Kimi K3 vs Fable 5 vs Sol: Benchmarks and Pricing

Kimi K3 just dropped and the numbers are genuinely surprising. At launch, it was the first open-weight model to score within 3 points of Claude Fable 5 on the Artificial Analysis Intelligence Index. And it took the #1 spot on Arena.ai's Frontend Code Arena, beating Fable 5 outright.

No open model has been this close to the frontier before. K2.6 was good. K3 is in the conversation with Sol and Fable 5. And at $0.94 per task versus Fable 5's $3.50, the pricing makes this a serious contender for production workloads.

Here's the full breakdown: benchmarks, Moonshot's own demos, pricing, where it falls short, and whether you should switch.

Update (September 24, 2026): Kimi K3 now scores 44 on the Artificial Analysis Intelligence Index. AA rebased the index in September 2026 (now v4.3.2), so every model scores lower than at launch, and the comparison tables below use the new numbers. K3 is no longer the top open-weight model on AA's index. Xiaomi's MiMo-V2.6-Pro scores 46 at $0.43/$0.87, and Z AI's GLM-5.3 scores 45 at $1.40/$4.40, both far cheaper than K3's $3/$15. K3 still codes well: it scores 52 on AA's new Coding Agent Index v1.5, two points behind GLM-5.3 (54). Launch-week cost per task and Moonshot's own benchmarks are left as published in July. For live rankings, see the LLM Leaderboard.

Update (September 26, 2026): Through the API, Kimi K3 takes over a minute for a full answer at every tested setting. For current open-weight picks with real costs per task, see the LLM leaderboard.

Kimi K3 by the Numbers

K3 launched at 57 on AA's old v4.1 scale, third overall behind Fable 5 (60) and Sol (59). That 3-point gap to Fable 5 was the tightest an open model had come to the frontier. K2.6 scored 44 on the same scale, so K3 was a 13-point jump in one generation. On today's v4.3.2 index, K3 scores 44, Fable 5 50, and GPT-5.6 Sol 47.

Why K3 Matters: The Open-Weight Story

This is a 2.8 trillion parameter model, and the full weights are now public. At launch, the previous leading open-weight models were Kimi K2.6 (44), DeepSeek V4 Pro (~44), and Thinking Machines' Inkling (41) on the v4.1 scale. K3 beat all of them by 13+ points.

What "open weight" means for you:

• You can self-host and run it on your own infrastructure
• No API dependency on Moonshot
• Fine-tuning and customization become possible
• The community can optimize inference (vLLM support is already in progress)
• No vendor lock-in on a frontier-class model

The catch: 2.8T parameters is enormous. You'll need significant hardware to run it. Moonshot recommends 64+ accelerators in a supernode configuration. This isn't something you'll run on a single GPU. But for teams with the infrastructure, or using inference providers that host it, the economics change significantly.

Coding Performance

This is where K3 makes its strongest case.

Frontend Code Arena #1. K3 debuted at the top of Arena.ai's Frontend Code Arena with 1679 Elo, beating Fable 5 (1631) and Sol (1618). It ranked first in 6 of 7 domains: Brand & Marketing, Reference-Based Design, Data & Analytics, Consumer Product, Simulations, and Content Creation Tools. Fable 5 held the top spot only in Gaming.

Arena called it a 17-place jump from Kimi K2.6 (#18 to #1). Arena evaluations are blind human judgments, not self-reported benchmarks, which makes the result harder to dismiss.

Benchmark breakdown from Moonshot's own testing:

K3 leads on Program Bench and SWE Marathon. Sol leads on Terminal Bench (narrowly) and DeepSWE. Fable 5 leads on FrontierSWE.

The pattern: K3 is strongest on long-horizon coding tasks. SWE Marathon, which tests sustained multi-hour engineering sessions, is where K3 beats both Fable 5 and Sol. For quick, contained tasks, Fable 5 and Sol still edge ahead.

One caveat: harnesses differ by benchmark. K3 mostly ran in Kimi Code, and the others in Claude Code, Codex or Terminus 2, per Moonshot's footnotes. Moonshot also notes Fable 5 hit fallbacks on 35% of SWE Marathon tasks. Harness quality affects results. The Arena score is the more apples-to-apples comparison since it uses blind human evaluation.

Real-World Coding: What Using K3 Feels Like

Benchmarks tell one part of the story. Moonshot's own demos show the rest.

Long-horizon coherence is the standout. In Moonshot's demos, K3 designed a small chip in a single 48-hour autonomous run, and reproduced an astrophysics result in about two hours that Moonshot says would take an experienced researcher one to two weeks.

FrontierSWE score hints at code taste. K3 scored 81.2 on FrontierSWE, slotting between Fable 5 (86.6) and Sol (71.3). That benchmark focuses on whether code is likely to merge, not just whether it solves the problem. A 10-point lead over Sol suggests K3 writes code that looks and feels better, not just code that works.

3D and game dev capabilities. Moonshot's launch post shows K3 turning concepts, images and videos into playable browser games, from a voxel colosseum to an FPS arena. It also demos chip design, kernel optimization and video editing.

Kernel optimization results are notable. Moonshot gave each model up to 24 hours in an identical sandbox to optimize four GPU kernels. K3 performed about as well as Fable 5 and clearly beat Opus 4.8 and GPT-5.6 Sol. Late in development, an early K3 handled most of Moonshot's own kernel work. An open-weight model this capable at ML infrastructure work has real implications.

Token use runs a little high. Artificial Analysis says K3 at max effort generated 160M output tokens running its Intelligence Index, somewhat above the 140M median for open-weight models of similar size (AA's K3 page). Cheaper per-token pricing makes up for it: its cost per task ends up level with GPT-5.6 Sol's.

Where K3 Falls Short

Being honest about the gaps:

Hallucination rate is high. At launch, AA-Omniscience scored K3 at +18, which is a big improvement over K2.6's +6. But the hallucination rate rose from 39% to 51%. K3 answers more questions correctly (accuracy went from 33% to 46%), but when it's wrong, it's confidently wrong. Fable 5 leads here with a 16% hallucination rate. For high-stakes factual work where wrong answers are costly, Fable 5 is still the safer pick.

Agentic performance gap. GDPval-AA v2 (AA's scores, as cited in Moonshot's table) shows K3 at 1668 Elo versus Fable 5's 1760 and Sol's 1748. A 92-point gap on agentic tasks is real. If your workflow is heavily agent-based with multi-step tool calling, Fable 5 and Sol remain ahead.

Speed. At 37 tok/s on AA's current measurement, K3 is the slowest of the three. Fable 5 runs at 67 tok/s and GPT-5.6 Sol at 73 tok/s. For interactive use where you're waiting on responses, that difference is noticeable.

Excessive proactiveness and UX roughness. Moonshot themselves flag this limitation. K3 was trained heavily on long-horizon tasks and sometimes makes unexpected decisions on your behalf during ambiguous situations. Moonshot's fix is explicit limits in the system prompt or AGENTS.md. It also says K3 shows a noticeable gap in user experience compared with Fable 5 and GPT-5.6 Sol, and that quality can become unstable if a harness drops its thinking history.

Subscription limits. The full 1M-token context window is available through the API. Kimi's subscription plans meter usage separately, so check the current tiers before you rely on one for heavy coding.

Kimi K3 Pricing and Cost Breakdown

K3's API pricing on the official Kimi API:

Moonshot reports 90%+ cache hit rates in coding workloads. That means effective input cost drops to roughly $0.57/1M for repeated or related queries. For sustained coding sessions where you're iterating on the same codebase, the cache economics are strong.

Cost comparison per Intelligence Index task (from Artificial Analysis):

In July, K3 offered the best intelligence-per-dollar at the frontier tier. Sol was close at $1.04, and Fable 5 cost roughly 3.7x more per task. Among open models, that title now belongs to MiMo-V2.6-Pro: 46 on the index at $0.43/$0.87 per million tokens.

Today's numbers: On AA's current index at max effort, K3 costs about $2.00 per task, level with GPT-5.6 Sol ($1.99) and less than a quarter of Fable 5 ($8.75).

One thing to watch: the $15/1M output cost isn't cheap. K3 matches Sol's cost per task despite lower token prices because it burns more tokens per task. The savings come from the cache hit rate and the lower input pricing, not from output efficiency.

Use our cost calculator to estimate your monthly spend based on volume.

Vision Capabilities

K3 has native multimodal vision built into the architecture: one model handles text, images and video. This isn't bolted-on vision. It's trained from the ground up.

The vision benchmarks are competitive:

K3 leads on OmniDocBench (document understanding) and beats Fable 5 on PerceptionBench, though Sol edges it there. For image analysis workflows, it's a viable alternative to Sol or Fable 5. See our full vision model rankings for a broader comparison.

Architecture: What's Under the Hood

For those interested in the technical details:

• 2.8T total parameters with Mixture of Experts (MoE), activating 16 of 896 experts
• Kimi Delta Attention (KDA): An attention mechanism designed for efficiency at scale
• Attention Residuals (AttnRes): Selectively retrieves representations across depth rather than accumulating them uniformly
• Stable LatentMoE: The sparsity framework that keeps 896 experts balanced during training
• Quantization-aware training from the SFT stage using MXFP4 weights with MXFP8 activations
• 2.5x scaling efficiency improvement over K2, meaning the model converts compute into intelligence more effectively

The technical report is linked from the Hugging Face model page. We haven't reviewed it for this update.

How to Use Kimi K3 Today

If you don't want to self-host, you have two options:

Kimi subscription (kimi.com): Paid monthly plans give you K3 in Kimi's own apps. Usage limits differ by plan, so check the current tiers before you rely on one for heavy coding.

Kimi API (platform.kimi.ai): Standard pay-per-token pricing at $3/$15. You can bring the API key into tools like OpenCode, which also lets you use K3 inside other harnesses. The API gives you the full 1M context window.

At launch, K3 ran only at max thinking effort, with low and high modes promised later. AA now also lists a low setting: it scores 34.5 against 43.6 at max and still takes over a minute per answer, so dialing reasoning down doesn't fix the speed.

Kimi K3 Weights Download and Self-Hosting

The full 2.8T parameter weights are out on Hugging Face. Downloading well over a terabyte takes a while, so most teams start with a hosted provider: Artificial Analysis now tracks K3 on Fireworks, Baseten, Nebius, Amazon Bedrock and others.

Hardware requirements are steep. Moonshot recommends 64+ accelerators in a supernode configuration. In its native MXFP4 weight format, 2.8T parameters come to roughly 1.4TB. That's not fitting on any consumer hardware.

Don't assume third-party hosting is cheaper. AA lists more than a dozen providers for K3, and prices vary between them, so compare before you pick one (AA's provider comparison).

The real value of open weights isn't cheaper hosting. It's control: running on your own infrastructure, fine-tuning for your use case, and not sending proprietary code to a third party. Check community quantization benchmarks before committing infrastructure.

Security Implications

A frontier-class model with no usage restrictions changes the security picture. Anthropic and OpenAI both put cyber safeguards on their frontier models. Anthropic offers its Mythos models only by invitation, and Moonshot's own footnotes note that 10% of one in-house benchmark's tasks tripped GPT-5.6 Sol's cyber guard.

Once K3 runs on your own hardware, nobody can enforce restrictions like these. That's useful for defenders running security audits, and the same capability is available to attackers now that the weights are public.

Moonshot's launch post doesn't mention safety or security at all, and we haven't found a system card. Check whether the technical report covers safety evaluations or training-time mitigations before deploying.

The Competitive Landscape

K3's arrival reshuffled the top of the leaderboard in July. Here's where it stands as of September 26, 2026, on AA's v4.3.2 index:

K3 still slots between GPT-5.6 Sol and Terra, but it's no longer near the top, and two newer open-weight models score higher for far less money. For the latest rankings on specific tasks, use our LLM Leaderboard or the model selector to find the right fit.

Should You Switch to Kimi K3?

Switch if: You need frontier coding at lower cost than Fable 5 or Sol, you're building frontend applications (Arena #1 at launch), you want open weights for self-hosting or customization, or you do sustained long-horizon coding sessions where K3's SWE Marathon score matters.

Stay on Fable 5/Sol if: You need the lowest hallucination rate (Fable 5 wins by a mile), your workflow is heavily agentic with complex tool calling, you need faster response times for interactive work, or you rely on the Claude/OpenAI ecosystem and tooling.

Look at MiMo-V2.6-Pro or GLM-5.3 first if: You want open weights and the best score per dollar. Both outscore K3 on the current index at a fraction of its price. GLM-5.3 also edges K3 on AA's Coding Agent Index v1.5 (54 vs 52).

Wait if: You want to self-host but don't have 64+ accelerators. The weights are out, but check community benchmarks on inference optimization and quantized versions before committing infrastructure.

How Kimi K3 Compares to GPT-5.6

Since the GPT-5.6 Sol vs Terra vs Luna comparison is our most-read post, here's the direct comparison:

At launch, K3 cost about 10% less than Sol per task ($0.94 vs $1.04). On today's index it scores 3 points lower (44 vs 47), and it's cheaper per token ($3/$15 vs Sol's current $4/$20). On coding, K3 beats Sol on several benchmarks (Program Bench, SWE Marathon) but trailed on DeepSWE and the July Coding Agent Index overall (76 vs 80). GPT-6 Sol now scores 48 at $2/$10, which undercuts K3 on both input and output. The biggest K3 advantage over all GPT-5.6 variants is open weights. You can't self-host Sol.

For the full AI coding model rankings, see our dedicated comparison.

FAQ

Is Kimi K3 better than GPT-5.6 Sol?

On the Intelligence Index, no. Sol scores 47 versus K3's 44 on the current Artificial Analysis index (v4.3.2). But K3 beats Sol on the Frontend Code Arena (1679 vs 1618) and several coding benchmarks. K3 is also cheaper ($3/$15 vs $4/$20) and open weight. For frontend development and long-horizon coding, K3 has a case. For general intelligence and agentic tasks, Sol leads.

Is Kimi K3 open source?

K3 is open weight, not open source. The model weights are public, so you can download and run the model. But the training code, data, and methodology are proprietary. The weights release still allows self-hosting, fine-tuning, and community optimization.

How much does Kimi K3 cost?

Via the official Kimi API: $3.00/1M input tokens (cache miss), $0.30/1M (cache hit), $15.00/1M output tokens. With Moonshot's reported 90%+ cache hit rate in coding, effective input cost drops to around $0.57/1M. At launch, K3 came out slightly cheaper than Sol per Intelligence Index task ($0.94 vs $1.04).

Can Kimi K3 process images?

Yes. K3 has native multimodal vision trained into the architecture, and it takes both images and video. In Moonshot's launch table it scores competitively with Fable 5 and Sol on vision benchmarks like MMMU-Pro (81.6) and leads on OmniDocBench (91.1).

Are Kimi K3 weights available?

Yes. The full weights are public, and Moonshot worked with inference partners and open-source maintainers (including vLLM) on compatibility. Hosted versions run on Fireworks, Baseten, Nebius, Amazon Bedrock and others.

Can I run Kimi K3 locally?

Not easily. At 2.8T parameters (even with only 16/896 experts active), K3 requires significant hardware. Moonshot recommends 64+ accelerators in a supernode configuration. Quantized versions from the community may lower the bar, but don't expect to run this on consumer hardware.

Where can I download Kimi K3 weights?

The full weights are public on Moonshot's Hugging Face page. Expect about 1.4TB in the native MXFP4 format. If you'd rather not self-host, Fireworks, Baseten, Nebius and Amazon Bedrock all serve K3.

Is Kimi K3 good for 3D and game development?

It's a strength Moonshot highlights. Its launch post shows K3 building playable browser games from concepts, images and video, and K3 led the Frontend Code Arena at launch. It also tops Moonshot's BrowseComp table (91.2). Test it against Sol and Opus 5.5 on your own project before you commit.

Current scores, prices and speeds from Artificial Analysis Intelligence Index v4.3.2 (September 2026). Launch-era cost per task and Coding Agent Index v1.1 data from July, Arena.ai Frontend Code Arena, and Moonshot AI official benchmarks. Updated September 26, 2026. See the LLM Leaderboard for live rankings or our Best LLM in 2026 guide for the full comparison.