Kimi K3 vs Claude Fable 5.1: Pricing, Benchmarks & When to Switch
September 25, 2026 · kimi k3 vs claude fable 5.1 · claude fable 5.1 · fable 5.1 pricing · kimi k3 vs claude · kimi k3 pricing · claude opus 5.5 · ai model comparison · llm api cost
Claude Fable 5.1 is the stronger model, and Kimi K3 is the cheaper one, by roughly 2x to 3.3x depending on how you measure. Anthropic’s September 1 update kept Fable’s $10/$50 per-million-token price, cut cache reads to $0.25, and opened a clear lead over K3 on independent tests (51 vs 44 on the Artificial Analysis Intelligence Index). K3 still costs $3/$15, still wins on long-context reasoning and some agentic workflow tests, and is still the only one of the two with downloadable weights.
This piece is current as of September 25, 2026 and extends our July Kimi K3 vs Claude Fable 5 comparison rather than repeating it: what changed in 5.1, the full rate card with four worked cost examples, which benchmarks can be compared, the latency gap, and the twist Anthropic added three weeks later, Claude Opus 5.5.
What changed from Fable 5 to Fable 5.1
Anthropic announced Fable 5.1 and Mythos 5.1 on September 1, 2026. The model overview lists it as claude-fable-5-1, available on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS, with retirement “not sooner than September 1, 2027.”
| Claude Fable 5 | Claude Fable 5.1 | |
|---|---|---|
| Released | June 9, 2026 | September 1, 2026 |
| Input / output (per 1M) | $10 / $50 | $10 / $50 (unchanged) |
| Cache read (per 1M) | $1.00 | $0.25 (75% lower) |
| Cache write, 5 min / 1 hr | $12.50 / $20 | $12.50 / $20 (unchanged) |
| Context / max output | 1M / 128K | 1M / 128K |
| Knowledge cutoff | Jan 2026 | Jun 2026 |
| Thinking | Adaptive, always on | Adaptive, always on (default effort high) |
| Forced tool use | Supported | Returns a 400 error |
| Cyber safeguard interventions | Baseline | ~60% fewer per Claude Code session |
Three details matter for anyone weighing it against K3:
- The price cut is cache-only. Anthropic estimates Fable 5.1 costs “around 25%” less than Fable 5 on typical workloads and “up to around 45%” less on context-heavy agentic work, based on four weeks of August usage. Fresh input and output prices did not move.
- The safeguards loosened, but did not disappear. In July, Moonshot’s footnote showed Fable 5 falling back to Opus 4.8 on 35% of SWE-Marathon tasks. Fable 5.1 may now find software vulnerabilities, but penetration testing, exploit generation, binary vulnerability scanning and life-sciences R&D queries are still redirected to Opus models.
- Mythos 5.1 is the same model. The what’s-new page states both share specs and pricing. Mythos 5.1 has looser safeguards and is invitation-only through Project Glasswing; for now it is limited to a set of US organizations.
Kimi K3 vs Claude Fable 5.1 at a glance
| Kimi K3 | Claude Fable 5.1 | |
|---|---|---|
| Maker | Moonshot AI | Anthropic |
| Released | July 16, 2026 | September 1, 2026 |
| Architecture | 2.8T MoE, 104B active | Undisclosed |
| Context window | 1,048,576 tokens | 1M tokens |
| Max output | Not separately capped in docs | 128K (sync API) |
| Input modalities | Text, image | Text, image |
| Reasoning control | reasoning_effort low / high / max (max default), always thinking | effort levels, adaptive thinking always on, per-message effort (beta) |
| Weights | Downloadable, Kimi K3 License | Closed |
| Where you can run it | Moonshot API, Amazon Bedrock, 14+ third-party hosts, self-host | Claude API, Bedrock, Google Cloud, Foundry |
| Data retention | Varies by provider; zero-retention options exist | 30 days; no zero data retention unless Anthropic authorizes it |
The data-retention row is easy to miss: per Anthropic’s docs, Fable 5.1 carries 30-day retention with no zero-retention option unless Anthropic authorizes it. With K3, retention depends on the host, and some third-party K3 providers advertise zero retention.
Pricing: the full rate card
From Anthropic’s pricing page and Moonshot’s Kimi API pricing, with Claude Opus 5.5 included because it changes the decision.
| Per 1M tokens | Kimi K3 | Claude Fable 5.1 | Claude Opus 5.5 |
|---|---|---|---|
| Input (uncached) | $3.00 | $10.00 | $4.00 |
| Cache write, 5 min | $3.00 | $12.50 | $5.00 |
| Cache write, 1 hr | $6.00 | $20.00 | $8.00 |
| Cache read (hit) | $0.30 | $0.25 | $0.20 |
| Output | $15.00 | $50.00 | $20.00 |
| Batch API | Not published | 50% off ($5 / $25) | 50% off ($2 / $10) |
| US-only inference | Provider-dependent | 1.1x multiplier | 1.1x multiplier |
Two things stand out. First, Fable 5.1’s cache read is now cheaper than K3’s ($0.25 vs $0.30), which is the only line where Anthropic’s flagship undercuts Moonshot. Second, cache writes are where K3 quietly wins big: a 5-minute write on K3 costs the same $3 as normal input, while Fable 5.1 charges $12.50, a 25% premium over its own input rate. Kimi’s docs also note each cache hit refreshes the entry’s lifetime with no extra write charge.
One billing change landed this week. Per Anthropic’s docs, from September 24, 2026, Fable 5.1 refusals that arrive before any output are billed when they fall in categories with low false-positive volumes. A mid-stream refusal bills input plus the output already streamed. If you run security-adjacent agents, budget for this.
Worked cost examples
List-price math for four realistic workloads. These assume identical token counts on both models, which flatters Fable slightly (see the next section).
| Scenario | Kimi K3 | Claude Fable 5.1 | Claude Opus 5.5 | Claude Fable 5 |
|---|---|---|---|---|
| A. One long document: 400K input (uncached), 8K output | $1.32 | $4.40 | $1.76 | $4.40 |
| B. Agent session: 60 turns x 120K context, 95% cache hits, 3K output per turn | $5.83 | $15.21 | $6.77 | $20.34 |
| C. Offline batch: 10,000 docs x 5K in / 1K out | $300 | $500 (batch) | $200 (batch) | $500 (batch) |
| D. Repo Q&A: 900K cached context, 100 queries, 2K output each | $32.70 | $43.75 | $26.50 | $111.25 |
How the numbers work:
- A: No caching, so Fable 5.1’s discount does nothing.
- B: 7.2M input tokens total: 6.84M cache reads plus 0.36M fresh tokens written to the 5-minute cache, and 0.18M output. Fable 5.1 saves 25% versus Fable 5 here, in line with Anthropic’s estimate, yet still costs 2.6x K3 because output dominates.
- C: Moonshot publishes no batch discount, so K3 is at full price. Fable 5.1’s 50% batch rate narrows the gap to 1.7x, and Opus 5.5 at batch rates beats K3 outright.
- D: 90M cached reads, one 0.9M cache write, 0.2M output. This is Fable 5.1’s best case, and it still loses to K3 by about a third.
The break-even rule. Ignoring writes, Fable 5.1 is cheaper than K3 only when your workload reads more than 700 cached tokens for every output token (the $0.05 read saving has to cover a $35 output premium). Scenario B runs at about 38:1 and Scenario D at 450:1. Very few real workloads get past 700:1.
The measured gap is closer to 2x
List prices assume both models use the same number of tokens. They do not. In Artificial Analysis’ head-to-head, K3 (max) used about 48K output tokens per task, 32K of them reasoning, while Fable 5.1 (high) used about 38K, 19K of them reasoning. The result: $2.00 per task for K3 vs $3.91 for Fable 5.1, a 1.96x gap rather than 3.3x. At max effort, Fable 5.1 costs $7.63 per task on the same index.
Tokenizers differ too. Anthropic says its current tokenizer produces roughly 30% more tokens than pre-Opus 4.7 models for the same text, and K3 uses its own 160K-vocabulary tokenizer. Per-token prices are not perfectly apples to apples, so run your own prompts through both before you commit.
Benchmarks: what can and can’t be compared
This is where most comparisons go wrong. There is no official benchmark table that includes both Kimi K3 and Fable 5.1. Anthropic’s launch table compares Fable 5.1 with Fable 5, Opus 5 and GPT-5.6 Sol. Moonshot’s July model card compares K3 with Fable 5. Chaining them together looks tempting, but it does not work, as shown below.
Anthropic’s numbers for Fable 5.1 (vendor-reported)
| Benchmark | Fable 5.1 | Fable 5 | Opus 5 | GPT-5.6 Sol |
|---|---|---|---|---|
| Terminal-Bench-Science 0.1 | 52.6% | 24.7% | 29.0% | 22.4% |
| Terminal-Bench 4.0 | 55.8% | 42.0% | 52.3% | 37.3% |
| GDPval-AA v2 (Elo) | 1853 | 1723 | 1824 | 1711 |
| OSWorld 2.0 (strict) | 41.7% | 36.1% | 39.6% | n/a |
| Humanity’s Last Exam (no tools / tools) | 60.9% / 65.0% | 57.8% / 63.8% | 56.6% / 63.6% | n/a |
| AutomationBench | 31.4% | 17.1% | 26.9% | 19.6% |
| CursorBench 3.2.0 | 73.4% | 70.5% | 70.0% | 67.2% |
Mythos 5.1 scored 60.9% on Terminal-Bench 4.0; Anthropic attributes the gap to tasks where the older, less precise cyber safeguards intervened on Fable 5.1.
Why you can’t bridge through Fable 5
Both vendors measured Fable 5, so it looks like a common anchor. It isn’t. On AutomationBench, Moonshot reported Fable 5 at 29.1 (K3: 30.8), while Anthropic reports Fable 5 at 17.1. Anthropic’s footnote explains the difference: in its run, Fable 5 “scored a zero on AutomationBench” wherever safeguards intervened. On GDPval-AA v2, Moonshot cited 1747 for Fable 5 and Anthropic 1723. OSWorld 2.0 is worse: Anthropic ran the benchmark authors’ August 2026 task release and warns its numbers “aren’t directly comparable to previously published OSWorld 2.0 results.”
So you cannot infer a K3 vs Fable 5.1 gap from the two vendor tables. You need one evaluator running both models.
The independent head-to-head
Artificial Analysis ran both through its Intelligence Index v4.3.2 (Fable 5.1 at high effort with default fallback; K3 at max):
| Evaluation | Kimi K3 (max) | Claude Fable 5.1 (high) |
|---|---|---|
| Intelligence Index v4.3.2 | 44 | 51 |
| Terminal-Bench 4.0 | 13% | 52% |
| Humanity’s Last Exam | 47% | 56% |
| AA-Omniscience | 20 | 41 |
| GDPval-AA v2.1 (Elo) | 1524 | 1617 |
| AA-Briefcase v1.1 (Elo) | 1505 | 1592 |
| CritPt | 23% | 30% |
| GDP.pdf | 22% | 27% |
| SciCode | 59% | 59% |
| AutomationBench-AA | 58% | 55% |
| AA-LCR v1.1 (long context) | 89% | 84% |
The honest reading: of the 10 component evaluations, Fable 5.1 wins 7, K3 wins 2, and 1 is a tie. The Terminal-Bench 4.0 gap (13% vs 52%) is the most striking number here, and it is the opposite of K3’s strong July showing on Terminal-Bench 2.1 (88.3 in Moonshot’s own run). Harness and task-version differences matter a lot on terminal benchmarks, so treat this single result carefully, but it is the newest independent data point. K3’s long-context lead (89% vs 84%) is consistent with what we saw at launch. For context, AA ranks K3 third among open-weight models in its large size class.
On real coding work from July, Together AI’s DeepSWE study found K3 at 68.5% pass@1 versus 69.9% for Fable 5 (xhigh), at $4.65 versus $13.41 per rollout. That was Fable 5, not 5.1; nobody has published the same test on 5.1 yet.
Speed: the cost nobody puts in the price table
Artificial Analysis’ latency numbers change the picture for interactive use:
| Metric | Kimi K3 (max) | Claude Fable 5.1 (high) |
|---|---|---|
| Output speed | 35 tok/s | 55 tok/s |
| Time to first token | 3.81 s | 20.93 s |
| Time to first answer token | 60.58 s | 20.93 s |
| End-to-end, 500-token answer | 74.78 s | 30.03 s |
| Average time per index task | 1,266 s (~21 min) | 442 s (~7 min) |
K3’s fast first token is misleading: it reasons first, and the first answer token arrives after about a minute at max effort. Fable 5.1 finishes an average task nearly 3x faster, and if a developer is waiting, that time can cost more than the token savings. Dropping K3 to high or low effort, per Moonshot’s reasoning-effort guide, cuts latency at some quality cost.
The Opus 5.5 twist
Three weeks after Fable 5.1, Anthropic undercut its own flagship. Claude Opus 5.5 shipped on September 22, 2026 at $4/$20 with $0.20 cache reads. According to VentureBeat’s coverage, Anthropic reports Opus 5.5 beating Fable 5.1 on Terminal-Bench 4.0 (66.4% vs 55.8%), CursorBench 4.0 (57.8% vs 51.8%) and GDPval-AA v2.1 (1846 vs 1735). On the current Artificial Analysis leaderboard, Opus 5.5 at max effort (58) sits above Fable 5.1 at max (53).
Even before Opus 5.5 launched, Anthropic’s own docs said: “For most workloads, start with Claude Opus 5… Use Claude Fable 5.1 for demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5 at higher effort still fall short.”
What this means for K3 buyers: the real Anthropic alternative to K3 is now probably Opus 5.5, not Fable 5.1. It costs 33% more than K3 per token rather than 233% more, and in our worked examples it lands within about 15–35% of K3 on uncached and agentic work, and below K3 on batch and heavy-cache workloads. Our K3 vs Claude Opus 5 piece covers the previous Opus generation.
Switching costs in both directions
Moving to Fable 5.1 is not a one-line change. The docs list three breaking changes:
tool_choiceofanyor a named tool returns a 400 error. Useautowith strict tool use or structured outputs instead.- Earlier Claude models cannot read Fable 5.1 thinking blocks, so a router that falls back from 5.1 to an older model loses the reasoning for those turns.
- Editing anything before a thinking block (system prompt, tools, earlier messages) invalidates it. This is enforced for accounts created on or after August 31, 2026, and partly exists to block distillation.
Prefilling and non-default temperature, top_p or top_k values are still rejected, and Anthropic notes 5.1 may issue one tool call per turn where Fable 5 batched several.
Moving to K3 from Claude is easier than it used to be. Moonshot’s API is OpenAI-compatible, several hosts expose Anthropic-compatible endpoints, and Kimi Code is Moonshot’s own agent harness. The friction is elsewhere. K3 reached Amazon Bedrock on September 18, 2026, but it is not in Google Vertex’s managed catalog, so teams tied to Google Cloud still need a third-party host or self-hosting.
When to switch: a decision guide
| Your situation | Pick | Why |
|---|---|---|
| High-volume agent loops where cost per task matters most | Kimi K3 | ~2x cheaper per measured task; 2.6x in list-price agent math |
| Long-document and whole-repo analysis | Kimi K3 | Wins AA-LCR (89% vs 84%); cheapest uncached input |
| Terminal-heavy coding agents, today | Fable 5.1 or Opus 5.5 | 52% vs 13% on AA’s Terminal-Bench 4.0 run |
| Hardest reasoning, knowledge recall, HLE-style questions | Fable 5.1 | Leads HLE (56% vs 47%) and AA-Omniscience (41 vs 20) |
| Interactive use where a person waits on the answer | Fable 5.1 | ~7 min vs ~21 min per average task |
| Offline batch jobs | Opus 5.5 (batch) | $2/$10 batch beats K3’s undiscounted $3/$15 |
| Self-hosting, data residency, fine-tuning | Kimi K3 | Only one with open weights |
| Must stay inside Google Vertex | Anthropic | K3 is on Bedrock and Foundry (via Fireworks), not Vertex |
| Security research (pen-testing, exploit work) | Neither Fable tier by default | Fable 5.1 still routes this to Opus; Mythos 5.1 needs vetted access |
| Already on Fable 5 | Upgrade to 5.1 | Same price, 25–45% lower bill, better scores; fix the breaking changes first |
A sensible stack: K3 for bulk agent and long-context work, an Anthropic model (likely Opus 5.5) for interactive coding, and Fable 5.1 only where cheaper models fail your evals.
FAQ
Is Claude Fable 5.1 better than Kimi K3? On most independent measurements, yes. Artificial Analysis scores Fable 5.1 (high effort) at 51 on its Intelligence Index v4.3.2 versus 44 for Kimi K3 (max), and Fable 5.1 leads heavily on Terminal-Bench 4.0 (52% vs 13%) and knowledge recall. K3 wins on long-context reasoning (AA-LCR 89% vs 84%) and AutomationBench-AA (58% vs 55%), and ties on SciCode (59%).
How much does Claude Fable 5.1 cost compared with Kimi K3? Fable 5.1 lists at $10 per million input tokens, $50 per million output tokens and $0.25 per million cache reads. Kimi K3 lists at $3 input, $15 output and $0.30 cache hits. Fable 5.1 is 3.3x more expensive on fresh input and output, slightly cheaper on cache reads, and offers a 50% batch discount that Moonshot does not publish for K3.
What changed between Claude Fable 5 and Fable 5.1? Input and output prices stayed at $10/$50, but cache reads fell 75% to $0.25 per million, which Anthropic estimates cuts typical bills about 25% and agentic bills up to about 45%. Anthropic also reports large benchmark gains, about 60% fewer cyber-safeguard interventions per Claude Code session, and three breaking API changes, including the removal of forced tool use.
Is Claude Fable 5.1 the same model as Claude Mythos 5.1? Yes. Anthropic says Fable 5.1 and Mythos 5.1 are the same underlying model with different safeguards and identical pricing. Fable 5.1 is generally available; Mythos 5.1 is invitation-only through Project Glasswing and trusted-access programs for vetted cyberdefenders and life scientists, currently limited to US organizations.
Does the cheaper Fable 5.1 cache make it cheaper than Kimi K3 for agents? Almost never in practice. Fable 5.1 reads cached tokens for $0.25 per million versus $0.30 on K3, but every output token costs $50 per million versus $15. Fable 5.1 only comes out cheaper when a workload reads more than about 700 cached tokens for every output token, and even then its $12.50 cache-write rate usually erases the edge.
Should I compare Kimi K3 with Claude Opus 5.5 instead? For most teams, yes. Anthropic released Opus 5.5 on September 22, 2026 at $4/$20, and its docs tell most users to start with Opus rather than Fable. Anthropic reports Opus 5.5 beating Fable 5.1 on Terminal-Bench 4.0 (66.4% vs 55.8%), and it tops the current Artificial Analysis index. That makes it K3’s real price-performance rival.
Is Kimi K3 faster than Claude Fable 5.1? No, despite a faster first token. Artificial Analysis measures K3 at 35 output tokens per second versus 55 for Fable 5.1, and K3 spends longer reasoning: its first answer token arrives after about 61 seconds versus 21 seconds for Fable 5.1, and its average benchmark task takes about 21 minutes versus about 7 minutes.
The bottom line
Fable 5.1 widened the quality gap over Kimi K3 and narrowed the price gap only for cache-heavy work. On independent tests it leads 51 to 44, finishes tasks about 3x faster, and is clearly ahead on terminal-style coding and hard reasoning. K3 answers with a measured cost per task about half of Fable 5.1’s, a long-context edge, $3 cache writes and open weights that no Anthropic model offers.
If you already pay for Fable 5, move to 5.1: same prices, smaller bill, better model. If you are choosing between K3 and Anthropic from scratch, benchmark K3 against Opus 5.5 first, because that is where Anthropic’s price-performance now sits. Keep Fable 5.1 for the jobs where both of them fail.
Keep reading: our original Kimi K3 vs Claude Fable 5 breakdown, every K3 API provider and price, the Kimi K3 open weights guide, what the Kimi K3 License actually allows, and how K3 stacks up against OpenAI’s latest in Kimi K3 vs GPT-6 Astra.