AI labs are shipping new models like they're trying to fill a quota, and Moonshot AI's Kimi K3 is the latest contender to enter the ring. At 2.8 trillion parameters, it's the largest open-source model on the market, designed for "long-horizon coding, knowledge work, and reasoning" - basically, all the stuff that makes humans feel inadequate. For context, that's nearly twice the size of DeepSeek V4 Pro, the model that first gave US labs a collective panic attack back in January 2025.

Kimi K3 managed to top Anthropic's Fable 5 on the Arena benchmark for front-end coding, which measures complex agentic coding tasks. However, Moonshot was quick to note that overall, Kimi K3 doesn't actually beat Fable 5 - it just rivals it and OpenAI's GPT-5.6 in a few individual benchmarks. So, a win is a win, but maybe not the kind that warrants a ticker-tape parade. Moonshot plans to release the model's weights by July 27, presumably so everyone can see how the sausage is made.

Why it matters: Kimi K3's competitive performance and record-breaking open-source size have revived the eternal debate about whether proprietary American frontier models are worth their hefty price tags. Fable 5 costs $50 per million output tokens, while Kimi K3 is a mere $15. But there's a catch: open-source models lack the safety guardrails of proprietary ones, and those risks are amplified by the fact that Moonshot is a Chinese startup - because nothing says "trust us" like a government that's perpetually suspicious of your home country.

The article also covers a slew of other model releases, including OpenAI's GPT-5.6 family, Meta's Muse Spark 1.1, GPT-Live-1, Anthropic's Sonnet 5, and the controversial Fable 5 and Mythos 5. The GPT-5.6 family includes Sol, Terra, and Luna, with Sol beating Fable 5 in adaptive and medium reasoning on UC Berkeley's Agents Last Exam. Meta's Muse Spark 1.1 focuses on "personal superintelligence," which apparently includes organizing dinner parties - because nothing says cutting-edge AI like planning your next potluck.

GPT-Live-1 aims to make ChatGPT voice conversations more natural, allowing you to interrupt, pause, and ask it to slow down, basically treating it like a human who's been told to listen for once. Sonnet 5 from Anthropic offers similar performance to Opus 4.8 but at a lower cost, though it ironically shows a higher rate of misaligned behavior than Mythos Preview. And the saga of Fable 5 and Mythos 5 continues, with both models being pulled by the US government after just four days, only to be partially restored later. Amazon researchers even jailbroke Fable 5, because apparently, safety measures are more of a suggestion.

Microsoft also released its first reasoning model, a 35-billion-parameter affair designed for multi-step agentic tasks, because it's never too late to join the party. And Anthropic's Opus 4.8 offers faster thinking modes for one-third the cost, with an emphasis on "prosocial traits" like supporting user autonomy - because if there's one thing we need from AI, it's lessons in being a decent person.