Do not let the models grade themselves as the sole quality control. Use deterministic checks where possible and a separately calibrated evaluator for subjective tasks.

Routing score

A simple decision score makes policy auditable:

complexity score = domain risk + ambiguity + tool depth + context size + reversibility

Score each dimension from 0–2. Suggested starting policy:

0–3: Luna
4–6: Terra
7–10: Sol
Regulated or irreversible action: human approval regardless of score

Tune thresholds with production outcomes, not benchmark headlines.

Cache-aware prompt design

Place stable content first: policy, schemas, tool definitions, long reference material. Put volatile user data and timestamps last. Track:

prefix hash and version;
cache writes and reads;
hit rate by route and tenant;
cache-adjusted input cost;
invalidations caused by prompt changes.

At a 90% cache hit ratio, 100K repeated input tokens on Sol have an average read cost near $0.076 per request after a single $0.50 cache write amortized across ten requests, versus $0.40 each without caching. Real savings depend on provider semantics and traffic reuse.

Migration experiment

Run a shadow evaluation across representative tasks:

Metric	Guardrail
Accepted outcome rate	No more than agreed quality loss
Cost per accepted outcome	Must improve, not merely cost per call
p95 latency	Within product SLO
Escalation rate	Stable and explainable
Retry amplification	No hidden loop growth
Long-context incidence	Below budgeted threshold

Start by moving deterministic and reversible work to Luna. Move medium-risk generation to Terra only after the low tier is stable. Keep Sol capacity for the cases where additional capability changes the business result.

Contract and forecast controls

Version the rate card and store its effective date.
Add an alert before Sol promotional pricing expires.
Model Standard, Batch/Flex and Fast-mode scenarios separately.
Include tool-call charges and regional processing uplift.
Forecast by workload shape: input, cached input, output and calls—not requests alone.
Re-run the routing evaluation after every model snapshot change.

Sources

Pricing and promotional terms change. Verify the provider pages before implementation.