Solar Mini 4: cheap tokens, expensive tasks, weak agentic coding

Solar Mini 4 is cheap per token but not cheap per job, because it generates a lot of text to finish each task. Artificial Analysis, which benchmarks AI models independently, reports that the new reasoning model from Korean lab Upstage costs about five times as much per task as GPT-6 Luna (max), even though the two have similar per-token prices. Solar Mini 4 is priced at $0.10 per million input tokens and $0.40 per million output tokens, a third less than Upstage charged before. A token is a small chunk of text, and API bills are counted in them, so a low price per token still adds up when a model uses many tokens per answer.

The same arithmetic explains its speed. The model writes 208 tokens per second, faster than GPT-6 Luna (max) at 152. But it averaged 88,000 output tokens per task on the Artificial Analysis Intelligence Index, so a task took 7.1 minutes on average. Its overall index score of 24 is 16 points above Upstage's previous flagship, Solar Pro 3, and the evaluator calls it strong for its size. Upstage says only 3 billion of its 35 billion parameters are active for each token, and Artificial Analysis notes it cannot independently verify the size of a proprietary model. The results are uneven: 83% on a long-context reasoning test, but 1% on Terminal-Bench 4.0, which the evaluator counts as agentic coding.