Prices verified 2026-09-02; Kimi K3 figures as of September 2026
Fable 5.1 pricing from Anthropic's September 2 release. Kimi K3 rates are Moonshot list prices as of September 2026.
Browse all Chinese model comparisonsFable 5.1 vs Kimi K3 · September 2026
Claude Fable 5.1 vs Kimi K3
Anthropic's September 2026 flagship against Moonshot's open-weights challenger — agentic coding, Chinese workloads, price, and self-hosting reality.
TL;DR
Fable 5.1 leads on frontier agentic coding (87.7% SWE-bench Verified) and hard reasoning. Kimi K3 wins on Chinese-language tasks (97.2%), open weights, and ~3x cheaper API pricing.
Specs at a glance
Fable 5.1 prices are Anthropic list rates; Kimi K3 rates are Moonshot figures as of September 2026.
| Spec | Claude Fable 5.1 | Kimi K3 |
|---|---|---|
| Model ID | claude-fable-5-1-20260902 | Kimi-K3 |
| Input price (per M tokens) | $10.00 | $3.00 |
| Output price (per M tokens) | $50.00 | $15.00 |
| Prompt cache read (per M tokens) | $1.00 | $0.75 |
| Context window | 1M | 1M |
| Max output | 128K | 64K |
| License | Closed API (proprietary) | Open weights — Modified MIT |
| Data retention | 30-day mandatory | None (self-hostable) |
Fable 5.1 carries mandatory 30-day retention. Kimi K3 is open-weights (Modified MIT, terms not fully confirmed) and can be self-hosted with a real inference fleet.
Benchmark comparison
Fable 5.1 leads on coding and reasoning benchmarks; Kimi K3 leads on Chinese evaluation suites.
| Spec | Claude Fable 5.1 | Kimi K3 |
|---|---|---|
| SWE-bench Verified | 87.7% | 71.2% |
| FrontierMath | 38.2% | 26.8% |
| Humanity's Last Exam (HLE) | 61.4% | 48.5% |
| Chinese evaluation suite | 91.5% | 97.2% |
| Needle-in-haystack (1M context) | 99.8% (1M) | 98.6% (1M) |
Fable 5.1 from Anthropic September 2, 2026 release. Kimi K3 SWE-bench 71.2% and Chinese eval 97.2% vendor-reported by Moonshot as of September 2026.
How to decide
Route by task language, deployment constraints, and agent difficulty — not brand loyalty.
You need frontier agentic coding
Fable 5.1's 87.7% SWE-bench and 38.2% FrontierMath are a tier above Kimi K3's 71.2% and 26.8%. Use Fable 5.1 for autonomous multi-file agents.
Your workload is Chinese-first
Kimi K3 scores 97.2% on Chinese eval vs Fable 5.1's 91.5%. Moonshot's China-hosted API and open weights suit mainland content, legal, and compliance workflows.
You want lower API cost with strong coding
At $3/$15 vs $10/$50, Kimi K3 is roughly 3x cheaper on input and ~3x on output. Good for high-volume tasks that don't need frontier SWE-bench scores.
You need self-hosting or air-gap
Kimi K3 open weights (~1.4 TB, MXFP4) can run on a cloud inference fleet — not a laptop project. Fable 5.1 is closed API only.
Retention or data sovereignty matters
Fable 5.1's 30-day mandatory retention blocks zero-retention teams. Kimi K3 self-hosted gives data control; China-hosted API suits data-residency requirements.
The ~3x API price gap narrows when hard agent tasks fail and retry on the cheaper model.
Estimate the real cost
Run both models through your own token mix before you commit to a production route.