Xiaomi MiMo-V2.6-Pro and Grok 4.7 Scored the Same 46 on the Same Index — One Cost 24× More to Run

Two models landed within hours of each other on 21 September 2026: Xiaomi MiMo-V2.6-Pro, the flagship checkpoint of Xiaomi's MiMo-V2.6 series, and Grok 4.7, the latest coding-and-knowledge model from SpaceXAI. Both now sit at 46 on the Artificial Analysis Intelligence Index, version 4.3.2 — the readings behind those rounded 46s are 46.32 and 46.45, a gap of 0.13 points. Putting each through that same index cost Artificial Analysis $206.66 for MiMo-V2.6-Pro and $4,967.35 for Grok 4.7. Roughly twenty-four to one, for a score you could not tell apart on the leaderboard.

That is the whole story of the night, and it is worth being precise about what it does and does not say. It does not say the two models are the same model. It says that on the one measurement both vendors are happy to be judged by, they are indistinguishable — and that the cost of getting that measurement out of them is not.

What actually shipped on both sides

Xiaomi published and open-sourced the MiMo-V2.6 series: MiMo-V2.6-Pro and MiMo-V2.6-Flash as the two native omni-modal checkpoints, plus a MiMo-V2.6-Pro-UltraSpeed serving tier of the same Pro weights. The weights went up on Hugging Face on 21 September, eighteen seconds apart, under the MIT licence, with a technical report attached.

SpaceXAI shipped Grok 4.7 the same day as a proprietary model — no weights, 500k-token context, text and image input, text out, with reasoning-effort levels from low through xhigh. The xhigh setting is the configuration Artificial Analysis scored.

So the two arrive with opposite business models and land on the same number. That is not a coincidence anyone engineered; it is what the frontier of September 2026 looks like from two directions at once.

The rate cards, side by side

Xiaomi's own price list, per million tokens in RMB, carries the V2.5 series rates unchanged — the vendor's stated position is better intelligence at the same price:

• MiMo-V2.6-Pro — ¥0.025 on a cache hit, ¥3.00 input, ¥6.00 output. Artificial Analysis lists the same model at $0.43 input and $0.87 output per million.

• MiMo-V2.6-Flash — ¥0.02 on a cache hit, ¥1.00 input, ¥2.00 output. Roughly a third of Pro on both ends.

• MiMo-V2.6-Pro-UltraSpeed — ¥0.25 on a cache hit, ¥30.00 input, ¥60.00 output. Exactly ten times standard Pro, for the same checkpoint served faster.

• Grok 4.7 — $2.00 input and $6.00 output per million tokens at the standard tier, unchanged from Grok 4.6. Above a 200,000-token prompt those rates double to $4.00 and $12.00.

On headline price alone MiMo-V2.6-Pro is about 4.6× cheaper on input and 6.9× cheaper on output. On the cache-hit line the vendor's own figure is ¥0.025 — the number Xiaomi is putting in front of agent builders, because a long agent loop re-reads the same context hundreds of times and that line is the one that decides the invoice.

Why the real gap is 24× and not 5×

The price ratio understates it, and the reason is worth internalising if you build cost models from rate cards. Grok 4.7 is a talkative model. Artificial Analysis reports that evaluating it on the Intelligence Index generated 240 million output tokens against a median of 94 million across the models on that board, and describes it as "notably slow and very verbose". MiMo-V2.6-Pro generated 140 million on the same evaluation, which Artificial Analysis calls only "somewhat verbose".

Multiply the token count by the rate and the sticker gap widens into the evaluation-cost gap: $4,967.35 against $206.66. A team budgeting from the rate card would predict a 5× difference and get something closer to 24× on a workload that streams long outputs.

Two more readings from the same page, both measured rather than claimed: MiMo-V2.6-Pro ran at 125 output tokens per second with a median of about 2.19 seconds to first chunk. Grok 4.7 sits at the slow end of the board. Speed and verbosity are the two variables a per-token price cannot show you, and they are the two that decide what a fixed monthly budget buys.

The "two extremes" thesis, tested

There is a reading of this week's releases that circulated quickly: that a model now has to pick one of two poles to be worth anyone's attention — be the best thing at the cheap end, or be genuinely frontier — and that anything in the middle goes unremarked. The MiMo-V2.6 launch and Grok 4.7 are a fair test of it, and the numbers mostly hold up.

Xiaomi chose the cheap pole deliberately and said so: its own announcement claims MiMo-V2.6-Pro beat Kimi K3 and Qwen3.8 Max on the Artificial Analysis index to become the strongest open-weights model on it, while conceding it still trails Claude Fable 5.1 and GPT-6 Astra. The open-weights board as captured on 22 September bears the first half of that out — MiMo-V2.6-Pro at 46.32 is the highest-scoring entry whose weights anyone can download, ahead of GLM-5.3 at 44.78 and Kimi K3 at 43.59. The closed models above it are Fable 5.1 at 53.35, GPT-6 Astra at 52.67 and Claude Opus 5 at 50.78.

SpaceXAI chose the other pole. Grok 4.7 is not open, is not cheap, and is not claiming to be — it is claiming the fourth-place lab position on aggregate and a coding-agent score that moved further than its general score did.

What the thesis misses is that the middle is not empty; it is just where the pricing arguments happen. MiMo-V2.6-Flash is the middle: a 309B-total, 15B-active checkpoint costing a third of Pro and, on Xiaomi's own benchmark tables, trailing it by a couple of points on most agent tasks. Flash beats Pro outright on one cybersecurity column. A model that is nearly as good for a third of the price is not undiscussed because it is uninteresting — it is undiscussed because "nearly as good, much cheaper" is a boring headline and a very good invoice.

What each one is, in config terms

• Parameters — MiMo-V2.6-Pro is a sparse mixture-of-experts at 1.02T total with 42B activated per token, on 70 layers, hidden size 6144, 384 routed experts with 8 active per token. Grok 4.7's parameter count has never been published.

• Weights — MiMo-V2.6-Pro is MIT-licensed open weights, roughly 573 GB of repository storage across the full checkpoint including its audio tokenizer and speculative-drafting stack. Grok 4.7 has no downloadable weights at all.

• Context — MiMo-V2.6-Pro declares a 1,048,576-token maximum position embedding. Grok 4.7 carries a 500k-token context window.

• Modality — MiMo-V2.6-Pro takes text, image, speech and video in and returns text. Grok 4.7 takes text and image in and returns text.

• Independent score — MiMo-V2.6-Pro 46.32 on Artificial Analysis Intelligence Index v4.3.2; Grok 4.7 (xhigh) 46.45 on the same version. Both are Artificial Analysis measurements, not vendor tables.

• Cost to evaluate on that index — $206.66 against $4,967.35, both figures published by Artificial Analysis on its own model pages.

What to do with this by Monday

The practical consequence is not "switch to the open model". It is that a single number on a leaderboard no longer tells you which of two tied models to run, and that the deciding variables — tokens per task, cache-hit rate, output speed — are published by the evaluator rather than the vendor. Pull them before you pull a rate card.

Neither MiMo-V2.6-Pro nor Grok 4.7 is routable through OrcaRouter as this is written, and it would be dishonest to imply otherwise: our catalogue carries 200+ models and currently includes neither of these two. What the catalogue does offer is the shape of the answer — one endpoint, provider list price passed through with 0% markup, so a vendor price cut lands on our side the same day, and the routing DSL for splitting a workload across models by task rather than by brand. When a tier like MiMo-V2.6-Pro-UltraSpeed prices the same weights at ten times the rate, the interesting engineering question stops being which model wins and becomes which requests deserve the expensive one — and that is a routing decision, not a purchasing decision.

The open question is whether the tie holds. Xiaomi has promised the V2.6 series at V2.5 prices and published a training ledger to go with it; SpaceXAI has kept Grok 4.7 at Grok 4.6's rates. If the next checkpoint on either side moves the index by five points, the price gap stops being the story and becomes the reason to re-run the comparison. Until then, the honest summary of 21 September 2026 is a night when two companies shipped models of equal measured intelligence and only one of them billed like it.