
Qwen4-Max vs GPT-6 Luna: 22× the Task Cost for 15 Index Points
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 187 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 413 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 87 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 119 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-079 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 89 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 135 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 78 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 116 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 114 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 66 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 91 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 41 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 212 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 111 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 168 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
Put a Qwen4-Max request and a GPT-6 Luna request through the same evaluation suite and the bill comes back at $3.27 against $0.15. Twenty-two times. Put the scores side by side and it is 43 against 28 — fifteen points, which is the difference between a model you would hand a multi-hour agent run and one you would hand a document. Both of those gaps are real, both are measured on the same index revision, and neither of them settles the question, because Qwen4-Max has been generally available since 16 November 2026 at $1.50 and $4.50 per million tokens while GPT-6 Luna has been sitting at a tenth of that since 22 September. The decision is not which model is better. It is which fifteen points you are actually going to use, and the answer changes twice on the way through this comparison.
The 22× and what it is made of
Start with the rate card, because the headline ratio is not the ratio you pay.
• Input — Qwen4-Max $1.50 per million in Singapore, $1.413 in the Global-scope regions. GPT-6 Luna $0.10. Fifteen to one.
• Output — Qwen4-Max $4.50, or $4.24 Global. GPT-6 Luna $0.50. Nine to one.
• Cached input — Qwen4-Max $0.075 Singapore, $0.071 Global, a 95% discount. GPT-6 Luna $0.01, a 90% discount.
• Blended — Qwen4-Max $0.80 per million on the tracker's 7:2:1 cache-hit/input/output weighting. GPT-6 Luna $0.077. Ten to one.
• Cost per Intelligence Index task — Qwen4-Max $3.27, GPT-6 Luna $0.15. Twenty-two to one.
Fifteen, nine, ten, twenty-two. Those four numbers do not agree, and the disagreement is the useful part: the per-task ratio is more than twice the blended-price ratio because Qwen4-Max has to emit more tokens to finish the suite, not because its rates are worse than the card suggests. The tracker's verbosity panel puts Qwen4-Max at 127 million output tokens against a 90 million median; Luna at 110 million against the same median. Both are on the verbose side. Qwen4-Max is 15% more so, and that 15% is doing real work in a 22× gap that a 10× rate card would have predicted.
All figures in this section are Artificial Analysis measurements on index v5.0.1 — the rebased index, so a Luna score from a July leaderboard is not comparable to the 28 quoted here — combined with each vendor's own published rate card. Neither vendor has published a matched head-to-head, and any comparison built from these numbers is arithmetic on two separate runs, not a benchmark.

The effort setting is the bigger lever
Here is what makes the 43-against-28 framing misleading, and it is not a small correction.
GPT-6 Luna ships with six reasoning settings — none, low, medium, high, xhigh and max — and it defaults to medium. Artificial Analysis scores every rung. The spread is not a rounding difference:
• Max — 28 on the index, $0.15 per task, 12.2 minutes, 103 tokens per second.
• Xhigh — 23, 7.4 minutes, 99 tokens per second.
• High — 21, 6.0 minutes, 91 tokens per second.
• Medium — 18, 3.5 minutes, 93 tokens per second. This is the default, and it is the setting almost every production call actually uses.
• Low — 10, 0.6 minutes, 87 tokens per second.
• Non-reasoning — 9, 2.0 minutes, 84 tokens per second.
So the honest comparison is not 43 against 28. It is 43 against 18 for a caller who sends no effort parameter, and 43 against 28 for one who turns the dial to the top and pays six times the latency for it. Twenty-five points at the default and fifteen at the ceiling, from the same model, with no price change on the card.
Qwen4-Max's ladder is three rungs — low, high, xhigh — and it defaults to high. There is no medium: Chat Completions and DashScope both map medium to high, so asking for the middle setting costs what the top one costs. That is a worse dial than Luna's, and it matters for the same reason — the rung you call decides the score, and Qwen4-Max gives you fewer ways to buy less of it.
What the tracker does not publish for Qwen4-Max is a score at each of its three rungs; the page carries 43 and stops. Luna's full six-rung ladder is public, which makes Luna the easier model to model and the harder one to quote. If you are building a cost forecast, a fully published ladder is worth more than a single better number, because you can see where the knee is.

Luna's 272,000-token cliff
Both models advertise a million-token context and only one of them charges the same rate across it, and this is the single most expensive thing a buyer can miss.
• Qwen4-Max — $1.50 and $4.50 flat across the full 1,000,000-token window, with no tiering at 32K or 128K. Documented input ceiling is 991,808 tokens non-thinking and 983,616 with thinking on; non-thinking output caps at 131,072 and thinking plus answer share 262,144.
• GPT-6 Luna — $0.10 and $0.50 up to 272,000 input tokens, then $0.20 input and $0.75 output above it. The repricing applies to the whole request, not the overflow: a 300,000-token prompt pays double on all 300,000, not on the last 28,000. Context window is 1,050,000 per OpenAI's documentation and 1M on the tracker; maximum input 922,000; output ceiling 128,000.
Run the long-context case and the 22× collapses toward 7.5×. A 400,000-token document with an 8,000-token answer costs $0.60 in and $0.09 out on Qwen4-Max — $0.69 at Singapore rates. The same request on Luna is $0.08 in and $0.006 out, $0.086. Seven and a half times, not twenty-two. Luna is still dramatically cheaper, and it is cheaper by a third of the margin the headline number implies.
That inversion has a second edge, and it cuts the other way. Below 272,000 tokens the cliff never applies, and most agent workloads — a codebase context, a support conversation, a tool-calling loop — live well under it. A team that reads "1M context" on both cards and budgets Luna at $0.10 for a 500,000-token job has under-budgeted by 2× on input and 1.5× on output, and the mistake only shows up in the invoice.
There is also a residency line worth knowing before the price matters: EU data residency on Luna is available only with Standard processing, which rules out the Batch and Fast discounts entirely. If a residency requirement is in scope, the effective Luna rate is the standard one, and the discount structure you were counting on is not available to you.
Video, and the row Luna cannot enter
The modality difference is smaller than the marketing suggests and larger than a spec-sheet glance implies.
• GPT-6 Luna — text and image in, text out. Not a video or audio model. Image generation, speech generation, transcription and translation are all documented as unsupported.
• Qwen4-Max — text, image and video in, text out, with vision handled natively rather than through an extractor. OpenAI Chat Completions, OpenAI Responses, Anthropic Messages and DashScope on the protocol side, against Luna's Responses, Chat Completions and Batch.
So Qwen4-Max is the only one of the two you can point at a screen recording, a lecture capture or a video review and ask a question about. That is a genuine capability difference and it is not a benchmark difference — no index on either card measures it, and no amount of effort-level tuning on Luna produces it. If video is in the workload, the comparison ends here and the 22× is the price of admission.
If video is not, the row is decoration, and the protocol breadth is the more interesting half. Qwen4-Max speaks four wire formats; Luna speaks three and gates its richest tool surface behind Responses — built-in tools such as web search, file search, code interpreter and hosted shell run there, and function calling on Chat Completions works only when reasoning effort is set to none. A pipeline that moves to Chat Completions to save money and still expects tool calls fails on the first request. That is a migration trap, not a capability gap, and it is the kind of thing that costs a week rather than a percentage.
This is the seam where a routing layer earns its keep rather than a model choice. OrcaRouter already carries openai/gpt-6-luna at OpenAI's list prices with 0% markup passed through — so a Luna price change is live here the same day it lands on OpenAI's card — and Qwen4-Max is not hosted here; it is reachable through Alibaba's own Model Studio API and several third-party platforms. One API for 200-plus models, one key, and a routing DSL that classifies a request and sends it to whichever of the two the row says it belongs on: Luna for the high-volume extraction tier, Qwen-class for the video and long-context work Luna structurally cannot take. Automatic failover covers the day the cheap tier has a bad hour, which is the day a single-provider pipeline discovers it has one.

Where each model is the right call
Three workloads, three different answers, and the third one is where most teams actually are.
High-volume, well-specified, checked. Classification over a corpus, extraction from document sets, summarisation, the support-triage rung of a pipeline. Luna at $0.15 a task against Qwen4-Max at $3.27 is a 95% cost reduction for work that never needed a 43. The 18 you get at Luna's default is above the median of its price band and the output is checkable, which is the whole argument. This is not a close call.
Long-horizon, unattended, expensive to get wrong. Multi-hour agent runs, repository-scale refactors, research that spans dozens of tool calls. Fifteen index points is the difference between a model that can hold a plan and one that loses it, and at $3.27 a task against a run that costs a engineer-hour to redo, the price is not the constraint. Qwen4-Max is also the only one of the two that will take a video input, which for a growing share of agentic work — screenshots, screen recordings, UI traces — is the difference between a workload that exists and one that does not.
Latency-sensitive, somewhere in the middle. This is the case that neither card answers. Luna decodes at 103 tokens per second at max effort and 93 at its default; Qwen4-Max at 59. Luna finishes an index task in 12.2 minutes at max and 3.5 at default; Qwen4-Max in 29.8. If a human is waiting, 59 tokens per second is the number that decides whether the product feels alive, and no per-task saving compensates for an interface that stalls. If nobody is waiting, 29.8 minutes is wall-clock you can schedule around.
The correct shape for most teams is not one of these two, it is both — and the reason is that the decision boundary is a request-level property, not a contract-level one. A pipeline that classifies on Luna and escalates to a frontier model when the classifier is unsure gets the 95% saving on the easy 90% and the fifteen points on the rest. That is what the routing DSL is for, and it is the only configuration in this comparison that does not require picking a loser.
The bill you can actually predict
Twenty-two times the task cost for fifteen index points. Divide one by the other and Qwen4-Max costs $0.076 per index point against Luna's $0.0054 — fourteen times more per point, which is the number that makes Luna look like the obvious buy and the number that is least useful, because nobody consumes index points.
What people consume is a workload, and the workload has three properties that decide this: how many tokens the context runs to, whether a human is waiting, and whether a wrong answer is cheap to catch. Luna wins the first two outright and loses the third so badly that the comparison inverts. Qwen4-Max wins the third and charges for it at a rate that only makes sense if the alternative is a re-run.
The prediction that survives all of it: if you are choosing between these two on price, you have already chosen wrong, because the 22× is not a premium you should be paying on every request in a pipeline — it is a premium you pay on the small fraction of requests that need it. Route the fraction. The rest is a rounding error you can spend on something else.
