Kimi K3 Just Beat Fable 5 on Agent Benchmarks

Moonshot AI released Kimi K3. On several agent benchmarks it cleared Fable 5 and GPT-5.6 Sol, while pricing itself like Sonnet 5. I have not seen open-weight models sit at that tier before.

Here is the picture in six parts.

Agent scores that are not niche

From the model blog: Program Bench 77.8, SWE Marathon 42.0, BrowseComp 91.2, SpreadsheetBench 2 at 34.8. Those are all ahead of Fable 5 on the numbers they published. FrontierSWE is 81.2, behind Fable 5 at 86.6 but well ahead of GPT-5.6 Sol at 71.3. Coding and browsing both look strong, which is the opposite of a one-trick specialist.

Sonnet-class price, not frontier-class price

API pricing is about $3 input and $15 output. That is roughly 30% of top frontier list prices. Compared with earlier Kimi versions the price jumped more than 3x, so this is not a free lunch relative to K2. You still have to measure tokens per completed job against Claude and Codex. Even so, if you run agents all day, token cost is payroll. The gap is hard to ignore.

The architecture is not a small bump

Kimi Delta Attention changes the attention path itself and claims up to 6.3x faster decoding at 1M context. Long context usually gets slower as it grows; they attacked that as a structural problem.

Attention Residuals reinforce information flow between layers. They add under 2% compute and report about 25% better training efficiency. Cheap trade for a large training gain.

The MoE design activates 16 of 896 experts. A 2.8T-parameter model that only turns on a couple percent of capacity at inference is how the API price can exist at all.

Together they report 2.5x better scaling efficiency than K2: same compute budget, more intelligence per unit.

It optimizes its own kernels

Kernels are the low-level GPU code that runs attention and friends trillions of times during training. Faster kernels mean cheaper, faster training.

Moonshot gave K3 a production-scale task: speed up an FLA Triton AttnRes kernel (96 layers, 8192-dim, 8192 tokens) without changing numerical results. Approximate shortcuts were disallowed because wrong math would change training.

Over 15 hours without human intervention, K3 redesigned a two-phase kernel algorithm and fused operations to cut GPU memory traffic. Reported speedup was about 2.5x. Fable 5 reached a similar end state on the same task; K3 improved faster per iteration.

The important detail is that this kernel is not a toy benchmark. It sits inside their real training stack. If a model can accelerate its own training path, the next version gets cheaper to train, which funds more infrastructure work. That is a self-evolving loop starting to move. K3 being near the top of the scoreboard matters partly because of that loop, not only because of one release.

Weights drop July 27

Parameter count is 2.8T. Prior open-weight highs were roughly 1T for Kimi K2 and 671B for DeepSeek-R1. Nearly 3x the previous open max, while frontier labs often refuse to state size at all.

Open weights mean you can host, fine-tune, and keep data on your own hardware. Today you can use Kimi.com and the API at platform.kimi.ai. After the 27th, self-hosting becomes a real option for teams that need it.

Third shock after DeepSeek-R1 and GLM-5.2

DeepSeek-R1 and GLM-5.2 proved that strong performance can land at a brutal price. K3 is different in shape: multi-benchmark leadership, about 30% of frontier price, and open weights on a giant model. If tokens-per-task and wall-clock also hold up under real harnesses, the market pressure gets real.

I am putting Kimi K3 through harness runs this week. If frontier labs stay quiet until weights ship, lower-cost models may matter more than people budgeted for.

I still want independent confirmation on long-running agent jobs, not only leaderboard slices. Benchmarks move markets; production harnesses decide whether you switch.