Kimi K3 matches Claude Fable 5 in a three-task coding test at lower cost
A three-task test by The New Stack found Kimi K3 produced comparably correct changes to Claude Fable 5 for $2.13 versus $5.98, but took 28 minutes versus under seven. The result is useful workload evidence, not a general ranking: it covers one Rust repository, three tasks and two different coding-agent CLIs.
The New Stack tested Moonshot AI's Kimi K3 and Anthropic's Claude Fable 5 on the same three coding jobs in the Rust-based fd repository. Both models delivered working changes, but Kimi cost $2.13 versus $5.98 and took about 28 minutes 18 seconds versus 6 minutes 49 seconds.
That is a concrete cost-latency tradeoff, not proof that the models are equivalent across software engineering. The comparison used one repository, three prompts and different native agent tools: Kimi Code for K3 and Claude Code for Fable 5.
What the test measured
The tasks covered a bug fix, a multi-file refactor and a new --count feature. Both models found the same one-line bug fix and passed all 70 tests. On the refactor, both split the large configuration function into a new module and passed 264 tests. Both also implemented a working count flag.
Kimi used about 3.3 million tokens across the three tasks, compared with roughly 2.4 million for Fable 5. Its lower bill therefore came from lower API rates, not lower token consumption. Its slowest run was the refactor at 14 minutes 50 seconds, compared with 3 minutes 11 seconds for Fable 5.
The test author used identical prompts and fresh working folders, but the coding CLIs were not the same. Tool orchestration, context handling and retry behavior can influence both token use and elapsed time, so the result cannot isolate the model alone.
Broader comparisons remain mixed
Artificial Analysis currently scores Kimi K3 at 57 and the compared Fable 5 configuration at 60 on its Intelligence Index. Its comparison page lists blended prices of $2.31 and $7.70 per million tokens, respectively. BenchLM gives Fable 5 an 83.68 score versus Kimi K3's 80.96, while warning that only two of eight categories have comparable aggregates for both models.
Moonshot describes Kimi K3 as a 2.8-trillion-parameter multimodal model with a one-million-token context window. The model is available through Moonshot's services, while the company says the full weights will be released by July 27.
What teams should test
For production coding agents, repeat the comparison on representative repositories and measure:
- •task success and regression rates;
- •end-to-end latency, including tool calls and retries;
- •input, output and reasoning-token use; and
- •privacy, hosting and observability requirements.
Kimi K3's lower per-token pricing may suit asynchronous or high-volume work. Interactive development can put a higher value on turnaround time. The New Stack's result makes that tradeoff visible, but it does not settle it for every workload.
Key Points
- 1In one three-task Rust-repository test, Kimi K3 and Claude Fable 5 produced comparably correct changes, while Kimi cost $2.13 versus $5.98.
- 2Kimi took about 28 minutes 18 seconds compared with 6 minutes 49 seconds and used more tokens despite the lower bill.
- 3The test used different coding-agent CLIs, so teams should reproduce the comparison on their own repositories and workflows.
Scoring Rationale
The comparison provides useful task-level cost and latency evidence for two frontier coding models. Its practical value is meaningful, but the result is limited to one repository, three tasks and different coding-agent CLIs, so it should not be treated as a universal model ranking.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
