Published · 4d agoProduct2 min read
The Number That Ties Is 52, Not 1M: Qwen3.8 27B Matches Luna And Bills More Per Token
Artificial Analysis puts an open 27B model and OpenAI's GPT-5.6 Luna (max) at the same Intelligence Index score. Serving the free one costs about 2.7x more per output token.
Not a builder's beat, but builders have a standing stake in it.See today for builders
What happened
- Qwen3.8 27B scores 52 on the Artificial Analysis Intelligence Index, a composite benchmark covering reasoning, knowledge, mathematics and coding.
- GPT-5.6 Luna (max), created by OpenAI, scores 52 on the Artificial Analysis Intelligence Index.
- Qwen3.8 27B has 27 billion parameters and was created by Alibaba.
- Qwen3.8 27B is open weights: the model weights are publicly available and can be downloaded for self-hosting.
- GPT-5.6 Luna (max) is proprietary; the model weights are not publicly available.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
Artificial Analysis scores Qwen3.8 27B at 52 on its Intelligence Index [1], the same score it publishes for GPT-5.6 Luna (max) from OpenAI [2]. Qwen's model has 27 billion parameters and open weights that anyone can download and self-host [3][4]; Luna is proprietary, with weights that are not publicly available [5].
One correction before the rest. The million-scale figures in the material behind this brief are context windows, not downloads: Luna has a 1.0M token context [6], Qwen3.8-27B is native to 262,144 tokens and extensible to 1,000,000 [7], and Alibaba says a hosted Qwen Cloud version will default to 1M context once the service ships [8]. No download count appears in our source material, so this desk is not going to assert one.
What the tie turns on is the bill. Across providers serving Qwen3.8 27B, the median price is $0.45 per 1M input tokens and $3.20 per 1M output tokens [9]. Luna, on OpenAI's own API, is $0.20 and $1.20 [10]. That is roughly 2.7x more per output token for the open model [11], and on the blended 7:2:1 rate Artificial Analysis publishes, $0.45 against $0.17, about 2.6x [12][13][14]. Free to download is not the same as cheap to run, and nothing in the material says what hardware the weights actually need.
Volume compounds it. Qwen generated 160M output tokens working through the index, against a 43M median for comparable open-weight models [15]; Luna generated 130M against a 60M median [16], at a total evaluation cost of $172.17 [17]. More expensive tokens, more of them.
Read the peer groups too. Qwen's 52 sits above a median of 9 for open-weight models its size [18]; Luna's 52 sits above a median of 17 for reasoning models in its price tier [19]. Same score, different crowds.
Watch the hosted Qwen service and whether provider prices for the open weights drift below Luna's, since $0.45/$3.20 is a median, not a floor [8][9].
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Qwen3.8 27B scores 52 on the Artificial Analysis Intelligence Index, a composite benchmark covering reasoning, knowledge, mathematics and coding.
- [2]
GPT-5.6 Luna (max), created by OpenAI, scores 52 on the Artificial Analysis Intelligence Index.
- [4]
Qwen3.8 27B is open weights: the model weights are publicly available and can be downloaded for self-hosting.
ReportedView cited source - [5]
GPT-5.6 Luna (max) is proprietary; the model weights are not publicly available.
ReportedView cited source
Sources & coverage · 2 publishers
The reporting this story was synthesized from, earliest first. Every link goes to the original.
- huggingface.co4d agoQwen/Qwen3.8-27B · Hugging Face
- artificialanalysis.ai4d agoQwen3.8 27B - Intelligence, Performance & Price Analysis
- artificialanalysis.ai4d agoIntelligence, Performance & Price Analysis
