A new AI model just grabbed the number four spot on the most respected independent benchmark, and it costs half of what Claude Opus 5.5 charges. Kimi 3.1 Pro is Moonshot’s newest release, and it might be the best cheap AI model you can try right now. There’s a catch though, and it involves a download measured in terabytes.
What Kimi 3.1 Pro actually is
Kimi 3.1 Pro is the latest model from Moonshot AI, the Chinese lab behind the Kimi chatbot and the Kimi Code coding assistant. It launched on October 23 with the usual launch-day fanfare, but this time the numbers have some independent receipts behind them.
Under the hood, it’s a mixture-of-experts model with 2.8 trillion total parameters and about 104 billion active at any moment. If that sounds like gibberish, here’s the plain-English version: the model is enormous, but for each question it only switches on the small slice of itself that’s relevant. So you get big-model brains without paying to run the whole thing on every prompt. It also handles a 1 million token context window, which means it can read an entire codebase or a small novel in one go without forgetting the start.
Moonshot describes it as a heavily post-trained version of Kimi K3. Basically they took last year’s model and poured far more training compute into refining how it follows instructions, writes code, and handles long tasks. That approach showed up in the results, just not evenly everywhere.
The benchmark numbers that matter
The most credible independent scorecard here is the Artificial Analysis Intelligence Index, which runs models through a battery of real tasks. Kimi 3.1 Pro scored 55 and landed in fourth place overall. For context, Claude Opus 5.5 sits at 58, Claude Sonnet 5.5 and Meta’s Muse Wonder both hit 56, and Claude Fable 5.1 scored 53.
Read that again. An open-weights-bound model from a lab most Americans couldn’t name scored above Fable 5.1 and within three points of the most expensive Claude. That’s the real story here: the gap between the frontier and everyone else just shrank to almost nothing. When the weights land, Artificial Analysis says the lead of the best closed model over the best open one drops from 12 points to 3.
Here’s how the two models you’d actually pick between stack up:
| Kimi 3.1 Pro | Claude Opus 5.5 | |
|---|---|---|
| Artificial Analysis Intelligence Index | 55 | 58 |
| Input price (per 1M tokens) | $2 | $4 |
| Output price (per 1M tokens) | $10 | $20 |
| Open weights | Coming Nov 4, 2026 | No |
Coding results back this up too. On WebDev Arena, where models build real websites and users vote on the better result, Kimi 3.1 Pro jumped to 8th place with 1,692 points, up nine spots from its predecessor. Moonshot also claims 65% fewer hallucinations on financial work, though that one is a vendor claim. Independent testing on a different benchmark (AA-Omniscience) still found hallucinations dropping from 53% to 44%, so the direction is real even if the exact number varies.
What it costs (and the trap in the pricing)
Standard API pricing is $2 per million input tokens and $10 per million output tokens. Cached input costs $0.20 per million. Compare that with Opus 5.5 at $4 and $20, and the appeal is obvious if you’re building anything that chews through tokens.
There’s also a Highspeed tier that generates output 2 to 3 times faster, and it costs exactly twice as much per token. For chat, standard is fine. For long-running agents where you’re burning the clock, the speed bump can pay for itself.
But here’s the trap, and it’s the part launch-day hype skips: cheaper tokens don’t automatically mean cheaper finished work. On the Intelligence Index tasks, the model generated about 55,000 output tokens per task, roughly 14% more than Kimi K3 did. So the cost per completed task stayed basically flat at about $1.98, even though the per-token price dropped by a third. On the cybersecurity suite it actually cost more per task than K3 because it consumed twice the input tokens. If you’re budgeting a project, budget in tasks completed, not tokens purchased.
Where it actually shines: agents and 3D
Independent testers noticed something interesting about where the gains landed. Kimi 3.1 Pro didn’t get uniformly smarter across the board. Its biggest jumps were on agentic tasks, the long multi-step jobs where a model has to plan, use tools, and keep going without hand-holding. Knowledge and science questions improved less. So if your mental model of AI is “ask a question, get an answer,” the upgrade will feel modest. If you run agents that work for hours, it’s a different story.
Moonshot leaned into that with a new service called Kimi Hosted Agents. It’s a managed environment where your agent runs around the clock even with your laptop closed, and it supports multiple agents collaborating on the same job. For anyone automating repetitive work (monitoring, report generation, scheduled research), this is the part of the launch that matters most, and it’s live now on the standard plans.
The flashy demos are worth a look too. Moonshot showed the model building entire 3D scenes in Blender, and early users posted playable 3D websites and interactive animations built from a single prompt. Independent rankings on 3D tasks are still being tallied, so treat the demos as promising rather than proven.
The catch: you can’t run it at home
Not yet, and honestly not ever on consumer gear. Those 2.8 trillion parameters add up to roughly 5.6 terabytes at full precision, or about 1.4 terabytes even heavily compressed. Nobody’s spare bedroom has that much VRAM.
The open weights arrive November 4, 2026, which is still a big deal for developers and companies who want to self-host on serious hardware or fine-tune the thing. Just know that “open weights” this time means “needs a small rack of accelerators,” not “runs on your gaming PC.” If you want an open model that actually fits on hardware you own, DeepSeek V4.1 Flash is the one to try first, or check our guide on open vs closed models before picking a side.
One more honesty note: the launch attracted allegations of undisclosed paid promotion on social media. Those claims are about marketing behavior, not about the benchmark results, which came from independent testers. Still, it’s worth knowing the hype around this launch isn’t all organic.
Should you actually switch?
Here’s my take. If you write code or run agents and you’re watching your API bill, Kimi 3.1 Pro is absolutely worth trying this week. You can use it free on kimi.com, or plug the API into your tools and see if the quality holds on your actual workload. For the price of half an Opus, you’re getting a model that lands within three points of it on independent testing.
I wouldn’t cancel a Claude subscription over this. Opus 5.5 still scores higher, and workflow familiarity counts for something. But for any high-volume, cost-sensitive job (batch summarizing, bulk code review, data extraction), Kimi is now the default recommendation over its own predecessor and most mid-tier models.
The takeaway
Kimi 3.1 Pro just made frontier-level AI a budget option, and the open weights land November 4. Try it free at kimi.com today, and if it holds up on your work, check the API pricing before your next invoice from anyone else arrives.
