New AI Models

GPT-5.6 Terra

GPT-5.6 Terra is OpenAI's cost-balanced model in the GPT-5.6 family, designed for workloads that weigh intelligence against cost and roughly corresponding to the mini tier of earlier GPT-5 families. It provides a 1,050,000-token context window, 128,000-token max output, text and image input, and reasoning effort settings from none to max, at $2 per million input tokens and $12 per million output tokens. Terra is available via the OpenAI API, serves as the recommended replacement for GPT-5.4 in Codex, and works with site tools in the ChatGPT desktop app.

Wan2.7-Reference-To-Video-2026-06-12

Wan2.7-Reference-To-Video is the reference-to-video model in Alibaba's Tongyi Wanxiang 2.7 series. It accepts up to 5 mixed image and video references plus an optional voice reference, and outputs 720P or 1080P clips of 2 to 15 seconds with more stable character, prop, and scene consistency. It is suited to single-character performance and multi-character interaction. This 2026-06-12 snapshot is pinned for stable production use.

Wan2.7-Text-To-Video-2026-06-12

Wan2.7-Text-To-Video-2026-06-12 is the June 12, 2026 snapshot of Alibaba's Wan2.7 text-to-video model, a fixed version suited to stable production use. It generates fluid video from text prompts with upgraded acting: nuanced emotional scenes, hard-hitting action, and more dramatic, well-paced shot transitions. Output covers 720P or 1080P at 2 to 15 seconds, with audio uploaded or auto-generated to match the visuals.

Claude Sonnet 5

Released on June 30, 2026, Claude Sonnet 5 is Anthropic's balanced-tier model and its most agentic Sonnet yet, pairing top-tier intelligence with coding and everyday professional work. It improves on Sonnet 4.6 across reasoning, tool use, coding, and knowledge work, and approaches Opus 4.8 performance at lower cost on agentic evals like BrowseComp and OSWorld-Verified. It suits multi-step software engineering, autonomous computer use, and business automation.

Gemini Omni Flash

Gemini Omni Flash is Google's high-performance multimodal model for fast video generation and conversational video editing, introduced in public preview on June 30, 2026. It natively processes text, image, audio, and video, and uses the Interactions API to turn text prompts or still images into videos of 3 to 10 seconds with audio. Results can be refined through natural language conversation, with support for video extension, resolution upscaling, and interpolation.

Nano Banana 2 Lite

Nano Banana 2 Lite (model code gemini-3.1-flash-lite-image) is the efficiency tier of Google's image generation family, released to GA on June 30, 2026. It targets sub-2 second end-to-end latency and significantly reduced compute costs, making it suited to high-volume interactive and real-time consumer applications. The model supports interleaved text and image generation, is optimized for 1K (1024x1024px) output, and handles fast multi-turn local edits such as color swaps, stickers, and background adjustments, with SynthID watermarking always on.

LongCat-2.0

LongCat-2.0 is Meituan's open-source MoE large language model with 1.6 trillion total parameters and about 48 billion activated per token. It introduces LongCat Sparse Attention and an N-gram Embedding module, pretrained on over 35 trillion tokens across 50,000+ domestic chips, with native million-token context. Scoring 70.8 on Terminal-Bench 2.1 and 59.5 on SWE-bench Pro against frontier closed models, it targets coding, agent, and long-context workloads.

Leanstral 1.5

Leanstral 1.5 is Mistral AI's updated Lean 4 formal proof engineering model, an Apache-2.0 open weights upgrade of the original Leanstral for automated theorem proving and autoformalization, with 119B total and only about 6B active parameters. It saturates miniF2F (100%), solves 587 of 672 PutnamBench problems, and sets a new state of the art on FATE-H (87%) and FATE-X (34%). Available free as the leanstral-1-5 API endpoint and on Hugging Face, it targets formal mathematics, proof engineering, and code verification, including uncovering real bugs in Rust repositories.

grok-4.5

Grok 4.5 is xAI's smartest model to date, built to excel at coding, agentic tasks, and knowledge work. It is served at fast-model speeds of 80 TPS, delivers roughly 2x the token efficiency of comparable leading models, and is priced at $2 per million input tokens and $6 per million output tokens. It is available today in Grok Build, in Cursor on all plans, and through the xAI API, and handles Office work in Excel, PowerPoint, and Word with similar care.

Kling 3.0

Kling 3.0 is the standard video generation model from Kuaishou's Kling AI, built on a unified multimodal training framework that consolidates text-to-video, image-to-video and reference-based generation into one model. It generates clips up to 15 seconds with flexible duration, adds Multi-Shot storytelling and Element consistency control, and outputs Native Audio in Chinese, English, Japanese, Korean and Spanish, including dialects and accents, alongside improved prompt adherence and image realism. It suits multi-shot cinematic storytelling, multi-character dialogue scenes and e-commerce advertising.

Kling 3.0 Omni

Kling 3.0 Omni is the all-in-one video generation model in Kuaishou's Kling 3.0 family, upgrading Kling VIDEO O1 with a deeply unified multimodal training framework for native audio and video output. It supports Native Audio and Multi-Shot storytelling, generates 3 to 15 second videos with shot-level control over duration, framing, perspective, and camera movement, and its Element system accepts video and voice references to keep characters consistent in both look and sound, fitting narrative work such as ads and short dramas.

Kling Image O1

Kling Image O1 is a reasoning image generation model that draws on a broad knowledge base and multimodal reasoning to interpret complex creative intent. It supports up to 10 reference images, keeping subject contours, core elements, and tonal qualities consistent across a series. Text instructions handle detail edits and style transfer while preserving original lighting and texture. Internal evaluations showed higher win rates than Nano Banana, Flux 1.0 Kontext Pro, and Seedream 4.0 on instruction-transformation and multi-image reference tasks, suited to IP character design and brand visuals.

GPT-5.6 Sol

GPT-5.6 Sol is OpenAI's flagship model, launched in limited preview in June 2026 alongside the balanced Terra and fast, affordable Luna tiers. It adds a new max reasoning effort and an ultra mode that leverages subagents, sets a new state of the art on Terminal-Bench 2.1, and outperforms GPT-5.5 on GeneBench v1 with fewer tokens while matching Mythos Preview on ExploitBench with roughly a third of the output tokens. Preview access runs via the API and Codex for select partners, priced at $5 input and $30 output per 1M tokens.

Qwen-Image-2.0-Pro-2026-06-22

Qwen-Image-2.0-Pro-2026-06-22 is the latest dated snapshot of the full-strength Qwen-Image-2.0 Pro image model, released on Alibaba Cloud Model Studio on June 25, 2026. It unifies image generation and image editing in one model. Compared with the April snapshot, it renders text more accurately, accepts prompts up to 1k tokens, and delivers finer photorealistic detail with stronger semantic adherence. The pinned version number suits production workflows that need reproducible outputs.

Seed2.1

Seed2.1 is ByteDance's next-generation agentic model for real-world productivity, released in two sizes, Pro and Turbo. It brings major improvements to general agents and code engineering, covering end-to-end requirement understanding, coding, debugging, and validation in real development workflows. It also strengthens visual understanding, spatial reasoning, and long-context processing, and achieves SOTA results across multiple video understanding benchmarks.

Mistral OCR 4

Mistral OCR 4 is Mistral AI's document-understanding OCR model and the successor to Mistral OCR 3. Beyond extracted text it returns bounding boxes, typed-block classification (titles, tables, equations, signatures), and inline confidence scores across 170 languages. Mistral reports a 72% average human-preference win rate over leading OCR systems, 85.20 on OlmOCRBench, and 93.07 on OmniDocBench. It ships via API, Mistral Studio, Amazon SageMaker, and Microsoft Foundry, with a single-container self-hosted option, suiting RAG ingestion, enterprise search, and document pipelines.

GPT-5.5 Instant

GPT-5.5 Instant is OpenAI's updated default model for ChatGPT, rolling out to all users and available in the API as chat-latest. In internal evaluations it produced 52.5% fewer hallucinated claims than GPT-5.3 Instant on high-stakes prompts in medicine, law, and finance, with gains across visual reasoning, math, and science. Answers are tighter and less overformatted, and the model makes better use of context from past chats, files, and connected Gmail to personalize responses.

Kling 3.0 Turbo

Kling 3.0 Turbo is the speed- and cost-focused tier of the Kling 3.0 video model family from Kuaishou, released alongside upgraded video editing in Kling 3.0 Omni. It supports text-to-video and image-to-video with first-frame input, generating 720P or 1080P clips of 3 to 15 seconds, with prompt-driven multi-shot storytelling of up to 6 shots and native audio output. Official docs position it as faster output at lower cost with reliable audio-visual sync, suited to high-volume video production.

Fun-ASR-Flash-2026-06-15

Fun-ASR-Flash-2026-06-15 is the dated June 2026 snapshot of the Bailing large-model ASR line on Alibaba Cloud Model Studio, a fit for production workloads that need consistent output. It transcribes audio up to 5 minutes across 30 languages, supports all seven major Chinese dialect groups, and refines punctuation prediction and text normalization so numbers, dates, and amounts follow standard written formats. Recognition of classical Chinese poetry is specifically optimized, and usage is billed per second of audio from USD 0.00003.

GLM-5.2

GLM-5.2 is Zhipu AI's flagship model for long-horizon tasks, open-sourced under an MIT license with a solid 1M token context window. An IndexShare sparse-attention design trims per-token FLOPs by 2.9x at full context, and coding strength takes a major step up over GLM-5.1: 81.0 on Terminal-Bench 2.1 and 62.1 on SWE-bench Pro, the best among open-source models and second only to Claude Opus 4.8 on FrontierSWE and PostTrainBench. Adjustable thinking effort levels balance performance against latency for coding agents and long-running engineering work.