GPT-5.6 Luna & Terra API Review 2026: Pricing, Speed Tiers & Subagent Routing

Research-based — not hands-on tested. Our scores are editorial judgements compiled from vendor documentation, published pricing and independent user reports. How we review.

See how these tiers stack up against Gemini 3.6 Flash and Claude Haiku 4.5, our three-way speed-tier comparison, or every current endpoint in the LLM API pricing reference.

1. Launch, access, and a recent price cut

GPT-5.6 launched on June 26, 2026, and the flagship Sol tier initially shipped as a limited, government-vetted preview with no public pricing. That restriction has since lifted: Sol’s own API documentation now lists standard pricing of $5.00 per million input tokens and $30.00 per million output tokens, with no waitlist or access gate mentioned. Terra and Luna were the production-facing tiers from the start, and on July 30, 2026 OpenAI cut Luna’s price by roughly 80% and Terra’s by roughly 20% — a change large enough that any pricing published before that date, including earlier coverage on this site, is now stale. The figures below are current as of this review’s last update.

We focus this review on Luna and Terra rather than Sol because they’re the tiers most API integrations actually route traffic to day-to-day: Sol’s $30.00 output rate is roughly 2.5× Terra’s and 25× Luna’s, which only makes sense for the hardest planning and reasoning tasks in a routing pipeline, not the bulk of the traffic.

2. Current pricing (retrieved August 2026)

Both tiers carry the standard cached-input discount — roughly 90% off the uncached input rate — so a pipeline that reuses the same system prompt or tool schema on every call pays the cached rate for most of its traffic, not the sticker price above.

3. Real-world coding performance: the DeepSWE benchmark

DeepSWE is a 113-task benchmark of original, long-horizon software-engineering problems spanning five languages, graded by hand-written verifiers rather than static patch-matching — a harder, more contamination-resistant test than most agentic coding leaderboards. On the public DeepSWE leaderboard, the GPT-5.6 family currently occupies the top three positions:

What stands out is how close Luna — the cheapest tier by a wide margin — sits to Terra and Sol on this specific benchmark. For workloads that are mostly about resolving well-scoped coding tasks rather than deep multi-step planning, Luna’s DeepSWE score makes a genuine case for starting there before paying for Terra.

4. Subagent routing pattern

The practical pattern most teams land on: route high-frequency, low-risk calls — classification, log parsing, routine tool calls — to Luna by default, and escalate to Terra only when a call is flagged as high-risk (a destructive action, a multi-file refactor, or a planning step that needs more careful reasoning). At Terra’s $12.00 output rate versus Luna’s $1.20, that escalation gate is where most of the budget control in a Luna/Terra pipeline actually happens.

A concrete version of this pattern shows up in support-ticket triage pipelines: every incoming ticket runs through Luna first for classification and routing-tag assignment, and only the tickets Luna flags as ambiguous, technical, or high-priority get handed to Terra for a fuller response draft. Because Luna’s output price is a tenth of Terra’s, this two-stage design keeps the bulk of the traffic on the cheap tier while still getting Terra-level judgment exactly where it earns its cost — rather than paying Terra rates for volume that never needed the extra reasoning in the first place.

5. Final verdict

GPT-5.6 Luna is now priced aggressively enough — and scores well enough on DeepSWE — to be the default tier for high-volume API work, with GPT-5.6 Terra reserved for tasks that genuinely need the extra reasoning budget. Given how recently OpenAI moved this pricing, re-check the current rate at OpenAI’s own pricing page before committing a production budget to either tier.

Pricing per OpenAI’s GPT-5.6 Luna and Terra API documentation and OpenAI’s July 30, 2026 pricing announcement, retrieved 2026-08-04. DeepSWE scores per the public DeepSWE leaderboard, retrieved 2026-08-04. Both are subject to change — confirm current figures at platform.openai.com/docs/pricing before budgeting.