Kimi K3 vs Fable 5 vs Sol: Benchmarks and Pricing

Kimi K3 just dropped and the numbers are genuinely surprising. It's 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, real-world testing, pricing, where it falls short, and whether you should switch.

Kimi K3 by the Numbers

K3 scores 57 on the Intelligence Index, placing it third overall behind Fable 5 (60) and Sol (59). That 3-point gap to Fable 5 is the tightest an open model has ever come to the frontier. For comparison, K2.6 scored 44. That's a 13-point jump in one generation.

Why K3 Matters: The Open-Weight Story

This is a 2.8 trillion parameter model with full weights releasing by July 27. The previous leading open-weight models were Kimi K2.6 (44), DeepSeek V4 Pro (~44), and Thinking Machines' Inkling (41). K3 blows past 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.

This is the first time an open-weight model has taken #1 on this leaderboard. 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 Terminal Bench, Program Bench, and SWE Marathon. Fable 5 leads on FrontierSWE and DeepSWE. Sol leads on Terminal Bench (narrowly) and DeepSWE.

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: each model was tested in its own harness (K3 in KimiCode, Fable 5 in Claude Code, Sol in Codex). 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. Early hands-on testing from developers tells the rest.

Long-horizon coherence is the standout. In early testing, K3 completed over 120 tasks from a single paragraph prompt during a 3-hour codebase porting session before hitting context limits. No open-weight model has come close to staying coherent for even half that length on real-world migration work.

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. This is where K3 goes beyond what other frontier models focus on. It can generate 3D models, textures, and fully playable browser games. Early testers report it creating fish models, submarines, and functional 3D environments from text prompts. Anthropic and OpenAI have focused heavily on code completion. Moonshot appears to have trained K3 on a broader capability set, including chip design, kernel optimization, and video editing workflows.

Kernel optimization results are notable. K3 was used internally by Moonshot to optimize GPU kernels during training. After 15 hours of optimization work, it slightly exceeded Fable 5's improvements across four kernel optimization categories. This is one area where Anthropic has explicitly restricted Fable and Mythos. An open-weight model this capable at ML infrastructure work has real implications.

Token efficiency improved 21% over K2.6. K3 uses roughly 23,000 output tokens per Artificial Analysis task. That's more than Sol's ~15,000 but significantly less than DeepSeek V4 Flash's ~45,000. Combined with cheaper per-token pricing, the actual cost per task ends up competitive with Sol.

Where K3 Falls Short

Being honest about the gaps:

Hallucination rate is high. AA-Omniscience scores 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 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 62 tok/s, K3 is the slowest of the three frontrunners. Fable 5 runs at 71 tok/s and Sol at 85 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. Early users report another quirk: when sub-agent workflows finish, K3 sometimes responds to the workflow status rather than to you, the person who kicked it off. It'll say "no new report here, just an idle status ping" instead of summarizing what it found. These are solvable with strict system prompts, but the polish gap with Fable 5 and Sol is real. Moonshot admits the model "still has a noticeable gap in user experience" compared to the frontier.

Subscription context limits. The 1M token context window is available via the API, but the Kimi subscription tiers ($20/$40/$100/$200) don't give you the full window in all configurations. The $40 tier can burn through its usage window quickly under heavy coding. The $200 tier is much more generous.

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):

K3 offers the best intelligence-per-dollar at the frontier tier. Sol is close at $1.04. Fable 5 costs roughly 3.7x more per task.

Real-world cost example: A full day of heavy K3 coding (migrations, security audits, UI redesigns) came to roughly $63 in API usage. That compares favorably to $300-400/day on Fable 5 for similar work volume.

One thing to watch: the $15/1M output cost isn't cheap. K3 uses roughly 2x the output tokens Sol uses per task, which cancels out much of the per-token discount. 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. Images and video are encoded via a hierarchical patch encoder and processed alongside text in the same decoder. 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 PerceptionBench. 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

Moonshot says the full technical report drops alongside the weight release on July 27. We'll update this post when it lands.

How to Use Kimi K3 Today

Until the weights release on July 27, you have two options:

Kimi subscription (kimi.com): Plans at $20, $40, $100, and $200/month. The $40 tier can burn through its 5-hour usage window in about 30 minutes of heavy K3 usage. The $200 tier is much more generous and hasn't been exhausted by developers doing sustained coding work.

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 only runs at max thinking effort. Low and high effort modes are coming in a future update. This means you can't dial down reasoning for simple tasks yet, which contributes to the speed and cost overhead.

Kimi K3 Weights Download and Self-Hosting

The full 2.8T parameter weights will be available by July 27, 2026. Expect them on Hugging Face, though downloading 2+ terabytes will take some time.

Hardware requirements are steep. Moonshot recommends 64+ accelerators in a supernode configuration. At FP8 precision, the model is roughly 1.4TB. That's not fitting on any consumer hardware. To put it in dollar terms, 64 H100s represents roughly $2.6 million in hardware costs.

Don't expect third-party hosting to be dramatically cheaper. When K2.6 weights landed on providers like Together and Fireworks, pricing was only 11-15% below Moonshot's own API. The model is simply massive, and that hardware cost gets passed through.

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. Wait for community quantization benchmarks after July 27 before committing infrastructure.

Security Implications

A frontier-class model with no usage restrictions changes the security landscape. Both Anthropic and OpenAI have invested heavily in preventing their models from assisting with offensive security work. Fable and Mythos explicitly restrict kernel optimization and exploit discovery. Sol has similar guardrails.

K3 has none of these restrictions. Early testing shows it will readily perform security audits, spinning up multiple sub-agents for discovery and verification passes. That's great for defenders, but the same capability will be available to attackers once the weights release.

Moonshot's blog post for K3 doesn't mention "safety" at all, and "security" only appears in demo descriptions. No system card has been published. The technical report due July 27 may address this, but for now there's no public information about safety evaluations or training-time mitigations.

The Competitive Landscape

K3's arrival reshuffles the top of the leaderboard. Here's where everything stands as of July 17, 2026:

K3 slots between Sol and Terra on intelligence but offers the cheapest output pricing of the top 3. 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), 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.

Wait if: You want to self-host but don't have 64+ accelerators. The model weights drop July 27. Wait for 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:

K3 costs 40% less than Sol per task ($0.94 vs $1.04) while scoring 2 points lower on Intelligence. On coding, K3 beats Sol on several benchmarks (Program Bench, SWE Marathon) but trails on DeepSWE and the Coding Agent Index overall (76 vs 80). 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 59 versus K3's 57. But K3 beats Sol on the Frontend Code Arena (1679 vs 1618) and several coding benchmarks. K3 is also cheaper ($3/$15 vs $5/$30) and will be 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 will be released by July 27, 2026, meaning 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. That makes K3 roughly 2x cheaper than Sol per Intelligence Index task.

Can Kimi K3 process images?

Yes. K3 has native multimodal vision trained into the architecture. It handles images and video through a hierarchical patch encoder. It scores competitively with Fable 5 and Sol on vision benchmarks like MMMU-Pro (81.6) and leads on OmniDocBench (91.1).

When will Kimi K3 weights be released?

Moonshot has committed to releasing full model weights by July 27, 2026. They're working with inference partners and open-source maintainers (including vLLM) to ensure compatibility. The technical report will drop alongside the weights.

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 after the weight release, but don't expect to run this on consumer hardware.

Where can I download Kimi K3 weights?

Weights are not available yet. Moonshot will release the full model weights by July 27, 2026. They'll likely be hosted on Hugging Face. The download will be 1.4TB+ at FP8 precision. Third-party inference providers like Together, Fireworks, and others will likely offer hosted versions shortly after.

Is Kimi K3 good for 3D and game development?

Early results are surprisingly strong. K3 can generate 3D models, textures, and playable browser games from text prompts. It also scores well on browser and computer use benchmarks. This is a capability gap where Anthropic and OpenAI haven't focused, making K3 the leading option for 3D/game dev tasks.

Data from Artificial Analysis Intelligence Index v4.1, Arena.ai Frontend Code Arena, and Moonshot AI official benchmarks. Updated July 17, 2026. See the LLM Leaderboard for live rankings or our Best LLM in 2026 guide for the full comparison.