AI Models Leaderboard

Specifications and live ecosystem intelligence.

Unrestricted frontier reasoning, Glasswing partners only

Anthropic has launched Claude Opus 5.5, a new AI model featuring stronger safeguards against risky behaviors like sandbox escape attempts, following recent rogue AI hacking incidents.

OpenAI is expanding its Daybreak cybersecurity initiative and introducing new specialized models, including GPT-5.6-Cyber and GPT-5.6 Sol, designed for authorized defensive security work, vulnerability discovery, and exploit validation.

1M token context introduction, native Agent Teams support

Agentic workflow driver, near-Opus coding efficiency

Chinese artificial intelligence developer MiniMax Group Inc. has released M3.1 Flash, a faster, cheaper update to its M3 large language model designed to significantly reduce the operating costs of AI agents.

Leading open-weights video foundation model generating fluid motion and natural lighting in text-to-video and image-to-video modalities.

Open-weights mixture-of-experts reasoning model developed by DeepSeek utilizing large-scale reinforcement learning for algorithmic code synthesis.

Lightweight multimodal model by OpenAI delivering fast token throughput and affordable inference for structured code generation and unit test drafting.

Cost-efficient reasoning model by OpenAI offering tiered reasoning effort parameters optimized for competitive programming and STEM tasks.

Flagship 70B parameter open-weights model by Meta AI offering state-of-the-art instruction following and competitive programming capabilities.

Edge-optimized lightweight text model for local deployment

OpenAI has launched GPT-Live-1 in its API, offering developers access to its full-duplex voice model at $0.05 per minute for the voice layer. The model supports simultaneous listening and speaking while delegating reasoning to paired tools.

Low-cost high-speed execution, automated model switching

A novel hybrid SSM-Transformer architecture designed for infinite context.

Anthropic multimodal foundation model demonstrating strong visual reasoning, long-form video question answering, and advanced coding.

A foundational dual-text-encoder latent diffusion model operating at native 1024x1024 resolution with extensive fine-tuning support.

A multi-resolution diffusion transformer featuring bilingual English and Chinese text encoders for fine-grained concept comprehension.

xAI's flagship frontier reasoning model trained on the Memphis Colossus supercluster, delivering competitive results in STEM and coding.

Google DeepMind high-definition video generation model emphasizing prompt semantic adherence, visual consistency, and cinematic controls.

Microsoft's compact 3.8B parameter edge model featuring native function calling and strong mathematical reasoning.

NVIDIA 14-billion parameter world foundation model generating 120 future video frames conditioned on visual and textual states for physical AI.

The ~3B parameter on-device model powering iOS 18.

Pretrained on 18T tokens, strong multilingual reasoning

NVIDIA and the UW Medicine Institute for Protein Design debuted the open protein-design model RFD4-Proteina at GTC Berlin.

A highly advanced text-to-image open model with a Multimodal Diffusion Transformer (MMDiT) architecture.

The model that powered the initial viral launch of ChatGPT.

NVIDIA released Cosmos3-H-Dreams, a real-time action-conditioned generative world model for robotic surgery fine-tuned on clinical cholecystectomy recordings from the CMR Surgical Versius system.

Flagship 123B parameter foundation model by Mistral AI optimized for advanced multi-file reasoning, test generation, and tool use.

Alibaba Cloud flagship open-weights vision-language model featuring dynamic-resolution ViT and 3D RoPE for hour-scale video understanding.

Shanghai AI Lab open-source multimodal model coupling a 6B vision encoder with a 72B language backbone for leading Video-MME and MVBench accuracy.

OpenAI is expanding its Daybreak cybersecurity initiative and introducing new specialized models, including GPT-5.6-Cyber and GPT-5.6 Sol, designed for authorized defensive security work, vulnerability discovery, and exploit validation.

Google high-speed multimodal reasoning model engineered for real-time video comprehension, live perception, and high-throughput QA.

A 12-billion parameter step-distilled rectified flow model engineered for 1 to 4-step rapid image generation under an open-source license.

A text-to-image model featuring dense caption following, robust typography generation, and native integration with ChatGPT and OpenAI API.

A unified vision foundation model using a prompt-based sequence-to-sequence structure for captioning, object detection, and visual segmentation.

A closed-source flagship text-to-image rectified flow transformer hosted via dedicated API endpoints for enterprise-grade generative synthesis.

Google flagship text-to-image diffusion model delivering photorealistic rendering, fine detail, and SynthID watermarking.

ByteDance Seed video diffusion model delivering high-aesthetic 720p text-to-video synthesis with integrated synchronized soundtrack generation.

Frontier reasoning model by OpenAI utilizing internal reinforcement learning chains of thought for competitive programming and complex algorithmic synthesis.

A 1.6-billion parameter linear diffusion transformer by NVIDIA synthesizing 4K resolution images with high operational efficiency.

A 6-billion parameter diffusion transformer utilizing GLM-4 text representations for high-resolution bilingual image generation.

ByteDance and LLaVA team open-source multimodal model trained on curated video instruction suites for dense event narration and QA.

14-billion parameter diffusion transformer tailored for high-resolution image-to-video generation at 720p with precise motion trajectory control.

104B parameter model tuned for enterprise RAG and tool use

Compact 14B parameter small language model by Microsoft trained on synthetic and curated data achieving top-tier mathematical and algorithmic reasoning.

Mid-tier open code model by Alibaba Cloud striking a balance between memory footprint, inference throughput, and multi-file code editing.

Alibaba Cloud ATH high-aesthetic text-to-video generative model leading silent visual preference pools in independent testing.

High-parameter general foundation model by Alibaba Cloud exhibiting robust multilingual coding, tool orchestration, and algorithmic execution.

StepFun 30-billion parameter open-weights diffusion model capable of generating up to 204 frames with deep 16x16 spatial and 8x temporal VAE compression.

Flagship 14-billion parameter open-weights diffusion transformer model by Alibaba supporting 720p text-to-video and bilingual visual text rendering.

The widely deployed 512x512 latent diffusion model based on the CompVis architecture, extensively utilized in open-weights workflows.

Initial public preview release of OpenAI test-time reasoning architecture evaluated across competitive coding and repository bug resolution benchmarks.

Experimental open reasoning model by Alibaba Cloud applying test-time reasoning steps to resolve complex coding and mathematics benchmarks.

The initial 2-billion parameter multimodal diffusion transformer checkpoint released by Stability AI featuring dual text encoders.

Lightweight 3.8B parameter model by Microsoft featuring a 128K context window optimized for memory-constrained on-device code generation.

Frontier testbed architecture for high-complexity operations

Ultra-compact 1.5B code model by 01.AI supporting repository-level context retrieval and efficient local developer terminal assistant workflows.

Compact dense 6.7B parameter open-weights model designed for developer workstations and local interactive code completion pipelines.

A latent diffusion model based on the ChatGLM language backbone supporting bilingual English and Chinese text-to-image synthesis.

OpenAI flagship diffusion transformer model generating photorealistic video clips with complex physics and cinematic lighting.

Google enterprise text-to-image diffusion model supporting high-resolution image generation, inpainting, and outpainting in Vertex AI.

An open diffusion transformer architecture supporting multi-resolution generation with flexible aspect ratios and high text alignment.

Commercial creative video foundation model offering specialized multi-shot camera work, character performance continuity, and native sound effects.

Frontier-class 405B dense model by Meta AI designed for large-scale synthetic data generation, code model distillation, and complex system engineering.

xAI generative video engine providing high aesthetic preference rankings and physical dynamics inside the Grok platform ecosystem.

Commercial text-to-video diffusion checkpoint engineered by MiniMax for high cinematic realism and character physical motion.

Community-governed 15B code foundation model developed by BigCode trained on 4T tokens from Software Heritage The Stack v2 across 600+ languages.

Compact 8B parameter model by Meta AI offering high-efficiency inference for routine programming syntax transformations and script generation.

Post-trained optimization of MiniMax H3 by fal.ai delivering enhanced prompt fidelity and lower API per-minute costs.

Balanced 7B parameter open-access model by BigCode trained with Grouped Query Attention and fill-in-the-middle support.

An open-weights rectified flow transformer architecture developed by fal.ai for high-fidelity text-to-image generation.

Mid-tier enterprise code model by IBM trained across 116 programming languages under an Apache 2.0 license for corporate environments.

Sakana AI has announced the launch of Fugu Max and Fugu Ultra v2, two models built on a learned orchestration architecture designed for cost efficiency and peak capability.

Lightricks low-latency real-time video latent diffusion model producing 5-second 1080p clips in under 28 seconds.

Initial budget-tier model in the Claude 3 generation

Peking University and YuanGroup 8.5B parameter sparse DiT video foundation model employing WFVAE 8x8x8 temporal-spatial downsampling.

ShengShu Tech foundation video generation engine utilizing Universal Visual Model architectures for dynamic scene consistency.

ByteDance Seed high-throughput video foundation model balancing aesthetic fidelity with reduced generation latency.

Open-source Mixture-of-Experts video diffusion architecture offering enhanced cinematic motion scaling trained on expanded multimodal corpora.

Lightricks flagship foundation model checkpoint engineered for high-fidelity 1080p generation with enhanced prompt adherence.

NVIDIA has released Kumo Tabular, a new family of tabular foundation models designed for classification and regression tasks without the need for traditional training, feature engineering, or hyperparameter tuning.

Nums AI has released Causilo, a pretrained tabular foundation model for classification and regression that features a scikit-learn interface and achieves the top TabArena Elo among single models.

Illumina has introduced SpliceAI2, a genomic AI model developed by its BioInsight AI Lab designed to predict how genetic variants alter RNA splicing. In testing, the model identified 17% more disease-relevant variants compared to other splicing models.

Google's anticipated Gemini 3.5 Pro model has failed to launch after a three-month delay and a recent leadership change, leaving its release status uncertain.

A highly efficient edge model with a massive 128k context window.

The smaller variant of Gemma 2 optimized for local execution.

A revolutionary model trained to spend time thinking before responding.

A robust, open-source automatic speech recognition system.

An open general-purpose LLM created by Databricks to power enterprise custom models.

A highly performant open-weights model built from Gemini technology.

An open-source Mixture-of-Experts model that outperforms proprietary coding models.

A faster

Microsoft's hyper-optimized 3.8B parameter edge model.

A generative AI music model capable of creating full, radio-quality songs up to 4 minutes long.

Offering a 128k context window and multilingual support, the 70B variant is widely deployed by enterprises seeking high performance without the massive compute overhead of the 405B model.

A high-fidelity AI music generator focused on crisp audio and structural control.

A model capable of creating a synthetic voice from just a 15-second audio sample.

An earlier Anthropic model pioneering the 200k context window.