Grok 4.7 and MiMo-V2.6-Pro tie at 46 on Artificial Analysis. Grok's output price is 6.9x Xiaomi's, its bill for the same test was 19x, and most of that gap is on the input side.

• Same score, 46.33 against 46.32. Running the whole index cost $3,881 on Grok 4.7 at xAI's default effort and $207 on MiMo-V2.6-Pro. Output price ($6.00 against $0.87) explains less than half of that. Grok writes more tokens, and its cache hits cost $0.50 to Xiaomi's $0.0036.
• Grok answers faster than Xiaomi's own MiMo endpoint, but a third-party MiMo host beats it on speed at MiMo's price, and Claude Sonnet 5.5 outscores both for about a third of Grok's bill.

Two rate cards that didn't move

Neither company changed its price for the new model. Grok 4.7 launched on September 21 on Grok 4.6's card, and xAI's launch post says so in as many words. MiMo-V2.6-Pro arrived the next day, Beijing time, on the same $0.435 and $0.87 that MiMo-V2.5-Pro charges.

xAI pricing docs and Xiaomi's MiMo-V2.6 launch page and API catalogue, read October 4, 2026. xAI's US regional endpoint adds 10%. Grok 4.7 Fast ($4 / $1 / $12) exists only inside Cursor and Grok Build, not on the API. Xiaomi says cache writes are free for a limited time.

Line by line Grok costs 4.6x as much on input, 6.9x on output and 139x on a cache hit. That last number is the one people skip, and it's the one an agent with a long system prompt pays most often. On xAI's own API there's no Batch discount either: xAI gives 20% off batch to Grok 4.3 and the 4.20 models, and the Grok 4.7 model page lists Batch as not supported. Amazon Bedrock, which added Grok 4.7 on September 28, is the exception. It sells the same $2 / $6 card and a Flex tier at half that, $1 and $3.

One test, two bills

Artificial Analysis runs every model through the same set of tasks and publishes what the run cost. Here is the total for the full Intelligence Index, with each model's score in front.

• 46.4Grok 4.7 xhigh$4,967
• 45.4Qwen3.8 Max$4,935
• 46.3Grok 4.7 high (default)$3,881
• 43.6Kimi K3 max$3,658
• 44.8GLM-5.3 max$2,503
• 44.3Grok 4.6 high$2,352
• 46.8Claude Sonnet 5.5 high$1,215
• 39.5DeepSeek V4.1 Flash max$477
• 46.3MiMo-V2.6-Pro$207
• 37.9MiMo-V2.6-Flash$109

Artificial Analysis Intelligence Index v4.3, score and total cost to run the index in USD, from the model pages, read October 4, 2026. Grok 4.7 is measured on xAI's API, MiMo on Xiaomi's.

At high, which is what the xAI API runs if you don't set effort, Grok 4.7 scored 46.33 for $3,881. MiMo scored 46.32 for $207. That's 18.8x the money for a hundredth of a point. Turning Grok up to xhigh adds 0.11 points and $1,086, which nobody should pay for.

The breakdown on the xhigh run shows where the money goes. Grok spent $1,442 on output and $3,525 on input and cached tokens. MiMo spent $125 and $82. So the output side is 11.5x, which is the 6.9x price gap times the 1.67x extra output tokens Grok produced (240.4M against 143.7M). The input side is 43x. If you're pricing the switch off the output column, you're missing most of it.

Grok 4.7 costs 65% more than Grok 4.6 at the same price

This is the part xAI's "same price and speed as Grok 4.6" line leaves out. Both at high effort, Grok 4.6 used 94.2M output tokens on the index and Grok 4.7 used 197.7M, 2.1 times as many. The run went from $2,352 to $3,881. The score went from 44.31 to 46.33.

Two points for 65% more is not a terrible trade at the top of a leaderboard. But if you upgraded the model string and kept the budget, your bill is about to tell you it wasn't free. Xiaomi went through the same thing more violently: MiMo-V2.5-Pro scored 26 for $138, and V2.6-Pro scores 46 for $207. Same card, 50% more per run, 20 more points.

What an agent turn costs on each

Take a coding-agent turn with a 100,000-token cached prefix, 2,000 new input tokens and 1,000 output tokens. On Grok 4.7 that's $0.050 for the cache read, $0.004 for the new input and $0.006 for output: $0.060. On MiMo-V2.6-Pro it's $0.00036, $0.00087 and $0.00087: about $0.0021. A thousand turns is $60 against $2.10, three and a half cents on the dollar, on a workload where the output price barely matters.

Then there's xAI's 200K line. A 199,000-token prompt with 5,000 tokens out costs $0.428 on Grok. Make it 201,000 and the whole request flips to the long-context rate: $0.864, double, for 2,000 more tokens. Bedrock's Grok 4.7 card shows no such tier, so long prompts are one reason to buy Grok there. MiMo has no line. The same two requests cost $0.091 and $0.092 anywhere up to its 1M window.

You can run your own mix through the calculator. If your prompts get anywhere near 200K, price the Grok side at the long-context rate as well.

Where Grok earns some of it back

Speed. Artificial Analysis clocks Grok 4.7 at about 78 output tokens a second at high, against about 40 for MiMo on Xiaomi's own endpoint. End to end, a Grok response took 42.2 seconds and a MiMo one 66.6. Ignore the time-to-first-token column when you compare them: Artificial Analysis records no separate reasoning phase for Grok, so its 35.8 seconds includes the thinking, while MiMo's 4.0 is the first reasoning token and its first answer token lands at 54.1.

Open weights cut into that advantage. PrimaLabs serves MiMo-V2.6-Pro at $0.43 and $0.87, within half a cent of Xiaomi, and measured 256.8 tokens a second with the first answer token at 8.5 seconds. That's faster than Grok at about a fifth of Grok's input price. Novita, DeepInfra and GMICloud sell it too. Grok 4.7 has more homes than it did a month ago, with xAI, Bedrock and Google's Vertex partner catalogue, but all of them sell the same weights at xAI's card or above. Azure Foundry stops at Grok 4.6.

On the coding benchmarks each lab publishes, they're close. xAI reports 71.0 on DeepSWE v1.1 and Xiaomi reports 71.9. On Terminal-Bench 4.0, xAI claims 37.6 and Xiaomi 34.9, though Artificial Analysis's own run put Grok 4.7 at 25.8. We wouldn't pick either model on those numbers alone.

The row above both of them

Look at the chart again. Claude Sonnet 5.5 at high scored 46.75, a little above both, for $1,215. That's about a third of Grok 4.7 at high, at 97 tokens a second. It is still close to six times what MiMo cost.

So Grok 4.7 is squeezed from both sides. If you want a closed model from a US lab at this score, Sonnet 5.5 is cheaper and faster on this index. If you want the cheapest 46 there is, MiMo-V2.6-Pro is, and it runs on hardware you choose. We'd want a specific reason to pick Grok 4.7: committed spend on xAI or Bedrock, a Grok-specific tool, or Cursor, where Grok 4.7 Fast lives.

What we read

• xAI: API pricing and Grok 4.7 model page - rates, 200K tier, regional and priority multipliers, Batch support, effort levels
• xAI: Grok 4.7 announcement - launch date, pricing statement, xAI-reported benchmarks
• AWS: Grok 4.7 on Bedrock - Global and US prices, Flex tier
• Xiaomi: MiMo-V2.6 launch and API model catalogue - prices in USD and CNY, context, cache-write policy, UltraSpeed
• Hugging Face: MiMo-V2.6-Pro model card - parameters, licence, Xiaomi-reported benchmarks
• Artificial Analysis: Grok 4.7, MiMo-V2.6-Pro and MiMo providers - index scores, cost to run, token counts, speed by host
• OpenRouter: MiMo-V2.6-Pro - third-party endpoints and prices