DeepSeek’s newest open-weight model posts a score on the Artificial Analysis Intelligence Index that’s more than double the median for comparable models. According to Hacker News, DeepSeek V4.1 Flash 1017 (Max) scored 45, while the median for models in its class is 19. It also costs less than most of its peers and runs quickly. The catch is that it talks a lot.
The listing climbed to 189 points on Hacker News. Developers want open models that can compete with closed models without costing as much, and this one looks like a serious contender.
📊 The Numbers at a Glance
Here’s how V4.1 Flash compares with the median for similar open-weight models, as reported on Hacker News:
- Intelligence Index score: 45 vs. a median of 19
- Input price: $0.30 per 1M tokens vs. a median of $0.43 (about 30% cheaper)
- Output price: $1.20 per 1M tokens vs. a median of $1.68 (about 29% cheaper)
- Tokens used to run the full index: 210M vs. a median of 140M (50% more)
- Speed: 230 tokens per second, which the listing calls “notably fast”
- Average cost per index task: $0.26
The model accepts text and images, outputs text, and has a 1M-token context window. That’s enough to load an entire codebase or a large set of documents into a single prompt.
🔍 How the Test Works
The Artificial Analysis Intelligence Index is a composite benchmark. It runs a model through a set of evaluations covering reasoning, knowledge, math and coding, then combines the results into one score. It also records how many tokens the model produces along the way. That’s why verbosity shows up as its own metric.
The “(Max)” in the name most likely refers to a maximum reasoning-effort setting. In that mode, the model spends more tokens thinking before it answers. That helps explain both the high score and the high token count.
💬 Why Verbosity Matters
This is the part most people will skip, and they shouldn’t. The per-token prices look good. But a model that writes 50% more tokens to reach an answer gives back part of that discount. Output tokens are the expensive ones, so every extra paragraph of reasoning shows up on your bill.
The $0.26 per task figure already includes those extra tokens, so it’s a more honest number than the headline prices. Even so, verbosity has costs beyond money:
- Latency: At 230 tokens per second, a long answer still takes a while to finish. A higher generation speed doesn’t help much if the model generates twice as much text.
- Context usage: In agent loops, long outputs fill the context window quickly, even one as large as 1M tokens.
- Post-processing: If your app parses the model’s responses, more text means more to filter out.
🛠️ What You Can Do With This
If you’re choosing a model for production work, V4.1 Flash is worth testing. A few practical steps:
- Benchmark on your own tasks. Index scores are averages. Measure accuracy and tokens per completed task on your actual workload.
- Compare cost per task, not cost per token. A cheaper model that writes twice as much can end up costing more.
- Try lower effort settings. If a non-Max mode exists, check whether it holds up on simpler queries at a fraction of the tokens.
- Cap output length. Use max-token limits and prompt instructions to keep responses short where you can.
⚠️ Caveats
The source data comes from a single benchmark index. It doesn’t break down performance by task type, and it doesn’t give the model’s parameter count. “Comparable models” is a moving category, so the medians will change as new releases arrive. Treat these numbers as a reason to test the model yourself, not as a final verdict.
What stands out is the direction things are moving. Open-weight models are now beating their class medians by wide margins at below-median prices. That puts more pressure on closed-model pricing with every release. The next thing to watch is whether DeepSeek can cut the verbosity without losing the score.
Full benchmark details are available in the original listing on Hacker News.