Open weights versus closed weights usually turns into a pride contest. The numbers here don't leave much room for that. Kimi K3 scores 60.4% on SWE-bench Verified, against 95% for Claude Fable 5. The gap is real and large, and so is the price ratio: Fable 5 costs 3.3 times more per token, both on input and output.
That doesn't close the question. It opens the right one: in which slice of your work do those 34.6 points of difference actually show up, and in which slice are you paying 3.3x for headroom you never use?
How do Kimi K3 and Claude Fable 5 compare on benchmarks?
Three independent measures, three different sources, same direction: Fable 5 leads on all of them. The question that matters is the size of the gap on each one.
| Benchmark | Kimi K3 | Claude Fable 5 | Source |
|---|---|---|---|
| SWE-bench Verified | 60.4% | 95.0% | Moonshot AI / Anthropic |
| Artificial Analysis Intelligence Index | 44 | 53 (5.1) | artificialanalysis.ai |
| WebDev Arena Elo (frontend) | 1691 (5th place) | 1788 (2nd place) | arena.ai |
Checked on 09/22/2026. SWE-bench Verified is where the gap hurts most: almost 35 percentage points is the difference between a model that resolves most issues in a real repository and one that resolves the overwhelming majority of them. On WebDev Arena, blind human voting comparing generated web apps, the gap shrinks to 97 Elo points, which is noticeable but not overwhelming.
How much does each one cost on the official API?
| Model | Input (per million tokens) | Output (per million tokens) |
|---|---|---|
| Claude Fable 5 | US$ 10.00 | US$ 50.00 |
| Kimi K3 | US$ 3.00 | US$ 15.00 |
Source: official launch pricing, Anthropic (anthropic.com) and Moonshot AI, checked on 09/22/2026. The ratio is exactly 3.3 times on both sides, input and output, which is unusual: output usually weighs more in the final bill because that's where reasoning burns tokens. Here the proportion stays even, so the cost multiplier doesn't shift depending on how much the model "thinks" before answering.
What's inside each model?
| Architecture | Kimi K3: MoE, 2.8T total parameters, 104B active per token |
| Weights | Kimi K3: open. Claude Fable 5: closed, architecture undisclosed |
| Context window | Kimi K3: 1M tokens. Claude Fable 5: 1M input / 128K output |
| Release | Kimi K3: July 16, 2026 |
K3's MoE activates 104B of its 2.8T parameters per token, a ratio of roughly 27 to 1. That's what lets a model this size get served at a price that still competes with a closed model, even though it's nearly four times larger in total parameters than GLM-5.2, another open-weight model from the same crop.
Where does Kimi K3 still lose, and by how much?
On hard tasks, it loses badly. The nearly 35 points of SWE-bench Verified don't distribute evenly between simple and complex issues. They concentrate at the top of the curve: refactors that touch many files, bugs whose root cause isn't obvious, business context the model has to infer instead of read. That's where paying 3.3x more for Fable 5 pays for itself.
An open model losing to a closed model by 35 points on a hard benchmark isn't bad news for open weights. It's the same distance that existed two years ago between any model and the state of the art. The question that decides isn't "which one is better," it's "how much of your ticket queue lives at that hard top."
How do you actually decide?
- Look at your queue's real distribution, not last month's worst ticket. If most of the work is maintenance, small features, and common bugs, the 35-point gap doesn't show up in practice.
- Separate frontend from algorithmic backend. On WebDev Arena the gap is 97 Elo, proportionally much smaller than on SWE-bench Verified. If most of the work is interface, the gap hurts less.
- Stop treating it as a single choice. Routing is the most common answer in practice: Kimi K3 for volume, Fable 5 reserved for the issue that already failed twice with the cheaper model.
The math the benchmark table doesn't show
Every number above assumes you can afford to let the model finish the task. In a long agentic session, with the 1M context filling up and several iterations until the test passes, the cost per output token is what decides whether the fifth attempt happens or someone cuts the session short.
The Verboo Code already runs Kimi K3 on dedicated GPU, with unlimited tokens, at an API price lower than Moonshot's official rate: US$ 2.50 input and US$ 9.00 output per million tokens, against US$ 3.00 and US$ 15.00 direct from the lab. The 35-point gap on the hard benchmark still exists. What stops existing is the token bill deciding whether you make it to the end of the easy task.
