Open vs Closed AI Models: What the Same Quality Costs in 2026
We matched 154 models' LMArena scores against the cheapest API price we track for each. Just below the frontier, open-weight models cost about a quarter as much as closed ones of the same quality. At the very top, there is still no open option.
Is an open-weight model actually cheaper than a closed one of the same quality, or only cheaper because it’s worse? We put every model that has both an LMArena Text Arena score and an API price on Inference Hub side by side: 154 models, 73 open-weight and 81 closed.
The short answer:
- At the frontier, there is no open choice. None of the 14 models scoring 1,490 or more is open-weight. The best open model, Kimi K3, scores 1,488, 39 points behind Gemini 4 Carbon.
- One step down, open models cost about a quarter as much. Among models scoring 1,470 to 1,489, the median closed model costs 4.0× the median open one. The cheapest open model in that band, GLM 5.3 Flash, costs about a seventh of the cheapest closed one, Gemini 3 Flash.
- Closed models win on price in places. In the 1,450–1,469 band, OpenAI’s GPT-5.6 Luna is cheaper than every open model of the same score, and GPT-6 Luna is the second-best value of all 154.
- Open models are sold by more providers. Above 1,440, an open model is available from 14 providers on average against 9 for a closed one, and that competition is a big part of why they’re cheap.
How we compared
- Quality: LMArena’s Text Arena score, from blind head-to-head votes by real users, as of 19 November 2026. It covers more models than any other benchmark we track. It measures chat quality, so it says less about coding or long agent tasks.
- Price: the cheapest offer we track for each model across every API provider on Inference Hub, as a blended price per million tokens: (3 × input + output) ÷ 4. That weighting assumes a typical chat or RAG workload where prompts are longer than answers.
- Open-weight: the model’s weights are published for download. API-only models from the same labs (for example the larger Qwen Max models) count as closed.
- Cheapest offer, not official price: for closed models, the cheapest offer is often a reseller such as Kie AI selling below the official rate. At official prices, the gap below would be wider.
Price by quality band
| Text Arena score | Open models | Closed models | Cheapest open | Cheapest closed | Median open | Median closed | Closed ÷ open (median) |
|---|---|---|---|---|---|---|---|
| 1,490 and up (frontier) | 0 | 14 | none | Gemini 3.8 Flash $0.45 | – | $2.86 | – |
| 1,470–1,489 | 7 | 19 | GLM 5.3 Flash $0.050 | Gemini 3 Flash $0.34 | $0.40 | $1.60 | 4.0× |
| 1,450–1,469 | 7 | 14 | Gemma 4 31B $0.15 | GPT-5.6 Luna $0.13 | $0.54 | $1.23 | 2.3× |
| 1,420–1,449 | 17 | 12 | DeepSeek V4 Flash $0.018 | GPT-6 Luna $0.06 | $0.37 | $0.89 | 2.4× |
| 1,380–1,419 | 20 | 11 | Qwen3.5 35B A3B $0.095 | Solar Pro 4 $0.053 | $0.44 | $0.55 | 1.2× |
| Under 1,380 | 22 | 11 | Llama 3.2 1B $0.006 | Nova Lite $0.11 | $0.13 | $0.36 | 2.7× |
Blended price per million tokens, cheapest offer we track, 22 November 2026.
Two things stand out. First, the frontier band is closed-only, and it isn’t only expensive models up there: Gemini 3.8 Flash scores 1,497 at a cheapest price of $0.45, cheaper than most open models in the band below it. If you need top-tier quality at a low price, a cheap closed “Flash” model can be the right answer.
Second, the closer you get to the top without reaching it, the bigger the open-model discount. In the 1,470–1,489 band, DeepSeek V4.1 Flash and GLM 5.3 Flash score the same as GPT-5.2 and GPT-5.4 (1,475) for a small fraction of the price.
The best value at each level
The models with a Text Arena score of 1,440 or more, ranked by cheapest blended price:
| Model | Open? | Text Arena | Cheapest blended price | Providers |
|---|---|---|---|---|
| GLM 5.3 Flash (Z.ai) | Open | 1,475 | $0.050 | 24 |
| GPT-6 Luna (OpenAI) | Closed | 1,443 | $0.060 | 10 |
| DeepSeek V4.1 Flash (DeepSeek) | Open | 1,475 | $0.105 | 30 |
| GPT-5.6 Luna (OpenAI) | Closed | 1,453 | $0.126 | 9 |
| MiMo V2.6 Flash (Xiaomi) | Open | 1,448 | $0.145 | 7 |
| Gemma 4 31B (Google) | Open | 1,452 | $0.153 | 10 |
| DeepSeek V4 Pro (DeepSeek) | Open | 1,464 | $0.198 | 31 |
| GLM 5.2 (Z.ai) | Open | 1,475 | $0.290 | 23 |
| Gemini 3 Flash (Google) | Closed | 1,473 | $0.338 | 10 |
| GLM 5.3 (Z.ai) | Open | 1,482 | $0.402 | 27 |
DeepSeek V4 Flash is cheaper still ($0.018 blended on its cheapest endpoint) but scores just under the cut-off at 1,439.
Why open models are cheaper
- Competition. Anyone can host an open model, so dozens of providers sell the same weights and undercut each other. DeepSeek V4 Pro is on 31 providers, GLM 5.3 on 27. A closed model is sold by its maker and a handful of resellers and clouds.
- Efficient architectures. Most of the cheap open models are mixture-of-experts designs that only run a small part of the network per token, which keeps serving costs low.
- Quantization. Some of the cheapest endpoints serve open models at reduced precision (fp8 or fp4). That’s usually fine, but check the provider’s notes if output quality matters; our model pages show each provider’s price and notes.
What this means if you’re choosing a model
- You need the very best chat quality: pay for a closed frontier model. Gemini 4 Carbon and Claude Opus 5.5 lead, and the cheapest frontier option is Gemini 3.8 Flash.
- You need “very good” at scale: an open model in the 1,470–1,489 band, such as GLM 5.3 Flash, DeepSeek V4.1 Flash or GLM 5.3, gives you near-frontier chat quality for a quarter of the price or less, from many providers.
- You want a closed model on a budget: OpenAI’s Luna models are the cheapest closed options with strong scores.
- You care about data control: open models can also be self-hosted. See running open models locally and our ranking of open-source models.
Chat scores aren’t everything. For coding, the gap between open and closed models is wider (see the WebDev Arena column in our open-source ranking), so test on your own prompts before switching.
Data: LMArena Text Arena scores as of 19 November 2026; prices from the API providers we track, checked weekly. Each model page lists every provider’s price. You’re welcome to quote these numbers with a link to this page.