Artificial Analysis: MiMo-V2.6-Pro Intelligence Index, Speed, Pricing, and Latency

One-sentence conclusion

Under the Xiaomi provider, reasoning variant, and default production workload of 10,000 input tokens, Artificial Analysis measured MiMo-V2.6-Pro at an Intelligence Index of 46.324 (displayed as 46), output speed of 129.75 tokens/s, 17.58 seconds to the first answer token, and 21.44 seconds for an end-to-end 500-token output; pricing is in the low-cost range, but reasoning wait time accounts for most of the time to the first answer token.

Suitable use cases

• Suitable tasks: Tasks requiring strong general reasoning, long context, multimodal input, or Agent workflows while also prioritizing API unit cost and output speed; the AA page lists text, image, audio, and video as input modalities, with a 1M-token context window.

• Unsuitable tasks: Interactive short-form Q&A where time to first response is highly sensitive; rigorous cross-provider, cross-reasoning-level, or cross-random-seed comparisons; this page has only one Xiaomi provider and cannot represent other deployments.

• Applicable model version: The reasoning version of MiMo-V2.6-Pro. The results must not be extrapolated to MiMo-V2.6-Pro-RL, MiMo-V2.6-Pro-Ultraspeed, MiMo-V2.6-Flash, or other non-reasoning variants.

• Applicable client, Agent, or API: Artificial Analysis's independent benchmark environment and the Xiaomi API provider listed on the page; the page does not publish a complete, directly reusable request payload, tool schema, or evaluation harness.

• Recommended reasoning level and parameters: This source does not publish selectable reasoning levels, temperature, top-p, random seed, or complete API parameters; a retest should fix the reasoning version, Xiaomi provider, and 10,000-input-token workload, while recording each unknown parameter.

Test method / workflow steps

1. Use the Artificial Analysis model page to verify the identity of MiMo-V2.6-Pro, its reasoning label, and its association with the Xiaomi provider; do not mix results for Pro, Flash, RL, or Ultraspeed.

2. Use Artificial Analysis Intelligence Index v4.3.2 as the composite intelligence measure. The page says the index contains 10 evaluations: AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity's Last Exam, GDP.pdf, CritPt, AA-Omniscience, and AA-LCR v1.1.

3. Read the structured Dataset data embedded in the model page and record the index, output-token usage, cost per task, total evaluation cost, and speed; the index page's measurementTechnique explicitly says that Artificial Analysis ran the measurement independently on dedicated hardware.

4. On the Provider Benchmark page, use the default workload of 10,000 input tokens. The page's pricing explanation uses a cache-input-output ratio of 7:2:1 for the blended price; speed is output tokens/s; latency is measured by the first answer token and end-to-end output of 500 tokens.

5. Break down time to the first answer token according to the page's structured data: reasoning time + input processing time = time to the first answer token. Check end-to-end time as input processing + reasoning + output time for 500 tokens.

6. When comparing with other models, retain the reasoning labels shown on the page (for example, max, high, and xhigh) and each provider; do not treat the page ranking as a causal conclusion under the same reasoning level.

Original evidence and data

1. Model identity and measurement scope

2. Intelligence Index, speed, tokens, and cost

3. Latency and end-to-end response

Artificial Analysis's latency data explicitly separates reasoning time from input processing time. Its model-page Dataset gives the following MiMo-V2.6-Pro data:

Under this measurement scope, reasoning wait time accounts for approximately 15.4144 / 17.5817 ≈ 87.7% of the time to the first answer token; this is a calculation from the page's published components and does not represent other request lengths or providers.

4. Model specifications (for interpreting the measurement scope)

• Context window: 1,000,000 tokens.

• Total parameters: approximately 1.0T on the page; the structured model-size Dataset gives 978B passive parameters.

• Active parameters: 42B.

• Input: text, image, audio, and video.

• Output: text.

• License: MIT.

Scope and limitations

• This is an independent Artificial Analysis benchmark result, not a Xiaomi-published benchmark and not a local retest; it cannot support a claim that the same speed will be achieved across all APIs, regions, networks, or loads.

• The page's structured data says that the measurement ran independently on Artificial Analysis dedicated hardware, but it does not publish the complete hardware model, request payload, concurrency, network path, random seed, temperature, top-p, timeout policy, or failed samples. Reproduction can therefore only approximately verify the metrics and workload level.

• The model-page body lists 10 component evaluations for the Intelligence Index, but does not provide MiMo-V2.6-Pro's score for each item, per-item sample counts, confidence intervals, or per-question logs. The 46.324 is suitable as an aggregate comparison signal and should not be decomposed into conclusions about individual capabilities.

• The Provider page has only one provider, Xiaomi. The $0.13 per task, $0.18 blended price, and 17.58-second first-answer latency cannot be extrapolated to unlisted providers.

• The page's prices are public rates per million tokens; the $206.66 cost of the complete Intelligence Index is Artificial Analysis's evaluation expenditure and should not be treated as the price of an ordinary user session.

• The model-page #1 / 114 and the chart's “28 of 656 models” belong to different filtering/display contexts; they must not be combined into a single ranking.

• Artificial Analysis's X account, https://x.com/ArtificialAnlys, was opened through Tabbit, but the timeline returned “It looks like your connection is lost. We’ll keep trying to reconnect.” No post directly related to MiMo-V2.6-Pro was visible, so no data was added from its title, summary, or a paraphrase.

Source excerpts or observations (short excerpts for compliance only)

• The model-page Dataset's public description of the measurement method: Independent test run by Artificial Analysis on dedicated hardware.

• The Provider page notes that the default performance benchmark workload was updated to 10k input tokens.

• The Provider page's structured data also gives Xiaomi's 129.74879744631201 output tokens/s, 15.414401053139381 seconds of reasoning time, and 2.1672836154999802 seconds of input processing time; this piece preserves these original fields to sufficient precision and lists the rounded values separately as page-display values.