Releases

Krea released Krea 3, an image-editing model delivering native-2K edits in under eight seconds, with product and character consistency and support for up to six reference images.

OpenAI began rolling out GPT-Live-1.1 for voice in ChatGPT, Codex and dots, with new voices, fewer pauses during voice interactions and improved accents in French, German, Hindi, Mexican Spanish, Korean and Japanese. API access is planned later.

Black Forest Labs released FLUX 3 Image Klein, a 5.6B-parameter open-weight model for text-to-image, editing and multi-reference generation with native 4K output. The distilled INT8 build makes a 1MP image in about 4 seconds on an RTX 5090; free commercial use for companies under $5M revenue (FLUX Kommunity License), API from $0.009 per image.

Microsoft AI finished pre-training MAI-Base-2, a 2.8T-parameter mixture-of-experts foundation model (80B active) trained on 50T tokens across 16k GB200 GPUs over 38 days. It is now in post-training; Microsoft plans to license the weights to trusted partners via Foundry rather than release them openly, and claims better pre-training efficiency than DeepSeek-V4-Pro-Base.

NVIDIA released Nemotron 3.5 Super VL, an open 120B-parameter (about 12B active) multimodal agent model trained on ~26T tokens, plus Nemotron 3.5 Bolt 74B-A9B (native 2-bit, fits in 24GB). NVIDIA claims Super VL is 5x faster than the next-best open model in its class; Bolt E4B is due Oct 23. Available on Hugging Face, OpenRouter, ModelScope and as NIM.

Meta released Muse Spark Open, an 816B-parameter mixture-of-experts model (42B active) with a 1M-token context, handling text, image, video and audio and built for coding and agents. Weights are on Hugging Face, SageMaker, Vertex Model Garden and Microsoft Foundry under the Muse Model License (limited commercial terms); it runs locally on two M5 Ultra Mac Studios via MLX.

Meta Superintelligence Lab released Muse Wonder and Muse Wonder Full, larger than Muse Spark and positioned as Meta's first frontier models. Meta says they score comparably to Fable 5.1 and GPT-6 Astra and lead on its MemoryBench continual-learning test (company claims). Muse Wonder costs $4 / $20 per million input/output tokens on the Meta Model API; Wonder Full is limited to Muse Code subscribers. Both power the Muse agent and were trained at the Prometheus data center in Ohio.

Microsoft AI released Microsoft-Decision-1, a text-only model that returns calibrated probabilities over fixed answer options for routing, classification, verification and agent control instead of generating text. Post-trained from Alibaba's Qwen3.5-9B, 32K context; available as a proprietary API in Microsoft Foundry and via OpenRouter's Decisions API at $0.042 per 1M input tokens with free output.

Anthropic released Claude Haiku 5.5, its fastest and most capable small model for high-volume, cost-sensitive workloads. Anthropic says it costs about 75% less per task on average than Haiku 4.5, supports adjustable effort, and is available across the Claude Platform, AWS, Google Cloud, and Microsoft Azure.

Mistral AI launched Mistral Large 4 in public preview, describing it as a natively multimodal, open-weight mixture-of-experts model with about 1.05 trillion total parameters and 49 billion active parameters. The preview API is available through Mistral Studio, with model weights planned for release by the end of October.

Google DeepMind released EmbeddingGemma 2, an Apache 2.0-licensed open-weight multimodal embedding model that maps text, code, images, video and audio into a shared 768-dimensional space. The modular model scales from 270 million parameters for text and code to 740 million parameters for all modalities.

Reflection made the weights for its first model, Beam, publicly available, establishing its open model family.

Reflection released Beam, a 501B-total-parameter sparse mixture-of-experts model with 23B active parameters for coding, reasoning and agentic work.

Google lists October 5, 2026 as the shutdown date for antigravity-preview-05-2026 in the Gemini API. Developers are directed to migrate to antigravity-preview-09-2026.

SpaceXAI retired Grok Voice Transcribe 1.0 and routed requests to Transcribe 2.0.

Suno launched Speech in beta, an audio model that generates spoken voice and original background music together in one cohesive track.

Cohere launched its fifth-generation frontier embedding models in Pro and Fast tiers.

Perplexity released a contextual embedding method, model and benchmark for retrieving answers with their supporting context.

Google announces Gemini 4 Argon, initially making it available to trusted cybersecurity defenders and evaluators ahead of broader rollout.

OpenAI released GPT-6.1 Sol in the API as an updated workhorse model.

ElevenLabs released the low-latency Turbo variant of Eleven v4.

ElevenLabs released Eleven v4 and v4 Turbo, its most expressive and fastest text-to-speech generation.

Anthropic released Claude Sonnet 5.5 as a faster, lower-cost upgrade to Sonnet 5 for everyday tasks, coding, and document work.

Google launches Gemini 3.8 Live with Live Avatar through Gemini Enterprise for real-time audiovisual agent experiences.

Google releases Gemini 3.8 Flash TTS for expressive, controllable text-to-speech generation.

Black Forest Labs released an open-weight 7B world-action model and reported first place on the RoboLab benchmark.

Skild published a post-training result showing a foundation policy learn dynamic, dexterous robotic behavior through long-horizon simulated self-play.

Anthropic released Claude Opus 5.5, the first model in the Claude 5.5 family, with higher capability and lower operating cost than Opus 5.

OpenAI added GPT-6 Sol and GPT-6 Luna to the API as capability and efficiency tiers in the GPT-6 family.

SpaceXAI released Grok 4.7 for coding and knowledge work with a larger base model, longer reinforcement-learning runs and stronger safeguards.

SpaceXAI released its second-generation speech-to-text model with improved transcription accuracy.

After two years in stealth, TypeSafe AI released Jev in early access: a decision model that returns typed outputs with calibrated probabilities instead of text. TypeSafe prices it at $0.042 per million input tokens with free output, and claims up to 193.6x faster and 444.6x cheaper than frontier LLMs on its own workflow evals.

Oracle added on-demand multimodal generation and embedding models from Cohere in additional regions.

Google launches an Extended Thinking variant of Gemini 3.8 Live for deeper real-time reasoning.

Google launches Gemini 3.8 Live for low-latency, real-time multimodal voice and video interactions.

Genspark introduced Gen-1 Slides, a model designed specifically for professional presentation creation.

Cognition released SWE-2, advancing capability and inference-cost efficiency for agentic software engineering.

Oracle added on-demand access to xAI's Grok 4.6 model in OCI Generative AI.

DeepSeek retired the experimental V4 Flash Vision model and routed it to V4.1 Flash.

DeepSeek retired V4 Flash and temporarily routed its model name to V4.1 Flash.

DeepSeek released a smaller native-multimodal V4.1 architecture model with upgraded coding, reasoning and agents.

Cohere released a compact sovereign open-weight machine-translation model optimized for speed and cost.

Suno released v6 Mini as a faster and more efficient generation option.

Suno released v6 Wild for exploratory and less predictable music generation.

Suno released v6, v6 Wild, and v6 Mini as its new family of music models.

OpenAI released GPT-Image-2.5 Sunburst and Flare as a new generation of image models.

Microsoft makes MAI-Image-2.6 and its Flash variant available in public preview through Microsoft Foundry.

Microsoft AI launches MAI-Transcribe-2 with automatic language identification and high-speed transcription across 60 languages.

Google DeepMind introduces WeatherNext 3, advancing probabilistic and high-resolution operational weather forecasting.

OpenAI released GPT-6 Astra, a frontier model positioned for its most demanding reasoning and agentic workloads.