A summary of recent AI model developments, including Moonshot's Kimi K3, OpenAI's GPT-5.6 family, and various updates from Anthropic and Microsoft.
Key facts
- •Kimi K3 is the largest open-source model at 2.8 trillion parameters and costs $15 per million output tokens.
- •OpenAI's GPT-5.6 family includes the flagship model Sol, which outperformed Fable 5 in reasoning benchmarks.
- •Anthropic's Sonnet 5 starts at $2 per million input tokens, with a planned price increase to $3 in September.
- •The US Department of Commerce lifted export controls on Mythos 5 and Fable 5 on June 30.
- •Microsoft's new 35-billion-parameter model is designed for multi-step agentic tasks and trained on commercially safe data.
AI labs continue to release new models, with recent entries from Moonshot, OpenAI, Meta, Anthropic, and Microsoft. These releases focus on agentic tasks, reasoning capabilities, and cost efficiency, while navigating evolving government oversight and safety standards.
By the numbers
Moonshot's Kimi K3 and OpenAI GPT-5.6
Moonshot's Kimi K3 is a 2.8 trillion-parameter open-source model designed for long-horizon coding and reasoning. It outperformed Anthropic's Fable 5 on the Arena benchmark for front-end coding, though Moonshot noted its overall performance remains competitive rather than superior to Fable 5. Moonshot plans to release the model weights by July 27. OpenAI released the GPT-5.6 family, including the flagship model Sol, alongside Terra and Luna. OpenAI stated that Sol outperformed Fable 5 in adaptive and medium reasoning on the UC Berkeley Agents Last Exam benchmark. The company noted that the government was involved in the timing of this release, reflecting an informal partnership between labs and the Center for AI Standards and Innovation.
Anthropic Model Updates and Restrictions
Anthropic released Sonnet 5, which the company says performs similarly to Opus 4.8 but at a lower cost. Sonnet 5 is now the default for Free and Pro plans. Anthropic previously faced government-ordered access restrictions on its Mythos 5 and Fable 5 models, with the Department of Commerce lifting export controls on June 30. Fable 5, described as a 'defanged' version of Mythos, was subject to a temporary access extension for paid users through July 12.
Meta and Microsoft Developments
Meta released Muse Spark 1.1, which scored above Anthropic's Opus 4.8 and OpenAI's GPT-5.5 in financial analysis and job-related tasks. The model is positioned toward personal agentic use cases. Meanwhile, Microsoft introduced a 35-billion-parameter model designed for multi-step agentic tasks. Microsoft stated the model was trained on clean, commercially safe data to address enterprise concerns regarding copyright.
Timeline
- June 26The government re-allowed access to Mythos 5 for certain partners.
- June 30The Department of Commerce lifted export controls on Mythos 5 and Fable 5.
- July 1Anthropic began restoring global access to Fable 5.
- July 12Anthropic extended paid plan access to Fable 5.
- July 27Moonshot plans to release Kimi K3 model weights.
Advertisement
This article was independently rewritten by ManyPress editorial AI from reporting originally published by ZDNET AI.
