GPT-6 Luna vs Claude Opus 5: Fifty Times the Price and the Same Afternoon

GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output. Claude Opus 5 costs $5.00 and $25.00. That is a fifty-fold gap on both sides of the rate card, and it is the first thing anyone notices about this pairing — a model that the vendor shipped on September 22, 2026, priced against a model the company released on July 24, 2026, that has been on the market for two months. The fifty-times number is real. It is also close to the least useful fact in this comparison, because these two are not competing for the same call, and because of something that happened on the second model's side of the fence the same afternoon Luna launched.

GPT-6 Luna is the small tier of OpenAI's GPT-6 generation, sitting below GPT-6 Sol at $2.00/$10.00 and well below the flagship GPT-6 Astra at $10.00/$50.00. Claude Opus 5 was Anthropic's flagship until September 22, when Anthropic shipped Claude Opus 5.5 at $4.00/$20.00 and moved Opus 5 down a rung. Both of the models named in this article's title are, as of today, one step off the top of their own vendor's ladder — for reasons that have nothing to do with each other. That matters more than the price ratio, and it is where this piece starts.

The afternoon both vendors shipped

OpenAI announced GPT-6 Sol and GPT-6 Luna on September 22. Hours earlier, Anthropic announced Claude Opus 5.5. Two of the three models in that sentence cut prices by roughly half against their predecessors, which is not a coincidence so much as a market: the previous eighteen months of open-weight releases from Chinese labs had already pushed the cost of a competent token down hard, and both vendors were responding to the same pressure.

The practical consequence for anyone holding a Claude Opus 5 deployment is that the model they built against is no longer the model Anthropic is selling hardest. Claude Opus 5.5 is $4.00/$20.00 — a 20% list cut — with cached input at $0.20 per million, down from Opus 5's $0.50, and Artificial Analysis measures it at 58 on its Intelligence Index at maximum effort, the highest figure it has recorded. Claude Opus 5 measures 51 on the same index. If you are running Opus 5 today, the fifty-times comparison in this article's title is not the one in front of you. The one in front of you is whether you move to Opus 5.5, and the answer to that is a separate piece of arithmetic from anything GPT-6 Luna is doing.

That said, the fifty-times number still governs a real decision, because the workloads that GPT-6 Luna is built for are exactly the workloads where a Claude Opus 5 bill gets uncomfortable. Both facts can be true at once.

What each of these two actually is

The specs, one line per dimension, both sides on each:

• Price per 1M tokens — GPT-6 Luna $0.10 in / $0.50 out vs Claude Opus 5 $5.00 in / $25.00 out

• Cached input — GPT-6 Luna $0.01 per 1M, a 90% discount vs Claude Opus 5 $0.50 per 1M on the Opus 5 card, cut to $0.20 on Opus 5.5

• Context window — GPT-6 Luna 1,050,000 total with 922,000 maximum input vs Claude Opus 5 1M, both capped at 128,000 output tokens

• Long-context repricing — GPT-6 Luna doubles input and cache and lifts output 1.5x above 272K input tokens, applied to the whole request vs Claude Opus 5 flat across its window

• AA Intelligence Index — GPT-6 Luna 37 at max effort vs Claude Opus 5 51

• AA Intelligence Index, effort-matched — GPT-6 Luna 29 at its default medium effort vs Claude Opus 5 51 at max, the only setting Anthropic measures it at

• Output speed — GPT-6 Luna 153.9 tokens/sec vs Claude Opus 5 53.6 tokens/sec

• Cost to run the index — GPT-6 Luna $122.39 vs Claude Opus 5 $7,274.74

• Reasoning off switch — GPT-6 Luna offers none, low, medium, high, xhigh and max vs Claude Opus 5 has thinking that is adaptive and always on

• Released — GPT-6 Luna 2026-09-22 vs Claude Opus 5 2026-07-24

Three of those lines are doing more work than the others, and the first is the effort row. GPT-6 Luna's 37 is its score at maximum effort. Its default is medium, and at medium it scores 29. Claude Opus 5's 51 is a max-effort number, because Anthropic does not expose a lower setting that Artificial Analysis measures. So the headline gap of fourteen points is a comparison between one model's ceiling and another model's ceiling — and a team that installs GPT-6 Luna and changes nothing is running the 29, not the 37. That is a twenty-two-point gap, not a fourteen-point one, and it is invisible unless you go looking for the effort default.

The second is the index cost line. It is not a quote — it is what Artificial Analysis actually spent running the full evaluation suite on each model with its own money. $122.39 against $7,274.74 is a fifty-nine-fold difference in the cost of finding out how good the model is, and it is the cleanest available proxy for how the two behave at volume. Claude Opus 5 generated 140 million tokens during that run; GPT-6 Luna generated 150 million. The cheap model was more verbose and still cost fifty-nine times less.

The third is the off switch. GPT-6 Luna can be told not to reason at all. Claude Opus 5 cannot. That single spec line is why one of these models can sit behind a classifier, a router, or a moderation screen and the other cannot, and it explains a large part of the price difference without appealing to capability at all.

Where the cheap tier stops being cheap

There is a cliff on the GPT-6 Luna rate card, and it is the single most expensive thing to miss. Requests that exceed 272,000 input tokens are billed at twice the input and cache rates and 1.5x the output rate — for the entire request, not the portion that crossed the line. A 300,000-token call on GPT-6 Luna bills at $0.20 per million input for all 300,000 tokens, because it went over by 28,000. Split the same document into two calls and the cost halves with no change to the prompt.

Claude Opus 5 has no equivalent clause. Its $5.00/$25.00 is flat across the full 1M window, which means the fifty-times ratio is a fifty-times ratio everywhere — except above 272K input tokens, where GPT-6 Luna's effective input rate doubles to $0.20 and its output rate rises to $0.75, and the gap narrows to twenty-five times. Still enormous. Still worth knowing, because the workloads most likely to have a 300,000-token working context are exactly the agentic ones both vendors are selling into.

The second thing to price is effort. Running GPT-6 Luna at max effort for a task that would have been fine at medium is the most common way to spend the savings you just made. The 37-versus-29 spread is the size of the penalty for the wrong setting, and unlike a price clause it does not appear anywhere on a bill — it shows up as token volume, which looks like success.

The split that actually decides it

The honest version of this comparison is that the two models are not substitutes, and the price ratio is a distraction from a workload question that has a clear answer either way.

GPT-6 Luna belongs under anything where the output is consumed by a program rather than a person and the volume is the point. Classification, extraction, routing decisions, bulk summarisation, the inner loop of an agent that makes a thousand small calls per task. At $0.10/$0.50 with a one-cent cached input rate, and with Batch and Flex at half the standard rate, the economics are not a close call against anything in the Opus line. Artificial Analysis flags one genuine regression worth knowing before you migrate: GPT-6 Luna's Coding Agent Index fell two points to 41 against GPT-5.6 Luna's 43, with the declines concentrated in SWE-Atlas-QnA and DeepSWE v1.1. It also flags one genuine gain of the same magnitude — the model's hallucination rate on AA-Omniscience fell from roughly 93% to 77%, while its accuracy barely moved. It declines to answer more often rather than knowing more, and for anything with a downstream consumer that is the more valuable half of the trade.

Claude Opus 5 belongs where the output is the product. Anthropic's own published results — vendor-reported, not independently reproduced — put it at 96.0% on SWE-bench Verified and 79.2% on SWE-bench Pro, and the fourteen-point index gap is the kind of margin that decides whether a long-horizon agent finishes a task or quietly stops halfway. If you are at the edge of what the model can do, a model that is fifty times cheaper and twenty-two points behind at default effort is not a cheaper version of the same thing.

Running both without choosing

The reason this comparison does not resolve into a single recommendation is that the boundary between the two workloads is not stable. It moves every time either vendor ships, and it moved twice in the three weeks before this article — once when Anthropic cut the Opus rate to $4.00/$20.00, and once when OpenAI halved the price of its entire small tier. A team that hard-coded a winner in July is running the wrong configuration today.

Claude Opus 5 is on OrcaRouter at Anthropic's list price, under the pass-through pricing that means a vendor rate change is live on our side the same day rather than waiting on a reseller to renegotiate — which is the concrete reason the Opus 5.5 cut landed in our catalogue without an integration project. GPT-6 Luna is not on OrcaRouter as of this writing; you reach it through OpenAI's own API. So the honest setup for a team that wants both is one key for Claude Opus 5 alongside whatever you already use for OpenAI, with automatic failover covering the gap when a provider degrades. Moving a route between a one-cent classifier and a twenty-five-dollar reasoner by changing a config rather than a code path is worth more than picking correctly in September.

What to do this week

If you are running Claude Opus 5, the fifty-times gap in the title is a real number pointing at the wrong decision. Your decision is Opus 5.5 — a 20% list cut, a higher index score, and a cached-input rate that fell 60% — and GPT-6 Luna is a different question that happens to be on the same page.

If you are running a high-volume classification or extraction pipeline on an Opus-class model, GPT-6 Luna is the largest single cost move available to you this month, and the migration is smaller than the price cut implies: the same 1M context class, the same 128,000 output cap, and a per-token rate that is fifty times lower on both sides. Test it at medium effort before you assume the default, put your long-context calls on the right side of 272,000 tokens, and check the coding regression against your own eval rather than against a vendor chart.

And if you are running both, the thing to fix is not which one you chose. It is that changing your mind currently costs a code deployment.