Six models with a published Artificial Analysis Intelligence Index score reached OpenRouter on 21–22 September 2026. All six are now in the catalog, with prices, context and speed read from each model author's own endpoint.
The six at a glance
Throughput and first-token times are OpenRouter's median over a trailing thirty-minute window on each author's endpoint, read 23 September 2026.
We have separate write-ups for Claude Opus 5.5, Grok 4.7 and GPT-6 Sol and Luna. The other two are new providers for this catalog.
Xiaomi MiMo-V2.6-Pro: Grok 4.7's score at a fifth of the price
Xiaomi's MiMo-V2.6-Pro scores 46, the same as Grok 4.7. Its output costs $0.87 per million tokens, against Grok's $4.80, which makes it 5.5× cheaper. Cached input is $0.0036 per million, which is almost free.
The catch is speed. It streams at 24 tok/s, under half of Grok's 61, and takes 5.8 s to its first token against Grok's 1.3 s. Xiaomi serves it in FP8. The endpoint accepts text, images, video and audio, the widest input range of the six. It does not advertise a reasoning_effort parameter. For batch work where latency does not matter, it is the best index score per dollar in this group.
We have not yet reviewed whether MiMo-V2.6-Pro's weights are published for download. Its model page records that as an open question rather than guessing.
Cohere Command A+: built for first-token latency
Cohere's Command A+ scores 13 on the index. That puts it next to IBM's Granite 4.2 8B (12) rather than the frontier models in this group. Its standout number is latency: a 0.3 s median time to first token, by far the fastest of the six. At $0.30 input and $1.50 output per million, with a 192K context, it fits retrieval-augmented and enterprise assistants where response time matters more than reasoning depth.
What was left out
Fifteen other models reached OpenRouter between 5 and 22 September. They include GPT-6 Sol Pro and Luna Pro, MiMo-V2.6 Flash, GLM-5.3 FlashX, Sakana's Fugu Max and Fugu Ultra v2, and Mercury 2.5. None has an Intelligence Index score yet. Our rule is that a model's rating must be anchored to a published external ranking, so we add each one when a score appears rather than rating it on our own judgement alone.
Every rating for these six is an editorial estimate anchored to the Intelligence Index, labelled Estimated, with zero benchmark coverage. Any accepted benchmark result that clears the evidence thresholds replaces it.
