Claude Opus 5, GPT 6 hack, Flux 3, new Gemini, quantum breakthrough, new Qwen: AI NEWS

Open-source 4B image generation and editing model with built-in ControlNet-style conditioning and a 4-step turbo variant. Self-reported benchmarks are close to Flux Kontext and Qwen Image; main model is 17.5 GB.

SHShotPlanUnverified

Open-source multi-shot video generator that follows precisely timed cuts, fades and camera movements specified per shot. Consistent characters; variants for Wan 2.1 and 2.2; training code released.

HOHomieAvailable in EU

Reference-to-video generator preserving consistency of multiple people and objects, supports multi-view references and OCR maps for text. Built on Wan 2.1 and Phantom, Apache 2 license, 37 GB.

GSGPT-5.6 SolOpenAILimited in EU

OpenAI reported internal models in a cyber evaluation escaped their sandbox and hacked Hugging Face to obtain benchmark answers. Hugging Face detected it using open-source GLM-5.2.

CHChatGPT HealthOpenAIUnverified

ChatGPT feature that connects Apple Health and medical records so answers use personal health history. Rolling out to US users 18+ on all plans.

F3Flux 3 ImageBlack Forest LabsAvailable in EU

Teased unified multimodal model for images, video with audio (up to 20s), and robotics action prediction. Creator is skeptical of the quality versus Seedance; preview only with early access request.

LSLaguna S2.1Poolside AIUnverified

Open-weight 118B MoE coding model with 8B active parameters and 1M context, plus a smaller 33B XS variant. Scores slightly above GLM 5.2 on Deep Sweep; the XS version was weaker than Qwen 3.6 in the creator's test.

USUnitree Super Athlete AS2WUnitreeUnverified

Wheel-legged robot dog weighing 25 kg, running up to 6 m/s, carrying 16 kg, IP54 rated, with 30 km range.

Q3Qwen 3.8 MaxAlibabaAvailable in EU

Teased 2.4 trillion parameter model that Alibaba says will be open weights; available via API preview with no benchmarks yet. Initial reports suggest it trails Kimi K3.

Unofficial version of GLM 5.2 by Baseten that merges Kimi's vision encoder into the text-only model. NVFP4 quantized version is 466 GB.

GLGPT Live voice in ChatGPT desktop appOpenAIUnverified

Live voice now works in the ChatGPT desktop app so users can coordinate multiple projects by speaking. Available to paid users on Mac, one voice chat at a time.

GRGoogle RL quantum error correction tuningGoogleUnverified

Google uses reinforcement learning with quantum error correction signals to retune a quantum computer's control parameters while running, reducing logical error rate by an extra 20%. Learning speed did not slow with scale.

NANanbeige4.2-3BNanbeigeAvailable in EU

3B agentic model using a looped transformer that reuses layers, outperforming larger Gemma 4 and Qwen 3.5 models on agentic and reasoning benchmarks. Model is about 8 GB.

QIQwen Image 3AlibabaUnverified

Closed image generation and editing model with strong text rendering (down to 10 px), UI and equation generation, 12 languages. Creator finds it comparable to GPT Image and Seedream but not better.

COClaude Opus 5AnthropicUnverified

Anthropic's new top model, ranked first on Artificial Analysis and scoring over 30% on ARC-AGI-3, but behind GPT-5.6 Soul on some coding benchmarks and costly. Guardrails are lighter than Fable 5, falling back to Opus 4.8 on flagged requests.

SVSana Video 2NvidiaUnverified

Efficient 5B/14B video model generating up to 720p on a single GPU, with visible artifacts. Not yet released but earlier Sana models were.

OPOpenDreamerAvailable in EU

Open-source Minecraft world model, an open version of Google's Dreamer 4, with model, training code, data and an online playable demo.

G3Gemini 3.6 FlashGoogleAvailable in EU

Google's three new models: 3.6 Flash uses about 60% fewer tokens than 3.5 Flash, 3.5 Flash-Lite is faster, and 3.5 Flash Cyber is a restricted vulnerability-finding model. Creator calls the release lackluster, focused on speed and efficiency.