Kimi K3 has became so popular that the company has closed registration for new subscription plans. For a long time, it simply showed "Sold Out", and now there is a waiting list.
Opus 5 Frequently Makes Mistakes
https://www.anthropic.com/news/claude-opus-5
https://news.ycombinator.com/item?id=49038433
On July 24, Anthropic released Claude Opus 5, which performs close to Fable 5 on benchmarks but is more affordable. In Anthropic's demo, without access to "viewing" the blueprint, the model writes a CV pipeline from pixels and assembles a 3D FreeCAD model by itself. It also performs root cause analysis in a package manager, whereas other models only apply superficial fixes.
In practice, the main issue turned out to be this excessive agentic behavior. If you forget a Postgres password, instead of asking, it spins up a Kind cluster. If asked not to touch the sandbox, it breaks the rules. It writes specs, documentation, and research poorly because it rushes to "do" rather than clarify. False positives occasionally occur during code reviews. It also frequently burns tokens on clever but unnecessary workarounds.
The Kimi K3 Effect
https://stephen.bochinski.dev/blog/2026/07/18/the-kimi-k3-moment/
Stephen Bochinski describes what many are experiencing now: in standard coding tasks, K3 and Claude Fable are almost indistinguishable, offering similar quality and using a comparable number of tokens. However, the price is vastly different. The broader theme of the blog post is the failure of US AI policy: Fable has been heavily restricted and constrained by safeguards, while the Chinese open frontier model is publicly available without these limitations. The author’s conclusion is harsh: there is currently no reason to pay for Claude if K3 maintains the same level of quality.
Kimi K3 is Cheaper at 256k
https://news.ycombinator.com/item?id=49101852
https://www.kimi.com/code/docs/en/kimi-code/models
Moonshot released k3-256k in Kimi Code: the same quality within a 256k limit, but at half the price and quota consumption compared to the 1M context version. However, they are reserving this for current subscribers to avoid Anthropic's scenario of "accepting everyone and quietly cutting limits."
Running K3 locally is still out of reach for a standard laptop, requiring 1.5TB of VRAM on MXFP4. Therefore, the 256k discount is primarily relevant for cloud subscriptions, not self-hosting. The logic mirrors OpenAI's 272k threshold: a long KV cache is expensive, so a separate SKU with a smaller maximum context is cheaper to serve. Switching from 256k to 1M does not invalidate the cache (according to their documentation), allowing users to start cheaply and expand the window only when necessary.
Kimi K3 Released
https://huggingface.co/moonshotai/Kimi-K3
https://news.ycombinator.com/item?id=49065752
https://huggingface.co/chat/models/moonshotai/Kimi-K3
The promise from the announcement has been fulfilled: full weights are available on Hugging Face under the Kimi K3 License. It features a 2.8T MoE architecture (activating 16 out of 896 experts, with 104B active parameters), a 1M context window, native vision support (MoonViT-V2), KDA + AttnRes, and native MXFP4. A chat interface is available on HuggingChat. Although very few can run this locally right now, major GPU cloud providers have already stepped in.
Combining Kimi K3 and Fable
https://fireworks.ai/blog/kimik3-fable
Fireworks evaluated K3 and Fable 5 across 1,030 agentic tasks (SWE, terminal, algorithms, multi-language, legal) in a single harness. The results showed a near-tie in SWE (K3 92.4% vs. Fable 92.6%), but their specializations differ. K3 is stronger in symbolic math, developer tooling, and long-terminal tasks (security, cryptography, system administration), while Fable excels in web tasks, data visualization, and broader multi-language support.
The article concludes that instead of relying on a single SoTA model for everything, a better approach is combining them: the open-weight model (K3) initiates the task, the premium model (Fable) finalizes it, and a router handles the switching. Oracle routing (selecting the cheapest correct model) achieves up to 93% accuracy and up to a 50× cost reduction compared to using Fable alone for long tasks.
Fugu isn’t a model
https://sakana.ai/fugu/
Sakana Fugu is not a single LLM but a multi-agent system presented as a single model: it exposes an OpenAI-compatible API, while dynamically orchestrating a pool of external frontier models under the hood. Information regarding which specific models are used and how they are routed remains proprietary. This represents the opposite end of the spectrum of the same trend observed with Fireworks routing.
#kimi #anthropic #fable #fireworks #fugu #agenticcoding #openweights