Kimi K3's open weights land this week, and its benchmark card puts it in a real fight with Claude Fable 5 — not a curiosity next to a champion, but a trade of blows. So the question a founder building coding agents should actually ask is not "which is best?" It's "which one do I make the default, and which do I keep in reserve?"
Those are different questions, and the second one has a cleaner answer.
The scoreboard is close — and it specializes#
Across roughly 14 shared benchmarks, Fable 5 wins about 8 and K3 about 6. Read past the tally, though, and the wins aren't scattered:
- Fable 5 owns deep, single-turn reasoning. FrontierSWE 86.6 to K3's 81.2; a lead on DeepSWE. When the task is "crack one very hard problem in one pass," the closed frontier is still ahead.
- K3 owns sustained execution and frontend. SWE Marathon 42.0 to Fable 5's 35.0 — the long-horizon, multi-file regime an agent lives in. Terminal-Bench 2.1 at 88.3. And #1 in the Frontend Code Arena, ahead of Fable 5, if you generate UI.
Neither dominates. They specialize. That single fact is why the smart move is a router, not a religion.
The tiebreaker isn't the score#
For a team of one, a two-point swing on someone else's benchmark harness is not what decides the invoice. Two other things do:
Price. K3's API is $3 / $15 per million tokens — on the order of a fifth of a closed flagship's output price. At the volume an autonomous coding agent burns, that gap compounds into real money every month.
Openness. K3's weights are public. You can self-host for data residency, or pin an exact version a vendor can't deprecate under you. Fable 5 is closed — you rent capability you can't hold. For most founders that's fine; for some it's disqualifying.
On a benchmark card they trade blows. On a monthly invoice, one of them costs five times more than the other for the work they both do well. That's the tiebreaker.
The default that fits most founders#
Make K3 the workhorse. Point the bulk of your agentic coding at it — long loops, tool use, boilerplate, and especially frontend — because that's where it's competitive-to-winning and cheaper. You are not compromising on those tasks; you're paying less for a model that's at least as good at them.
Keep Fable 5 as the escalation target for the hard reasoning tail: the subtle refactor, the architect-level design problem, the one-shot puzzle where the closed frontier's extra points actually change the outcome. Route to it by task class, not by default.
Flip the default only if your work inverts the usual profile:
- Default to Fable 5 if your product is reasoning-dense one-shot problem-solving rather than long execution, or if you're deep enough in Anthropic's ecosystem — Claude Code, the Agent SDK, Skills — that the tooling premium pays for itself.
- Default to K3 if you value open weights, run high output volume, or ship frontend and web apps. That describes a large share of early-path builders.
The one prerequisite#
None of this works if your pipeline is welded to one vendor's SDK and prompt format. The whole strategy is a config-level routing decision, which means you have to keep the model swappable before you need to swap it. Put a thin router in front of your coding agent, measure cost per completed task rather than cost per token, and let the numbers set the threshold where a task earns the expensive model.
The open challenger caught the closed champion on the work that fills an agent's day. It didn't catch it everywhere — but "cheaper, open, and at least as good at most of what I do, with a premium model on call for the hard 10%" is not a close call for a founder. It's the whole playbook.
