GPT-6 Luna vs GPT-5.6 Luna: The Same Name, Half the Price, and a Worse Deliverable

Two models, one word in common, and an independent score that is effectively identical. GPT-6 Luna reaches 37.26 on the Artificial Analysis Intelligence Index; GPT-5.6 Luna reached 37.32. A 0.07-point difference is not a measurement, it is noise, and it is the single most important fact about this pairing. What separates the two models is not capability — it is $0.0681 per completed index task against $0.1783, and a set of knowledge-work regressions that the price cut does not mention.

The dates matter for reading the numbers. GPT-5.6 Luna shipped on July 9, 2026 at $0.20 per million input tokens and $1.20 per million output. GPT-6 Luna shipped on September 22, 2026 at $0.10 and $0.50 — exactly half the input rate and 58% off the output rate, which OpenAI has described as permanent rather than promotional. That is a genuine cut, and it is also the smaller half of the story. The larger half is that the newer model is a different model, not the same one repriced.

What a migration actually buys, in numbers

Line the two up on the dimensions that decide a migration and the shape is uneven. Some rows improve, some regress, and one improves a great deal.

• Price — GPT-6 Luna $0.10 in / $0.50 out per million tokens vs GPT-5.6 Luna $0.20 / $1.20. Half on input, 58% off on output.

• Cached input — $0.01 per million against $0.02. The same 90% discount on a halved base, so a repeated prefix costs half of what it cost in July.

• Cost per completed task — $0.0681 against $0.1783 on the same suite. A 62% reduction, which is larger than the token-price cut because the task mix is not identical.

• Output tokens per task — 50,537 against 41,235. The new model is 23% chattier per finished job, and 78% of its output is reasoning that never appears in the response.

• Context window — 1,050,000 tokens against 1,000,000. The same practical ceiling; both cap output at 128,000 tokens.

• Abstention — a 76.7% hallucination rate on AA-Omniscience against 92.6%. This is the largest single improvement in the set, and it is not a knowledge gain: accuracy moves from 42.7% to 43.8%, essentially flat. The newer model declines instead of inventing, which is a behavioural change rather than a capability one.

• Knowledge-work deliverables — AA-Briefcase v1.1 falls from 1,345.38 Elo to 1,299.23, and GDPval-AA v2.1 falls from 1,443.14 to 1,367.09. Both down, by roughly 46 and 76 points.

• Coding and long-horizon work — Terminal-Bench 4.0 goes from 11.6% to 12.6% and SciCode from 53.6% to 54.6%, the two rows here that rise. Against that, the Coding Agent Index falls from 43 to 41 and the SWE-style agentic evaluations move the same way.

• AutomationBench-AA — 53.2% against 50.2%. Up, and by more than any other agentic measure in the set.

The regression is in the artefact, not the reasoning

The pattern in those last two rows is worth naming, because it is the whole decision. The new model is better at single-step agentic actions and worse at handing back a finished document. Artificial Analysis attributes the knowledge-work decline to presentation quality and deliverables that skipped required rubric elements rather than to factual or reasoning failures — which is precisely the kind of regression that survives a smoke test and fails a review.

Put the two facts together and the migration splits cleanly by what consumes the output. If your pipeline reads structured output — JSON, a classification, an extracted field, a routing decision — the presentation regression costs you nothing, the abstention improvement is a real gain, and the 62% task-cost reduction lands in full. If a person reads the deliverable — a report, a memo, a client-facing document — the newer model is measurably worse at exactly the thing you are paying for, and the saving is buying you a reviewer.

The abstention number deserves the same treatment. A model that goes from inventing on 93% of what it does not know to inventing on 77% is a materially different object for a guardrail, a compliance screen, or any pipeline where a confident wrong answer propagates. That improvement is real, it is independently measured, and it is the strongest argument for moving work onto the new model that does not depend on price at all.

What changes in your code, and what does not

Less than a price cut of this size implies, which is the good news. The context window is the same class, the maximum output is identical at 128,000 tokens, and the family's long-context repricing clause is unchanged: requests above 272,000 input tokens are billed at 2x input and cache rates and 1.5x output, for the entire request rather than the portion above the line. A request that was expensive at 300,000 tokens on GPT-5.6 Luna is expensive on GPT-6 Luna too — half the rate, same cliff, so a workload that should have been split in July should still be split now.

The effort ladder is where behaviour changes rather than cost. GPT-6 Luna exposes none, low, medium (the default), high, xhigh and max, and the default matters: a pipeline that was tuned against one model's default effort is not automatically tuned against another's. Teams that pinned a setting explicitly are unaffected. Teams that took the default are running a different configuration than they were in July, and the visible symptom will be token volume rather than quality.

One trap is worth restating because it does not disappear with the new model. On the Chat Completions endpoint, function calling works only when reasoning effort is set to none; every other effort level, and every built-in tool, requires the Responses API. A team porting a tool-calling loop by changing the model string will find its tools silently unavailable at the default medium effort, and the fix is a rewrite of the call surface rather than a parameter.

What you can route today, and what you cannot

This is the part of the migration that has nothing to do with benchmarks. GPT-5.6 Luna is on OrcaRouter at its list price — $0.20 per million input tokens, $1.20 per million output, a 1,000,000-token context window, a 128,000-token maximum output, and a p50 time to first token of 1.60 seconds measured across live traffic on our own model page. We pass provider list pricing through with no markup, so the day OpenAI repriced this family the change was live on our side rather than at the next billing cycle.

GPT-6 Luna is not routable through OrcaRouter as of September 23, 2026. Availability is OpenAI's own API and the cloud catalogues that carry this family, and we do not list a model we cannot serve. The practical consequence is that a team planning this migration can move the pricing today and cannot move the routing today — the two halves of the change land on different dates.

What that makes possible in the meantime is the arrangement most teams will end up wanting anyway. Keep GPT-5.6 Luna on the paths where the deliverable is the product, run GPT-6 Luna wherever the output is consumed by code, and treat the boundary as a tier rather than a rewrite. Because the models you can reach sit behind one endpoint with 200-plus models on the same key and automatic failover across providers, adding or removing a tier is a config entry rather than a second integration — and the escalation path stops depending on any single model being available.

The migration decision, stated plainly

Move the work where the output is machine-read and the success condition is verifiable. Classification, extraction, routing decisions, guardrail calls, bulk transform passes and single-step agent actions all qualify: the composite is unchanged, the abstention behaviour is materially better, and the task cost is down about 62%. With Batch at half of standard rates and cached input at a tenth of a halved base, a pipeline that was marginal in July can be straightforward now.

Do not move the work where a person signs off on the artefact, and do not move coding-agent paths blind. A 46-point AA-Briefcase fall and a 76-point GDPval fall are small numbers pointing the same direction as a two-point Coding Agent Index drop, and the failure mode Artificial Analysis describes — skipped rubric elements, weaker presentation — is invisible to an automated check. Keep the previous model where polish is the deliverable.

And treat the 0.07-point index tie as the finding it is. This is not a smarter model at a lower price. It is a differently-shaped model at a lower price, and the question a migration plan has to answer is which shape your workload consumes — not which score is higher, because on the independent record those two scores are the same number.