Kimi K3 is competitive with Fable; Kimi K3 + Fable is SoTA
- source_url: https://fireworks.ai/blog/kimik3-fable
- source_type: article
- platform: blog
- author: Fireworks AI
- original_date: 2026-07-21
- added_date: 2026-07-22
- category: models
- tags: model-routing, kimi-k3, fable, agentic-benchmarks, cost-performance, open-models
- quality_score: 4
摘要（中文）
Fireworks 在约 1030 个真实 agent 任务（SWE终端运维算法多语言法律等）上对比开源 Kimi K3 与闭源 Fable 5：二者路由约 93% 准确率，长 agent 环上最高可比单用 Fable 便宜约 50；主张按工作类型路由而非绑定单一前沿模型收录理由：用统一 harness 给出可核对的开闭源互补与成本曲线，对生产 agent 选型与路由策略很有参考
Summary (English)
Fireworks evaluated open Kimi K3 against closed Fable 5 on ~1,030 real agentic tasks (SWE-style, terminal ops, algorithmic, multi-language, legal). Routing between K3 and Fable reached about 93% accuracy and was up to ~50 more cost-effective than Fable alone on long agentic loops; oracle routing defines the cost/performance ceiling. The post argues single-model commitment is wasteful and that practical systems should route by work type rather than pick one frontier model.
One-liner
Fireworks 在约 1030 个真实 agent 任务（SWE终端运维算法多语言法律等）上对比开源 Kimi K3 与闭源 Fable 5：二者路由约 93% 准确率，长 agent 环上最高可比单用 Fable 便宜约 50.
原文 / 抓取正文
Kimi K3 is competitive with Fable; Kimi K3 + Fable is SoTA
作者: Fireworks AI
原文链接: https://fireworks.ai/blog/kimik3-fable
发布时间: 2026-07-21
tl;dr
We ran Kimi K3 (open) against Fable 5 (closed) on ~1,000 agentic tasks finding:
1. We achieved 93% accuracy with routing between K3 and Fable. 2. Results were up to ~50X more cost effective than Fable alone on long agentic loops, and consistently lower cost across every use case.
K3 is a frontier quality open model at a fraction of the cost. Even bigger is that it complements Fable predictably, which makes it possible to get the highest quality intelligence by routing tasks.
How We Measured
We averaged benchmarks, each aimed at a different kind of work, and ran K3 and Fable 5 through the same harness. About 1,030 tasks in all, in real agent loops.
| Family | What it tests | Tasks | | --- | --- | --- | | SWE | Real repo bug-fixes (SWE-bench style) | 460 | | Terminal | Long agentic ops: security, crypto, reverse-eng, sysadmin | 89 | | Algorithmic | LeetCode / AtCoder-style problems | 100 | | Multi-Language | Implementation across six languages | 225 | | Legal | A legal-agent benchmark (lawyer-graded tasks) | 120 |
Oracle routing is a method for measuring the best theoretical performance by running the task through each model and then picking the cheapest correct option (the cost/performance ceiling). In a practical router, you don’t get to run your task against multiple models. The router makes a prediction of which model has the best cost and quality trade off.
In this study, oracle routing demonstrated K3 is selected for 72-96% of tasks. This suggests a near-perfect router might be achievable, by learning the difference between day-to-day tasks and the true long tail of frontier work.
K3 is a good model
From a high level, the head-to-head is nearly a tie. On SWE, K3 gets 92.4%, Fable 92.6%. Across the five types of tasks, the two models tend to stay within a few points of each other, with Fable pulling slightly ahead on multi-language coding breadth.
They have discretely better performance across different task types. On SWE split by domain, K3 is sharpest on symbolic math and dev tooling; Fable wins on web & data visualization. On multi-language, Fable's breadth carries Java, Python and C++, while K3 draws even on JavaScript and Rust.
For long-horizon terminal work, K3 cleared a batch of tasks Fable never cracked: a 7z hash, FEAL cryptanalysis, leaked secrets, a live vulnerability, runaway async jobs. Of 89 terminal tasks, K3 has 11 solo wins to Fable's 7, and takes the security and crypto cluster outright.
Cost: K3 can be up to 50x lower cost on Fireworks
While quality is a near-tie at a high level, price isn't close. Two drivers: token pricing, and the fact that which model runs longer depends on the task.
On SWE, K3 works much harder than Fable: roughly 55 turns and 1.3M tokens a task versus 21 turns and 130K. On long terminal tasks it's the other way around: Fable spirals, running up 64 turns and 1.5M tokens (sometimes into a timeout).
Prompt caching does most of the work of turning that effort into K3's price advantage: even when K3 reads ten times the tokens, with cache hits SWE runs still come in lower cost than Fable. Extra turns generally mean more wall-clock time; if you need an answer in two seconds that matters; if you're running agents in the background at scale, a fraction-sized bill matters more.
Don't pick a model. Route.
If you send every task to whoever handles it best, you don't land somewhere between the two models, you land above both. Per-task routing always outperforms any single model run. The oracle router chooses K3 for 72-96% of task traffic. Overall quality ends above either model alone at a cost close to just using the cost-optimized one.
Single Models Are Wasteful and No Longer SoTA
Kimi K3 + Fable routed together unlocks their best qualities at the best price. The single model provider, token-maxxing days are coming to an end. The task-level data says these models are specialists at very different prices. The best AI no longer comes out of a single lab — it's a mixture of models.
What this means in practice: 1. Open as the default. A much lower-cost open model like K3 should be your base case, since the oracle sends it most of the traffic anyway. 2. The router is your moat. A router must be tailored to your workload and learning that task/model split continuously is the best chance you’ll have at staying ahead.
Obsidian intake evidence excerpt
该内容文件由 AAIF content-fetcher 根据 active/high-score entry 与 OpenCLI web 抓取正文补齐。
- entry_id: 162a22c4
- title: Kimi K3 is competitive with Fable; Kimi K3 + Fable is SoTA
- source: https://fireworks.ai/blog/kimik3-fable
- existing_summary_zh: Fireworks 在约 1030 个真实 agent 任务（SWE终端运维算法多语言法律等）上对比开源 Kimi K3 与闭源 Fable 5：二者路由约 93% 准确率，长 agent 环上最高可比单用 Fable 便宜约 50；主张按工作类型路由而非绑定单一前沿模型收录理由：用统一 harness 给出可核对的开闭源互补与成本曲线，对生产 agent 选型与路由策略很有参考