Kimi K3 vs Claude Fable 5: Benchmarks, Pricing, and Which Model to Use

What you will learn in this article:

• What Moonshot AI actually shipped on July 16, and why a 2.8-trillion-parameter model can still be practical to serve.
• The launch scorecard: where K3 genuinely beats Fable 5, where it clearly doesn't, and one caveat about hallucination rates.
• What both models cost, and why the 3.3x output-price gap matters more for agents than for chat.
• What the July 27 open-weights release does and does not change if you were hoping to run K3 yourself.
• A workload-by-workload answer to the only question that matters: which one should you use?

Introduction

On July 16, 2026, Moonshot AI released Kimi K3: a sparse mixture-of-experts model with roughly 2.8 trillion total parameters, a 1-million-token context window, native vision, always-on reasoning, and a promise to publish full open weights by July 27. It is the largest model ever announced for an open-weight release, and Moonshot's launch materials aim it squarely at the current frontier leader: Claude Fable 5, the Mythos-class model Anthropic shipped earlier this month.

What makes this launch worth your attention is that it breaks the usual pattern. This is not another "open model closes in on last year's frontier" story. On launch-week numbers, K3 beats Fable 5 outright in several categories while charging roughly a third of the price, and its wins cluster in exactly the workloads teams care most about right now: agents, terminal work, and frontend coding. Fable 5 still holds the overall crown, and the gap is real, but for the first time an open-weight model is winning individual rounds against the reigning flagship in the same week it launched.

This post walks through the scorecard, the pricing math, the open-weights reality check, and a practical recommendation by workload. If you have been following our Claude Sonnet 5 and GPT-5.6 coverage, the same house rule applies here: launch-week numbers tell you what to test, not what to deploy.

What Kimi K3 Actually Is

The headline number of 2.8 trillion parameters needs immediate context, because it does not mean what it would have meant in a dense model, where every parameter participates in every token. K3 is a sparse mixture-of-experts (MoE) system. Its parameters are organized into 896 expert subnetworks, and a router activates only 16 of them for each token. Total capacity is enormous, but the compute per token corresponds to a far smaller active budget. That is the trick that lets a model this size sell for $15 per million output tokens instead of flagship prices.

• Architecture: sparse MoE with 896 experts, 16 active per token, roughly 2.8T total parameters. It is the first open model announced in the 3-trillion-parameter class.
• Context window: 1,048,576 tokens, with flat per-token pricing across the whole window rather than a long-context surcharge tier.
• Modalities: native vision alongside text, with always-on reasoning instead of a separate thinking toggle.
• Availability: API and consumer app went live at launch. Full open weights are promised by July 27, 2026.

Claude Fable 5, for its part, is Anthropic's new Mythos-class flagship, a tier positioned above Opus. Anthropic does not publish parameter counts, so the comparison that matters here is the commercial one: capability per dollar on the tasks you actually run.

The Scorecard: Where Each Model Wins

Different evaluation sites slice this matchup differently, so instead of one vendor's tally, here is the picture that holds up across Moonshot's own launch deck, BenchLM's independent suite, and LMArena's community rankings. The short version: Fable 5 leads the aggregates, K3 wins the execution-heavy rounds.

Read as a whole, the pattern is coherent: Fable 5 is the deeper model and K3 is the stronger operator. When the task is to know things and reason carefully, Fable 5 leads. When the task is to drive a terminal, automate a workflow, or ship a frontend, K3 wins rounds, sometimes convincingly.

One caveat deserves its own paragraph. The Decoder reports that K3's measured hallucination rate, the share of answers where a model confidently states things that are not true, rose to 51%, up from 39% for its predecessor, even as accuracy improved. If your workload punishes confident wrong answers, weigh that number as heavily as any benchmark win.

The usual launch-week discipline applies with extra force, because several of these figures originate from Moonshot's own materials and vendors choose which comparisons to publish. Independent replications routinely tighten or flip orderings within weeks. Our weekly-refreshed AI model leaderboard tracks the cross-vendor picture as third-party evaluations land, and our benchmarks explainer covers what these suites actually measure.

Pricing: A Third of the Cost, Flat Across 1M Tokens

That is a 3.3x gap on output tokens, and output is where agentic bills live. Agent loops generate reasoning traces, tool calls, retries, and diffs, so output volume dominates spend. The cache-hit input price of $0.30 per million matters for the same reason: agent frameworks resend large, mostly identical contexts on every step, and a 10x cache discount changes the economics of long sessions. The absence of a long-context surcharge is quietly significant too, since some competitors' rate cards step up sharply past a context threshold.

Before you promise anyone a 70% budget cut, two things should temper the math. First, list price is not effective price. Different tokenizers count the same text differently, and reasoning models differ in how many tokens they emit to finish the same task. Artificial Analysis measured K3 at about $0.94 per task on its test suite, using 21% fewer output tokens than K2.6, which is genuinely efficient, but your workload will have its own profile. Second, K3 is dramatically more expensive than its predecessor: K2.6 charged $0.95 per million input tokens and $4 for output, versus $3 and $15 now. The Decoder reads this as the end of the ultra-cheap Chinese frontier era, with prices rising as capability reaches the real frontier. Model both scenarios against your actual token volumes with our LLM cost calculator before deciding anything.

The Open-Weights Wildcard

The July 27 weights release is what separates this launch from every other frontier head-to-head this year. An open-weight model at genuine frontier capability changes the strategic calculus: self-hosting for data-sensitive workloads, fine-tuning, distillation into smaller models, and freedom from any single vendor's rate card.

But temper the excitement with arithmetic. Even aggressive quantization, compressing each stored weight down to around 2 bits, works out to roughly 700 GB of weights for a 2.8-trillion-parameter model, before runtime overhead, the KV cache (the working memory a model fills up as it generates), and serving buffers. Practical estimates land in the 650 GB to 1 TB range, which means a multi-GPU datacenter node rather than a workstation, and early self-hosted deployments will also miss Moonshot's serving-side optimizations. If you want to sanity-check what your own hardware can hold, our VRAM calculator will give you the honest answer quickly. For most teams, open weights will mean hosted by a provider you choose rather than running in your closet. That is still a meaningful upgrade in leverage.

Which Should You Use?

• If you run agentic coding or automation workloads: K3 is the test to run first. Its benchmark wins cluster exactly there, and the 3.3x output-price gap compounds across every loop iteration. Run your eval set before moving production traffic.
• If your workload is deep reasoning, research, or high-stakes correctness: Fable 5 keeps the crown. It leads the aggregate indexes and most knowledge suites, and K3's elevated hallucination rate is a real concern for exactly these tasks. Where one wrong answer costs more than a million tokens, the premium is easy to justify.
• If you process images or grounded documents at volume: the launch numbers favor K3 by the widest margin of any category, and at these prices the experiment is cheap to run yourself.
• If data control or vendor independence matters to you: K3 is about to become the most capable open-weight model available, full stop. Plan around hosted open-weight providers rather than self-hosting unless you genuinely have a terabyte-class inference node.
• If you are happy on Fable 5 (or Sonnet 5) today: nothing here forces a migration. Switching costs are real, several of these numbers are vendor-published, and independent evaluations are days to weeks away. Re-evaluate when the picture stabilizes. Our leaderboard will reflect it weekly.

FAQ

Is Kimi K3 better than Claude Fable 5?

It depends entirely on the workload. Fable 5 leads the aggregate intelligence indexes and most knowledge-and-reasoning benchmarks, so it remains the most capable model overall. K3 wins on agentic execution, frontend coding, and multimodal tasks at roughly a third of the price. "Better" is a per-workload question, and the honest answer is to run both on your own evals.

How much cheaper is Kimi K3 really?

On list price, about 3.3x on both sides: $3 versus $10 per million input tokens and $15 versus $50 on output, with a $0.30 cache-hit input rate that further favors agent-style usage. Effective cost depends on tokenizer differences and how many reasoning tokens each model emits for your tasks, so treat 3x as the ceiling of the saving rather than a guarantee.

Can I download Kimi K3 now?

Not yet at time of writing. The API and app launched on July 16, and Moonshot has committed to publishing full open weights by July 27, 2026. Even then, running a 2.8T-parameter model locally requires roughly 650 GB to 1 TB of memory under aggressive quantization, so expect hosted open-weight providers to be the practical route.

Does the 1M context window really cost the same throughout?

That is Moonshot's published position: flat per-token pricing across the full 1,048,576-token window, with no long-context surcharge tier. Note that some consumer and coding plans cap usable context lower, at 256K on entry tiers, so check the plan matrix if you need the full window.

Should I trust these benchmark numbers?

Partially. Several head-to-head figures come from Moonshot's launch materials, and vendors choose which comparisons to publish. The LMArena frontend ranking is community-driven, which makes it harder to cherry-pick, and BenchLM runs its own suite. Still, the rule from our GPT-5.6 and Sonnet 5 coverage stands: launch-week numbers tell you what to test with your own eval set, not what to deploy. Independent evaluations typically reshape the picture within weeks.

References

• Tom's Hardware: Moonshot releases 2.8-trillion-parameter Kimi K3, largest open-weight AI model ever
• Fortune: Moonshot's Kimi K3 pushes Chinese AI into Fable-level territory
• The Decoder: Kimi's open model K3 nears GPT-5.6 Sol and Fable 5 while signaling the end of super-cheap Chinese AI
• BenchLM: Claude Fable 5 vs Kimi K3, benchmarks, pricing, speed
• OpenRouter: Kimi K3 vs Claude Fable 5 model comparison
• Codersera: Kimi K3 benchmarks vs Fable 5, GPT-5.6 Sol and Opus 4.8
• ModemGuides: Can you run Kimi K3 locally? The 2.8T hardware reality

Key Takeaways

• Kimi K3 launched July 16, 2026: a 2.8T-parameter sparse MoE (896 experts, 16 active per token) with a 1M-token context, native vision, always-on reasoning, and open weights promised by July 27. It is the largest open-weight model ever announced.
• Fable 5 leads the aggregates (Artificial Analysis Intelligence Index of about 60 versus 57, BenchLM 83.7 versus 81.0), while K3 wins the execution-heavy rounds: BenchLM's agentic suite, a number 1 debut on LMArena's Frontend Code Arena, and the multimodal category by a wide margin.
• Fable 5 is the stronger thinker and K3 is the stronger operator. K3's hallucination rate rose to 51% from 39%, which matters if your workload punishes confident wrong answers.
• K3 costs roughly a third of Fable 5 ($3/$15 versus $10/$50 per million tokens), with cache-hit input at $0.30 and no long-context surcharge. Those economics specifically favor agent workloads, though K3 is also a large price hike over K2.6.
• The open-weights release changes vendor leverage more than self-hosting reality: a 2.8T model needs roughly 650 GB to 1 TB of memory even heavily quantized. Most launch numbers are vendor-published, so test on your own evals and watch the leaderboard as independent results land.

Quiz