What just happened in frontier model releases?

Multiple frontier model releases and updates occurred. Moonshot AI released Kimi K3, a 2.8 trillion parameter open-source multimodal model with 1M context, approaching top closed-source models at one-third the price [5336, 5138, 4761, 3846]. Kimi K3 beats Claude Opus 4.8 on four of five benchmarks and matches Fable 5 and GPT-5.6 Sol on several [6117]. It built a walkable voxel town with drivable cars in 3 shots for $3.24 API cost, cheaper than Fable 5 ($10.80) and GPT-5.6 Sol ($6) [6118]. API priced at roughly a third of Fable 5 [6312]. Qwen 3.8 Max is now available on the chat.qwen.ai app [5334]. Anthropic released Claude Sonnet 5, replacing Sonnet 4.6 as default, with agentic performance close to Opus 4.8 at 40% API price [354]. OpenAI released ChatGPT Work, an agent powered by Codex and GPT-5.6 [1041]. GPT-5.6 is now the preferred model in Microsoft 365 Copilot [3562]. LingBot-Video, a 30B-param MoE video model with 3B active params, outperforms Wan2.6, Seedance 1.5 Pro, and Cosmos3 Super on RBench [1042]. GPT 5.6 Sol passed the Replit Benchmark by creating a full Replit clone in one prompt [993]. OpenAI expanded Daybreak to democratize patching vulnerable software, including a Codex Security plugin and GPT-5.5-Cyber model [159]. OpenAI's GPT-5.6 Luna outperforms GPT-5.5 at its highest reasoning setting while costing 25x less [1346, 2279]. OpenAI evolves its Bio Bug Bounty into an ongoing private program with doubled rewards up to $50K [1347]. Andrew Ng released a free 2-hour course on building agentic skills with Claude, covering pre-built skills, MCP, subagents, and long-running agents [2888, 1851, 1771, 1460]. Google upgraded Gemini with capabilities to read web pages, compare products across tabs, edit images, auto-browse, connect personal data, and integrate with business tools [1864], including voice-controlled image editing, new image templates, and Business Notebooks for company-specific answers [1865]. Kimi K3 release video was generated using Kimi K3 itself, demonstrating creative capabilities [3392]. Claude Code team released a free course on loop engineering with Fable 5, covering agentic loops and deployment [3391]. ExLlamaV3 v1.0.0 production release brings major performance improvements including new attention kernel, tensor-parallel support, and optimized kernels [2605]. Google's Gemini Omni Flash model debuted at #1 on the Artificial Analysis Text to Video and Image to Video Leaderboards, surpassing ByteDance's Seedance 2.0 [2489]. Unsloth officially supports AMD GPUs for inference, fine-tuning, RL, and deployment, with up to 70% less VRAM usage [6298].[28]

Events in order: Moonshot AI released Kimi K3 [5336, 5138, 4761, 3846]; Kimi K3 beats Claude Opus 4.8 on four of five benchmarks [6117]; Kimi K3 built a voxel town cheaply [6118]; Kimi K3 API priced at a third of Fable 5 [6312]; Qwen 3.8 Max available [5334]; Anthropic released Claude Sonnet 5 [354]; OpenAI released ChatGPT Work [1041]; GPT-5.6 preferred in Microsoft 365 Copilot [3562]; LingBot-Video outperformed competitors [1042]; GPT 5.6 Sol passed Replit Benchmark [993]; OpenAI expanded Daybreak [159]; OpenAI's GPT-5.6 Luna outperforms GPT-5.5 at lower cost [1346, 2279]; OpenAI evolves Bio Bug Bounty [1347]; Andrew Ng released agentic course [2888, 1851, 1771, 1460]; Google upgraded Gemini [1864, 1865]; Kimi K3 release video generated using Kimi K3 [3392]; Claude Code team released free course on loop engineering [3391]; ExLlamaV3 v1.0.0 production release [2605]; Google's Gemini Omni Flash debuted at #1 on leaderboards [2489]; Unsloth supports AMD GPUs [6298]. Claims are emerging except ChatGPT Work [1041] and GPT 5.6 Sol [993] which are corroborated.[28]

Kimi K3 model weights will be released on March 27, as announced by Moonshot AI [3390]. Qwen3.8 is coming as an open-weight model, continuing the trend of open releases [5509].[2]