AI Model Launch Tracker

What we captured

Suno launched its V6 model and three new subscription tiers on Sept. 9, 2026. The rollout includes catalog-experimentation and revenue-sharing arrangements for participating BMG and Warner Music Group artists, while training-data compensation remains outside the announced model.

• • Suno described V6 as developed in partnership with BMG and Warner Music Group.
• • Participating artists can earn revenue when users experiment with their catalogs.
• • The announcement does not provide compensation for copyrighted music used in Suno's training data.

Source evidence

Confidence 99% / Completeness 87%

What we captured

OpenAI released two GPT-Image-2.5 image models for API use and made ChatGPT Images 2.5 available worldwide across ChatGPT and Codex.

• • Flare is positioned as the faster default model, with up to 50% lower latency than GPT-Image-2.
• • Sunburst is positioned for more demanding visual editing and generally takes longer to generate images.
• • The release adds xhigh and max quality tiers and new ChatGPT features including Sketch, templates, image comments, and shareable prompts.

Source evidence

Confidence 99% / Completeness 92%

What we captured

Thesys announced OUI-1, an open-weight DiffusionGemma fine-tune designed to generate OpenUI Lang interfaces and run locally for developers.

• • OUI-1 is a 26B-parameter model with 4B active parameters.
• • It generates OpenUI Lang programs constrained by developer-approved component libraries.
• • It scored 71.7% on Thesys’s structural-validity benchmark versus 13.0% for untuned DiffusionGemma.

Source evidence

Confidence 99% / Completeness 91%

What we captured

Nex-AGI launched the N2.5 family with 35B Mini, 397B Pro, and 1.6T Max models. Mini and Pro target multimodal computer use, while Max is text-only; Pro weights remained unavailable for download.

• • The family includes 35B Mini, 397B Pro, and 1.6T Max variants.
• • Mini and Pro are multimodal computer-use models; Max is text-only.
• • Pro weights were not available for download as of September 8, 2026.
• • Nex cites a two-H100 sample deployment for Mini and a 16-H200 reference deployment for Max.

Source evidence

Confidence 99% / Completeness 87%

What we captured

Superhuman AI reports the general availability of Astra, identified in the article as OpenAI’s most powerful model.

• • The article states that Astra became generally available.
• • The article describes Astra as OpenAI’s most powerful model.
• • The article says Astra’s launch video received 128 million views over the weekend.

Source evidence

What we captured

Alibaba’s Qwen team released Qwen-Drive 1.0, a multimodal driving model combining 3D spatial perception, traffic question answering, and route planning, through Hugging Face, ModelScope, and GitHub.

• • Qwen-Drive 1.0 is built on Qwen3.5-4B.
• • It adds modules for 3D environment mapping and route planning.
• • Reinforcement-learning retraining reduced simulated road departures from 24% to 12%.
• • The model’s explanations do not always correspond to its actual driving maneuvers.

Source evidence

Confidence 96% / Completeness 87%

What we captured

WeatherNext 3 is a newly released AI weather model that uses live geostationary satellite data to generate hourly forecasts at up to 5-kilometer resolution. It is available across Google weather products and through Google Cloud data services.

• • Uses live geostationary satellite data and hourly forecast updates.
• • Produces surface forecasts at up to 5-kilometer resolution.
• • Supports precipitation, wind, cloud-cover, and solar-irradiance forecasting.
• • Released across Google's consumer products and cloud/developer services.

Source evidence

What we captured

Google released Lyria 3.5 in the Gemini app and via API, with additional availability in Google Flow Music, Google AI Studio, and Google Vids.

• • The article states that Google released Lyria 3.5 in the Gemini app and via API on September 6, 2026.
• • Google describes the model as offering more expressive vocals and richer musical arrangements than its predecessor.
• • Google says the Lyria model was trained only on licensed content, without specifying the training data.

Source evidence

What we captured

Meta released Muse Voice Transcribe for real-time transcription, speaker separation, and sentence-boundary detection. The model is available in Meta AI, Muse Code, and through the Meta Model API.

• • Real-time model combining transcription, speaker separation, and sentence-boundary detection.
• • Supports more than 20 simultaneous speakers and over 70 languages.
• • Priced at $0.18 per audio hour.
• • Weights are not being released.

Source evidence

What we captured

IFM released six K2 Horizon language models from 0.9B to 375B parameters with weights and extensive training artifacts intended to support external inspection and reproducibility.

• • Six models were released, spanning 0.9B to 375B parameters.
• • The release includes weights, code, checkpoints, training materials, logs, and evaluation artifacts.
• • IFM disclosed benchmark exploitation that reduced the 375B-A23B TerminalBench score from 70.2% to 66.9% after audit.

Source evidence

Confidence 99% / Completeness 87%