Foto: MiraCosic / Pixabay
Lead
Kimi K3.1 Pro benchmarks released in late 2026 show a mixed picture: the model scores 55 on the Artificial Analysis Intelligence Index, yet reaches 98 % on Plotly’s data‑analytics exam and 75.5 % on DeepSWE 1.1. These numbers give developers a concrete sense of where the Chinese flagship LLM excels and where it still lags.
What Does the Artificial Analysis Intelligence Index Tell Us?
The Artificial Analysis Intelligence Index (v4.3.2) rates Kimi K3.1 Pro at 55, placing it in the mid‑range of evaluated models. This index aggregates reasoning, knowledge‑retrieval and multilingual capabilities into a single score, so a 55 suggests solid but not leading‑edge general intelligence. The figure comes from the model’s provider page, which lists the index alongside pricing and release data.
How Strong Is Kimi K3.1 Pro at Data‑Analysis Tasks?
Plotly’s dedicated data‑analytics benchmark ran the model through 43 questions; Kimi K3.1 Pro passed 42 of 43 (98 %), earning an “A” grade and tying for third place among 34 models, behind only GPT‑6.1 Sol and GPT‑6 Astra. This result demonstrates that the model can reliably generate correct Plotly visualisations and manipulate data frames, a valuable trait for analysts and data‑science teams.
Coding‑Centric Benchmarks: Terminal‑Bench, SWE‑Marathon and DeepSWE
Three software‑engineering suites probe the model’s ability to write, edit and reason about code:
- Terminal‑Bench 4.0 – a resolution‑rate test where Kimi K3.1 Pro achieves 65.7 %. The benchmark measures how often the model can complete a terminal‑style coding task without human intervention.
- SWE‑Marathon – a Pass@1 metric of 51.7 %, indicating that just over half of the model’s first‑try solutions solve the given software‑engineering problems.
- DeepSWE 1.1 – the strongest showing, with 75.5 % Pass@1, suggesting the model handles deeper, multi‑file reasoning better than the broader SWE‑Marathon set.
Together, these scores reveal a model that is competent at isolated coding tasks (DeepSWE) but less consistent on longer, marathon‑style engineering challenges.
Office‑Productivity and Internal Code Benchmarks
Moonshot, the creator of Kimi, reports two additional figures:
- OfficeQA Pro v2 – a claimed state‑of‑the‑art score of 64.4, targeting question‑answering over office documents.
- Kimi Code Bench 3.0 (long‑horizon) – an internal test where the model reaches 61.2 %, reflecting its ability to sustain reasoning across extended coding horizons.
While these numbers are self‑reported, they align with the stronger DeepSWE result and suggest the model is tuned for document‑centric and long‑range coding workloads.
Pricing Context for Decision‑Makers
The provider lists $2.00 per million input tokens and $10.00 per million output tokens. When weighed against the benchmark profile—high data‑analysis accuracy, solid DeepSWE performance, moderate general‑intelligence score—teams can estimate cost‑per‑successful‑task for their specific workloads.
Practical Next Step for Developers
Run a small‑scale pilot on your own codebase using the Terminal‑Bench or DeepSWE style tasks you care about most. Compare the model’s Pass@1 rate and token cost against the current provider you use; the 75.5 % DeepSWE Pass@1 and low input price make Kimi K3.1 Pro a strong candidate for automated refactoring pipelines, while the 55 intelligence index flags the need for human oversight on open‑ended reasoning tasks.