Kimi K3 vs Claude Fable 5
Kimi K3 is the largest open-weight model ever at 2.8 trillion parameters. Claude Fable 5 is Anthropic's Mythos-class flagship — the same weights as the restricted Mythos 5, with production safeguards. Across 35 shared evaluations in Moonshot's launch table, Fable 5 wins 22, K3 wins 12, and they tie once. But K3 costs 70% less per token, leads on long-horizon coding, and took #1 on the Frontend Code Arena. This is the most interesting model matchup of 2026.
Model cards side by side
| Specification | Kimi K3 | Claude Fable 5 |
|---|---|---|
| Provider / release | Moonshot AI · July 16, 2026 | Anthropic · June 9, 2026 |
| API model ID | kimi-k3 | claude-fable-5 |
| Architecture | MoE · 2.8T total / 16 of 896 experts active | Proprietary (undisclosed); same weights as Mythos 5 |
| Attention mechanism | Kimi Delta Attention + Attention Residuals | Not publicly detailed |
| Context window | 1,048,576 tokens (1M) | 1M tokens |
| Maximum output | 131K default, up to 1,048,576 | 128K tokens |
| Reasoning modes | Max effort only at launch; lower modes planned | Adaptive thinking, always on |
| Input modalities | Text, image, video | Text, image |
| Tools | Function calling, structured outputs, required tool choice | Tool calls, JSON output, MCP, computer use |
| Weights / license | Open weights (Modified MIT), promised by July 27 | Closed, proprietary, API-only |
| Input / output price | $3 / $15 per 1M | $10 / $50 per 1M |
| Cached input | $0.30 per 1M, automatic | $1 per 1M after cache write |
| Safeguards | Standard content filtering | Fallback to Opus 4.8 on cyber/bio/chem queries |
Fable 5 is the public, safeguarded version of Anthropic's Mythos 5 — same underlying model, but with classifiers that fall back to Opus 4.8 on high-risk queries. K3 is Moonshot's first 3T-class model, with a published architecture (Kimi Delta Attention, Stable LatentMoE, Attention Residuals) and weights promised by July 27. Both support 1M-token context windows. Fable 5 has mature cloud deployment across Anthropic, Bedrock, Google Cloud, and Microsoft Foundry; K3 is API-only at launch with partner inference still ramping.
The 35-benchmark scorecard: Fable wins breadth, K3 wins key coding rows
Moonshot's official K3 launch table compares both models at maximum reasoning effort across 35 shared evaluations. The top-line count: Fable 5 wins 22, K3 wins 12, one tie. But the distribution matters more than the count.
Coding benchmarks (8 rows)
| Benchmark | Kimi K3 | Fable 5 | Leader |
|---|---|---|---|
| DeepSWE | 67.5 | 70.0 | Fable +2.5 |
| FrontierSWE | 81.2 | 86.6 | Fable +5.4 |
| Kimi Code Bench 2.0 (internal) | 72.9 | 76.9 | Fable +4.0 |
| Terminal-Bench 2.1 | 88.3 | 84.6 | K3 +3.7 |
| Program Bench | 77.8 | 76.8 | K3 +1.0 |
| SWE Marathon | 42.0 | 35.0 | K3 +7.0 |
| PostTrain Bench | 36.6 | 41.4 | Fable +4.8 |
| MLS Bench Lite | 48.3 | 49.9 | Fable +1.6 |
Coding splits 3-5 in Fable's favor, but K3's three wins are consequential: Terminal-Bench 2.1 (88.3 vs 84.6), SWE Marathon (42.0 vs 35.0 — a 7-point gap on sustained autonomous work), and Program Bench (77.8 vs 76.8). Fable's wins include the widest gap on any chart: FrontierSWE at 86.6 vs 81.2. Notably, Fable also wins Moonshot's own internal Kimi Code Bench 2.0 — a credibility-building detail that Moonshot published anyway.
Agentic benchmarks (12 rows)
| Benchmark | Kimi K3 | Fable 5 | Leader |
|---|---|---|---|
| GDPval-AA v2 (Elo) | 1,668 | 1,760 | Fable +92 |
| AA-Briefcase (Elo) | 1,548 | 1,583 | Fable +35 |
| BrowseComp | 91.2 | 88.0 | K3 +3.2 |
| DeepSearchQA (F1) | 95.0 | 94.2 | K3 +0.8 |
| Toolathlon-Verified | 73.2 | 77.9 | Fable +4.7 |
| MCP Atlas | 84.2 | 84.7 | Fable +0.5 |
| AutomationBench | 30.8 | 29.1 | K3 +1.7 |
| Job Bench | 52.9 | 57.4 | Fable +4.5 |
| APEX-Agents | 37.6 | 43.3 | Fable +5.7 |
| OfficeQA Pro | 63.3 | 69.9 | Fable +6.6 |
| SpreadsheetBench 2 | 34.8 | 34.7 | K3 +0.1 |
| DECK-Bench (internal) | 73.5 | 73.0 | K3 +0.5 |
Fable dominates the economically-weighted agent evals: GDPval-AA (+92 Elo), JobBench (+4.5), APEX-Agents (+5.7), OfficeQA Pro (+6.6). K3 answers with BrowseComp (+3.2) — a deep web research benchmark where an open model leading is genuinely new — plus narrow wins on AutomationBench, SpreadsheetBench 2, and DeepSearchQA.
Reasoning, knowledge & vision
| Benchmark | Kimi K3 | Fable 5 | Leader |
|---|---|---|---|
| GPQA-Diamond | 93.5 | 92.6 | K3 +0.9 |
| HLE-Full (no tools) | 43.5 | 53.3 | Fable +9.8 |
| HLE-Full (with tools) | 56.0 | 63.0 | Fable +7.0 |
| MMMU-Pro | 81.6 | 81.2 | K3 +0.4 |
| CharXiv (RQ) with Python | 91.3 | 93.5 | Fable +2.2 |
| ZeroBench with Python (pass@5) | 41.0 | 46.0 | Fable +5.0 |
| OmniDocBench | 91.1 | 89.8 | K3 +1.3 |
| PerceptionBench | 58.5 | 57.2 | K3 +1.3 |
Fable's largest leads are on HLE-Full (+9.8 without tools, +7.0 with tools) — the hardest expert-level benchmark. Vision is Fable's clearest category: it wins 8 of 12 vision rows. K3's vision wins are narrow (OmniDocBench +1.3, PerceptionBench +1.3, MMMU-Pro +0.4), but it has one product-level advantage the table doesn't capture: native video input through the first-party API.
Radar: six key shared benchmarks
● Kimi K3 ● Fable 5
The radar uses six benchmarks where both models have scores in Moonshot's table. Scores normalized: Term-Bench, FrontierSWE, BrowseComp, GPQA to 100; SWE Marathon to 50; GDPval-AA to 2000. It is a visual aid, not a new benchmark.
- K3 leads on Terminal-Bench 2.1, SWE Marathon, BrowseComp, and GPQA-Diamond.
- Fable 5 leads on FrontierSWE and GDPval-AA.
- The radar understates Fable's advantage because it can't show HLE, vision, and the other 29 rows where Fable leads.
- But it correctly shows K3's strength on the benchmarks most relevant to coding agents.
The Frontend Code Arena: K3 takes #1
The number driving the most discourse isn't on Moonshot's charts. Arena's community-voted Frontend Code Arena now has Kimi K3 at #1 with 1,679 points, ahead of Claude Fable 5. K3 ranked first in 6 of 7 frontend domains including Brand and Marketing, Reference-Based Design, and Data and Analytics. This is a 17-place jump from Kimi K2.6's #18 position. Unlike vendor-run benchmarks, this is thousands of blind human votes on real frontend tasks — which is why it traveled so fast.
Pricing: K3 costs 70% less at every tier
| Per 1M tokens | Kimi K3 | Claude Fable 5 |
|---|---|---|
| Fresh input / cache miss | $3.00 | $10.00 |
| Cached input / cache hit | $0.30 | $1.00 |
| Output | $15.00 | $50.00 |
| 10M fresh input + 2M output | $60.00 | $200.00 |
| 10M cached input + 2M output | $33.00 | $110.00 |
K3 costs 30% of Fable 5 at every token tier. Said another way: Fable costs 3.33× more. The gap is large enough to survive modest efficiency differences. But token price isn't cost per completed task — Fable may finish hard jobs in fewer attempts, and its 22 benchmark wins suggest the premium can buy higher success probability. K3's always-on max reasoning also means reasoning tokens are billed as output at $15/M. Measure total tokens, retries, and completion rate on your own workload.
Open weights vs closed maturity
K3: open weights by July 27
Moonshot has committed to releasing full weights under a Modified MIT license. At 2.8T parameters with 16 active experts, self-hosting requires serious infrastructure — Moonshot recommends at least 64 accelerators. Most teams will still use an inference provider.
The MIT license enables fine-tuning, air-gapped deployment, auditability, and sovereignty that closed models cannot match. Once the weights land, K3 stops being a cheaper API and becomes infrastructure.
Fable 5: closed, but Mythos-class
Fable 5 is the same underlying model as the restricted Mythos 5, with production safeguards added. It's available through Anthropic's API, Bedrock, Google Cloud, and Microsoft Foundry — mature deployment options K3 doesn't yet match.
One caveat: Fable 5's benchmark scores include fallback behavior. On cyber/bio/chem queries, it falls back to Opus 4.8. Moonshot labels Fable "with fallback" in its table. The Fable column measures the production system, not always the full model.
Which model should you use?
Choose Kimi K3 if…
- You need the #1 frontend coding model — K3 leads the Frontend Code Arena.
- Your workload is terminal-heavy, long-running, or browse-intensive — K3 wins Terminal-Bench 2.1, SWE Marathon, and BrowseComp.
- API cost is a first-order constraint — K3 is 70% cheaper at every token tier.
- You need video input, air-gapped deployment, or open weights for fine-tuning.
- You want to run high-volume, measurable coding tasks with clear pass/fail criteria.
Choose Claude Fable 5 if…
- You need maximum broad capability — Fable wins 22 of 35 shared evaluations.
- Your work involves hard repo surgery (FrontierSWE), expert knowledge (HLE), or complex vision tasks.
- You need mature cloud deployment across AWS, GCP, and Azure.
- You rely on Anthropic's tooling: Claude Code, MCP, computer use, dynamic workflows.
- Judgment, recovery from errors, and production reliability matter more than token cost.
Compare both models on real code
Run the same repository prompt, tool schema, and acceptance tests rather than trusting a leaderboard headline.
Try both in CodingFleet →Final verdict
Claude Fable 5 is the stronger broad-capability model. It wins 22 of 35 shared evaluations, leads on the hardest software engineering benchmarks (FrontierSWE +5.4, DeepSWE +2.5), dominates expert knowledge (HLE +9.8), and sweeps most vision tasks. It's the safer choice when the task is unpredictable and the cost of failure is high.
Kimi K3 is the more disruptive product. It wins the benchmarks most relevant to coding agents (Terminal-Bench 2.1, SWE Marathon, BrowseComp), took #1 on the Frontend Code Arena, costs 70% less, and will soon be available as open weights. For high-volume, measurable coding work, it's extraordinarily compelling.
The honest assessment: Fable 5 is the better model. K3 is the better deal. The gap between them is small enough that task-level routing beats picking one winner. Test both on your actual workload — the model that finishes your work with the least total friction is the right one.
Sources and methodology
- Moonshot AI — Kimi K3 Tech Blog. Official launch table with 35-row comparison, architecture details, and benchmark scores.
- LLM Stats — Kimi K3 vs Claude Fable 5: Complete Analysis. Detailed 35-benchmark breakdown with win counts and harness notes.
- LLM Stats — Claude Fable 5 Review. Official Anthropic benchmark scores, pricing, and fallback mechanism details.
- Valletta Software — Kimi K3 vs Claude Fable 5 Scorecard. 14-benchmark head-to-head with coding and agent chart analysis.
- Codersera — Kimi K3 Benchmarks vs Fable 5, GPT-5.6 Sol & Opus 4.8. Artificial Analysis Intelligence Index scores and independent analysis.
- BenchLM — Claude Fable 5 vs Kimi K3. Independent benchmark comparison with shared result analysis.
- AIHubMix — Kimi K3 pricing and eesel — K3 pricing analysis.
- All benchmark rows retain their source labels and harness details. Provider-reported results can differ because of harness, prompt, context, and reasoning-budget choices. The radar and bar charts are visual aids, not statistically normalized leaderboards. Win counts use Moonshot's launch table with Fable 5 in production configuration (fallback enabled).