Kimi K3 vs Claude Fable 5 and GPT-5.6: How the Open 3T Model Stacks Up
Moonshot's Kimi K3 is the first open 3-trillion-parameter model: fast, cheap, and huge on context. Here is how it actually compares to Claude Fable 5 and GPT-5.6 on the benchmarks, and where each one wins.
Moonshot's Kimi K3 is the release everyone is talking about, and for good reason. It's the first open model to reach the 3-trillion-parameter class, it ships with a 1-million-token context window, and it costs a fraction of the proprietary leaders. If you build with these models, the useful question is how it actually stacks up against the two at the top: Claude Fable 5 and GPT-5.6 Sol.
The short version: on raw coding benchmarks, K3 trades blows with both and wins a few outright. On overall intelligence and day-to-day polish, it still trails them. And on price and openness, nothing else here comes close. It isn't a clean winner. It's a genuine third option with a very different shape.
What Kimi K3 is
Kimi K3 is Moonshot AI's new flagship, and the headline is scale plus openness. It's a 2.8-trillion-parameter mixture-of-experts model that activates only about 50 billion parameters per token (16 of 896 experts), so it's far cheaper to run than its size suggests. It has native vision, an enormous 1-million-token context window, and it's trained with MXFP4 quantization, which shrinks the full model to roughly 1.4 TB of weights, about a quarter of what a normal 16-bit model this size would need. It launched on July 16, 2026 through the API and the Kimi apps, with the full open weights due by July 27. Moonshot itself is candid, in its Kimi K3 announcement, that overall performance still trails Fable 5 and GPT-5.6 Sol while calling K3 frontier-level, and that framing turns out to be about right.
Kimi K3 vs Claude Fable 5
Head to head with Fable 5, K3 splits the coding benchmarks. It's ahead on Terminal-Bench 2.1 (88.3 vs 84.6), Program Bench (77.8 vs 76.8), and SWE Marathon (42.0 vs 35.0, though Fable 5 reportedly hit fallbacks on about a third of those tasks, which likely dragged its number down). Fable 5 pulls clearly ahead on FrontierSWE (86.6 vs 81.2) and edges K3 on DeepSWE (70.0 vs 67.5).
On broader measures, Fable 5 leads: the Artificial Analysis Intelligence Index (about 60 vs 57) and GDPval-AA v2 (1,815 vs 1,687), a benchmark aimed at real economic-value work. The pattern is clear. K3 is genuinely competitive on hard coding, and Fable 5 stays ahead on general capability and on the harder software-engineering tasks.
Kimi K3 vs GPT-5.6 Sol
Against GPT-5.6 Sol it's a similar story with different edges. K3 essentially ties Sol on Terminal-Bench 2.1 (88.3 vs 88.8, half a point) and Program Bench (77.8 vs 77.6), and beats it on SWE Marathon (42.0 vs 39.0) and FrontierSWE (81.2 vs 71.3). Sol takes DeepSWE clearly (73.0 vs 67.5) and leads on overall intelligence (about 59 vs 57) and GDPval-AA v2 (1,748 vs 1,687).
Both K3 and Sol are strong agentic coders. K3's advantage is cost and context; Sol's is consistency and polish. For how Fable 5 and GPT-5.6 compare to each other, I went deep on that in Fable 5 vs GPT-5.6.
The comparison
| Dimension | Kimi K3 | Claude Fable 5 | GPT-5.6 Sol |
|---|---|---|---|
| Model type | Open-weights (2.8T MoE, ~50B active) | Proprietary | Proprietary |
| Price per M (in / out) | $3 / $15 | $10 / $50 | $5 / $30 |
| Context window | 1,000,000 tokens | large | large |
| Intelligence Index (Artificial Analysis) | ~57 | ~60 | ~59 |
| GDPval-AA v2 | 1,687 | 1,815 | 1,748 |
| Terminal-Bench 2.1 | 88.3 | 84.6 | 88.8 |
| Program Bench | 77.8 | 76.8 | 77.6 |
| SWE Marathon | 42.0 | 35.0 | 39.0 |
| DeepSWE | 67.5 | 70.0 | 73.0 |
| FrontierSWE | 81.2 | 86.6 | 71.3 |
Read the numbers with care
Two honest caveats before you take any of this to the bank. First, these benchmarks were run through different harnesses, KimiCode, Claude Code, and Codex, and several of the gaps are smaller than the margin of error. A half-point on Terminal-Bench is noise, not a verdict, and Fable 5's low SWE Marathon score came with a high fallback rate that likely hurt it. Second, Moonshot is upfront about K3's rough edges: it's sensitive to how a harness passes back its own reasoning, so switching a live session to K3 mid-stream can destabilise it, it can be overly proactive and act instead of asking when intent is ambiguous, and its conversational polish still trails Fable 5 and GPT-5.6 Sol. The benchmark parity is real. The day-to-day experience is not yet at the level of the two proprietary leaders.
Where each one wins
Reach for Kimi K3 when:
- Cost matters. At $3 in and $15 out per million tokens, with $0.30 for cached input, it undercuts both proprietary models, and the open weights mean you can eventually self-host.
- You need enormous context. A 1-million-token window is far beyond the others here, which suits whole-repository work and long documents.
- Hard, long coding is the job, especially sustained sessions where it posts its best numbers.
Reach for Claude Fable 5 or GPT-5.6 Sol when:
- You want the highest general capability and the most polished, predictable behaviour.
- The work is conversational or user-facing, where K3's UX gap shows.
- You need a mature, stable harness today rather than a model whose ecosystem is still settling.
The bottom line
Like the last two frontier releases, this isn't a pick-a-winner decision. It's a routing decision. Kimi K3 doesn't dethrone Fable 5 or GPT-5.6 Sol on overall quality, but it doesn't need to. It's the first open model that can sit in the same conversation as them on hard coding, at a fraction of the price, with far more context, and soon, weights you can run yourself. That combination is what makes it matter.
Match the model to the task: K3 for cost, context, and heavy coding you want to own; the proprietary leaders for polish and the highest ceiling. And if you're weighing those two against each other, my Fable 5 vs GPT-5.6 comparison breaks that decision down.
Written by Usman Nasir — control systems engineer, Stockholm.