Kimi K3 beats GPT-5.6 Sol at agentic work — and goes open-source July 27

The model that just outperformed OpenAI‘s flagship on the benchmark that matters most for long-horizon coding and knowledge work is available today via API. It is not from OpenAI, Google, or Anthropic. Kimi K3 is the latest release from Moonshot AI, a Beijing startup best known for the Kimi chatbot, and it is the largest open-weights language model ever released: 2.8 trillion total parameters in a sparse architecture that keeps costs down by activating only a fraction of them per request.

On AA-Briefcase — Artificial Analysis’s agentic evaluation designed to simulate real knowledge work rather than textbook problems — Kimi K3 scored 1,527, placing second only to Claude Fable 5 Max at 1,587 and beating GPT-5.6 Sol Max at 1,495. On the broader GDPval-AA benchmark, the model places third at 1,687, behind Fable 5 Max (1,815) and GPT-5.6 Sol Max (1,747.8). The performance gap between first and second on the agentic benchmark is narrower than the gap between second and fifth.

The pricing difference is harder to dismiss than the benchmark rankings. Kimi K3 charges $3 per million uncached input tokens and $15 per million output tokens. Claude Opus 4.8 charges $5 per million input and $25 per million output. For teams running high-volume agentic workflows, Moonshot reports cache hit rates above 90% on coding workloads, bringing the effective input cost to $0.30 per million tokens — a figure that changes the economics of deploying frontier-class AI at scale.

Two architectural innovations underpin the model. Kimi Delta Attention is a hybrid linear attention mechanism that the company says enables 6.3x faster decoding in million-token contexts versus standard attention. Attention Residuals is described as a drop-in replacement for standard residual connections that delivers consistent performance gains as the model scales. The 1-million-token context window — enough to hold roughly ten full-length novels at once — is live and functional, not a theoretical specification.

There is a distinction between ‘available today’ and ‘open-source’ that matters here. Kimi K3 is accessible now via API and the Kimi app, meaning requests pass through Moonshot’s servers. The model’s actual weights — the trained parameters that would let anyone deploy it on their own infrastructure — are not yet public. Moonshot plans to release them on July 27 under a Modified MIT license, the same terms applied to the previous K2 model. For most developers, the API is what they need; for organizations with data sovereignty or compliance requirements, the weights release is the relevant date.

Native multimodal support covers text, images, and video input within the same API call. The model’s 2.8 trillion total parameter count refers to the full set of sparse MoE parameters; the active parameter count per forward pass is considerably lower, which is how Moonshot keeps inference costs down. Running the full model locally would require hardware far beyond a consumer workstation. What the open weights will enable is deployment on enterprise-scale infrastructure without routing data through a Chinese-owned API.

The weights release July 27 will determine how much of the benchmark advantage survives in real deployment. When Moonshot released K2 under comparable open terms, developer adoption moved faster than the company expected — partly because the combination of near-frontier performance and MIT-style licensing removed legal friction for teams that needed both. K3 is a larger bet on the same strategy.

Tags: AI models, GPT-5.6, Local AI, Kimi K3, Moonshot AI, open-source AI