AI News Roundup – 22 September 2026: GPT-6, Xiaomi's New Frontier Model, and More

OpenAI Unveils GPT-6 Sol and Luna

The day's biggest announcement came from OpenAI, which introduced two new models under the GPT-6 family: Sol and Luna. According to OpenAI, the pair are designed to bring frontier intelligence to everyday work, each offering a different balance of capability and cost. Sol and Luna represent OpenAI's latest push to make high-end AI more accessible across a range of use cases, from demanding professional tasks to more cost-sensitive applications. No specific pricing or availability details were included in the announcement, but the two-model structure suggests OpenAI is deliberately segmenting its offering — a strategy familiar from the GPT-4 era.

Xiaomi's MiMo-V2.6-Pro Crowned Top Open-Weights Model — Trained for $3 Million

Xiaomi has made a significant mark on the open-weights AI landscape with the release of MiMo-V2.6-Pro 1T-A42B, which has been dubbed the new top open-weights model by AI newsletter Latent Space. What makes this release particularly striking is the reported training cost: just $3 million — a fraction of what comparable frontier models typically require. The result positions Xiaomi as a serious Chinese frontier lab in the global AI race, and raises important questions about the economics of training large models. The release is likely to intensify debate around open-weights versus closed models, especially given the performance-to-cost ratio on display.

OpenAI Sets Out Principles for Third-Party AI Safety Assessments

Separately, OpenAI released a policy document outlining its priorities and principles for third-party assessments of frontier AI models. The document calls for rigorous, secure, and independent evaluations of both models and their safeguards. This move comes amid growing regulatory and public pressure on AI labs to demonstrate safety through transparent, external review processes — not just internal testing. The principles document signals that OpenAI is attempting to shape what credible external oversight looks like, ahead of what many expect to be a wave of government-mandated assessments globally.

UK AISI and EvalEval Work to Make AI Benchmark Results Reproducible

Reproducibility in AI evaluation has long been a thorny issue, and a new collaboration between the UK AI Safety Institute (AISI) and EvalEval is trying to address it head-on. Hugging Face shared details of their joint effort to make benchmark results more reproducible — a crucial step if the broader research community is to trust leaderboard rankings and safety evaluations. The initiative matters particularly as models proliferate and organisations increasingly rely on benchmarks to make procurement and deployment decisions.

Hugging Face's Transformers Now Supports llama.cpp Quantised Models

In a practical update welcomed by the open-source community, Hugging Face announced that its widely-used Transformers library now runs llama.cpp quantised models. This integration lowers the barrier for developers who want to run efficient, compressed models locally without switching between separate toolchains. The move reinforces Hugging Face's position as a central hub for the open-source AI ecosystem, and is likely to be particularly useful for developers working on resource-constrained hardware.

oMLX Creator Joins Hugging Face to Support the MLX Community

Hugging Face also announced that Jun Kim — the creator and maintainer of oMLX — is joining the organisation to support the MLX community. MLX is Apple's machine learning framework optimised for Apple Silicon, and the oMLX project has been a key resource for developers building on that platform. Bringing Jun Kim in-house suggests Hugging Face is serious about deepening its support for Apple Silicon-based development, an increasingly important segment as more developers work on Macs.