AI Models, know what runs underneath

OpenAI's sixth-generation flagship model released 2026/9/3 (PT): billed as the world's most intelligent and aligned model, built for long-horizon agentic work and computer use; first model to reach the Critical cybersecurity threshold; 99.9% ARC-AGI-3, 97.6% FrontierMath Tier 4, 100% ExploitBench; API $10/$50

ByteDance flagship released 2026/6/23, rivaling GPT-5.5/Opus 4.7; 10+ hour autonomous engineering, 2-hour video understanding, 80% lower cost than Opus

A near-real-time voice conversation model released by Google DeepMind on 2026/9/15, with two variants: the standard version and Live Extended Thinking. The latter reasons while it speaks and targets complex multi-step tasks, ranking first on the Speech to Speech Quality Index at 82.6, scoring 68.6% on τ-Voice and 97.7% on Big Bench Audio. It supports mid-conversation switching across 97 languages, user interruptions, background tool calls and asynchronous long-running tasks, and embeds SynthID watermarks in its audio.

xAI's frontier model for coding, agentic tasks and knowledge work, 500K-token context, $2 input and $6 output per million tokens, AA-Briefcase score of 1657, with a fast variant limited to Cursor and Grok Build

Anthropic's first Claude 5.5 model, released 2026/9/22: matches Claude Fable 5.1 on most work; $4 per million input tokens and $20 per million output, cache reads down 60%, total operating cost about 40% below Opus 5 and output more than 30% faster

OpenAI light-tier GPT-6 model released 2026/9/22 (PT) for high-volume jobs with a clear goal such as summarization, extraction and short answers; API priced at $0.10 per million input tokens and $0.50 per million output, half of GPT-5.6 Luna, available to Free and Go users in the desktop app

OpenAI's mid-tier model launched 2026/10/27 at $1 per million input tokens and $5 output with $0.05 cached input, half the rates of GPT-6.1 Sol, built for computer use and reported to beat GPT-6 Astra on vision at about 47 percent the latency and 92 percent less cost

Google DeepMind's flagship model released 2026/10/28, two weeks after Argon, topping the AA Intelligence Index at 59 with a focus on cybersecurity and agents that run for days

Moonshot coding model released 2026/6/12, 1.1T MoE/32B active/256K context/native INT4; 3 coding benchmarks beat K2.6, 30% less token usage

Cursor self-developed coding model 2.5, based on Kimi K2.5 (1.04T MoE) + in-house RL, $0.10/task cheapest frontier coding model

Microsoft self-developed coding model, 2026/6 Build, MoE + Maia accelerator, strong on Rust/Haskell; replaces GPT-4 Turbo as Copilot default from 2026/8

Moonshot K3 post-train: 2.8T params, AA index 55, 2 dollars per M input tokens

Cohere open-source model released 5/22, 218B MoE, Apache 2.0, single B200, 48 languages, native citation

DeepSeek new-generation flagship model with improved reasoning, coding, and multilingual capabilities

Meituan's new large model preview launched on 2026/9/25, continuing the MoE architecture with native image understanding and deep adaptation to Claude Code

NVIDIA multimodal AI agent model, efficient inference for edge and robotics

FaceBrain open-source multimodal model, vision/audio/text, efficient on-device deployment

Ant Group large language model for enterprise and financial scenarios

DeepSeek's smallest new-architecture model, launched Sep 10, 2026 and refreshed to the 1017 checkpoint on Oct 17: 552B backbone plus 196B Engram (about 763B total weights), 8B/16B active params, native vision, 1M context, FP4 KV cache at ~890 bytes/token; the 1017 build scores 77.0 on DeepSWE v1.1, MIT-licensed with API input from 1 yuan per million tokens

Sakana AI orchestration model that routes across a swappable pool of open and specialized models; Fugu Max is priced at $2/$6 per million tokens

Xiaomi's native omni-modal open-weight flagship, a sparse MoE with 1.02T total and 42B active parameters, 1M-token context, top open-weight score of 46 on the Artificial Analysis Intelligence Index, MIT licensed

NVIDIA's open-weight hybrid model with 30B total and roughly 3B active parameters, up to 1M-token context and tool calling, built for high-frequency agent execution steps and local deployment

Flagship code and Agent model launched by Zhipu, with significantly enhanced post-training capabilities; weights are temporarily not open-sourced due to security hardening strategies

Google's new-generation model in the Flash series released on 2026/9/2 (the third Flash version within six weeks), focusing on long-horizon coding agents and professional workflows: DeepSWE v1.1 73.7%, Terminal-Bench 2.1 89.4%, Vals Finance Agent v2 61.4%; 1M context/64K output, intro price $0.75/$3.75 per million tokens; also launched Flash Cyber for defensive side at the same event

Tencent Hunyuan, a new generation flagship large model released and open-sourced on August 28, 2026 (Apache 2.0), with 770B total parameters, 49B activated, and 1M context, positioned as 'built for productivity'; internally, 163 expert engineering blind tests averaged 2.99/4, covering software engineering, office analysis, game development, and scientific research scenarios, with WorkBuddy, Yuanbao, and ima launched simultaneously

OpenAI's full-duplex speech model on the API (Sep 10, 2026): a single model that listens and speaks simultaneously, responds to interruptions in real time, and delegates deep reasoning and tool calls to backends like GPT-6 Astra, at $0.05/min for the voice layer

A mid-tier model released by Anthropic on September 28, 2026, with output speed more than 30% faster than Sonnet 5 and per-task cost reduced by up to 30%, while pricing remains unchanged.

A model released by OpenAI at DevDay on September 29, 2026, succeeding GPT-6 Sol, which had been online for only 7 days. Its capabilities are close to GPT-6 Astra, while its price is about one-fifth of it.

A frontier model released by Google DeepMind on September 30, 2026, the first in the Gemini 4 series; the output token limit has been increased from 64K to 1 million, with a focus on long-horizon software engineering, enterprise knowledge work such as law and finance, cybersecurity defense, and creative writing; it is first made available to trusted cyber defenders through the Fairwind program, and subsequently to paid API customers and Google AI Ultra subscribers

Anthropic released this small model on October 7, 2026, calling it the cheapest and fastest model it has shipped. Prompts up to 100K tokens cost $0.10 per million input and $0.50 per million output, rising to $0.50 and $2.50 beyond that, up to 90 percent below Haiku 4.5. It offers a 1M context window, up to 128K output tokens, and the first adjustable effort setting on a Haiku model, and targets high-volume work such as summarization, compaction, classification and database queries as well as subagent duty alongside Opus 5.5 and Sonnet 5.5

A multimodal MoE model released by MiniMax, with 428B total parameters, approximately 23B activated, a 1M context, focused on coding and agent workflows, priced at $0.15 per million input tokens

Inception Labs' diffusion-based large language model, claimed to achieve over 1000 tokens/s and 1M context, targeting low-latency, high-throughput scenarios

Open-source strongest coding and collaboration model, 2.4T parameters, first open-sourcing Qwen-Max level weights

Alibaba Tongyi Qianwen, open-sourced late at night on 2026/8/26, is a pioneer preview model based on the next-generation Qwen4 architecture: multimodal MoE, with a total main model parameter of 125B, activating only 6B per token, plus a 51B N-gram vocabulary layer, natively supporting 256K context expandable to 1M; training cost is nearly 90% lower than Qwen3.7-Plus, with weights open-sourced on Hugging Face and ModelScope

Alibaba's open-source 27.8B dense multimodal model, Apache 2.0 license, 262K context, Terminal-Bench 73.0, runs on consumer-grade GPUs

Open-weight 40B MoE coding model, 2.5B active, 262K context, runs on a laptop

33B dense agentic coding model with 512K context under Apache 2.0

Xiaomi MiMo text-to-speech model with natural voice synthesis

Flagship model released by Anthropic on 2026/9/1 (public version sharing the same foundation as Mythos 5.1), officially positioned as the world's most advanced model for programming and knowledge work: Terminal-Bench-Science 0.1 52.6%, CursorBench 73.4%, Terminal-Bench 4.0 55.8%; cache read price reduced by 75% to $0.25, typical load cost reduced by about 25%; 1M context/128K output

Anthropic's research-restricted flagship released on 2026/9/1 (sharing the same underlying model as Fable 5.1, with relaxed guardrails in dual-use domains), available only to cybersecurity and life sciences institutions reviewed under the Trusted Access Program; its network capabilities surpass the previous generation Mythos 5, with about half of 12 protein design targets hitting the mark; its capabilities also support Claude Security for enterprise customers

NVIDIA open-source 4B world model for robots and edge AI, real-time environment understanding and action generation

Xiaohongshu LLM series, dots-note 3.0 scored perfect 42/42 at IMO 2026; reads raw LaTeX end-to-end without formal languages

FaceBrain + OpenBMB open-source embodied AI model series, 1.5B VLA manipulation + 0.9B tracking

Meta's open-source 30B dense multimodal model, under Apache 2.0 license, designed for local agent scenarios, runs on a single consumer-grade GPU, supports image, text, and video input with tool calling

Six fully open model family (0.9B–375B-A23B) released by IFM (under MBZUAI) at the Abu Dhabi Foundation Model Institute on 2026/9/3, licensed under Apache 2.0 with public weights, training code, data ratios, and intermediate checkpoints; the flagship 375B-A23B activates about 23B per token, natively supports 512K context, and scores 70.2% on Terminal-Bench 2.1

Yutori released on 2026/8/26 a 27B general-purpose computer-use model: supports full desktop operations (Linux/macOS/Windows), no longer limited to browsers; OSWorld 2.0 65.2% (about $1.46/task), OSWorld-Verified 85.3%, MacAgentBench 83.1%; API input $0.5/million tokens, output $4

Reflection's first open-weight model (Oct 16, 2026): 501B MoE with 23B active, 1M-token context, text-only, Apache 2.0 weights in BF16/FP8/NVFP4, built for coding and agentic work; open-source Mirror CLI and 40+ day-0 partners, founded by ex-DeepMind researchers

Search-agent model family open-sourced by Xiaohongshu's AllSpark team on Sep 14, 2026: Iris-mini (35B, built on Qwen3.6-35B-A3B) and Iris-pro (397B, built on Qwen3.5-397B-A17B), both with 256K context; leads open models of comparable size across BrowseComp, BrowseComp-ZH, DeepSearchQA and HLE, with Iris-mini's 82.2 on BrowseComp approaching trillion-parameter models; trained with an SFT-to-RL climbing recipe, in-cluster judging and summarization, and resumable ultra-long trajectories; weights, paper and model cards are public

Scientific multimodal foundation model from Shanghai AI Laboratory, announced Sep 13, 2026 at the Pujiang Innovation Forum and released as a full open-source build on Hugging Face on Sep 14: 397B parameters focused on long-horizon research and agentic ability, with general knowledge, coding and agent skills in the top tier of open models; a pluggable Memory Decoder adds domain knowledge without changing base weights; scores 87.00 on FrontierScience-Olympiad, 93.56 on HMMT-2026 and 68.54 on SWE-bench-Pro; deeply co-optimized with the Ascend compute ecosystem and set to plug into the Shusheng Duanyan scientific discovery platform

Natively omnimodal model released by Alibaba Qwen on 2026/9/17 with text, image, audio and video input and a 1M token context; Alibaba reports the average score across 29 benchmarks is more than 25% above Qwen3.5-Omni-Plus while hourly audio input pricing falls over 98% and audio-video input over 93%, aimed at audio and video agent delivery

Lightweight open-source translation model Google released on 2026/9/15, trained from Gemini, supporting 55 languages and able to run offline; Google says its language technology now covers more than 300 languages reaching roughly 86% of the world population

Qwen's real-time speech interpretation model released 2026/9/18, built on a single-stream Interleave architecture with a Hybrid-MoE Thinker-Talker design, cutting average interpretation lag (LAAL) from 2.8 to 2.3 seconds; accepts text translation input in 60 languages with spoken output in 29, with built-in real-time speaker diarization, bilingual source-translation alignment and long-context disambiguation; served via Alibaba Cloud Bailian's realtime API

TypeSafe AI's large model built for high-frequency decisions: no dialogue or prose generation, only typed decisions with calibrated confidence, trained with the company's RLCD method; by its own workflow benchmarks up to 193.6x faster and 444.6x cheaper than conventional large models; access by application on the official site, with Vercel as the launch integration

The efficiency tier of the MiMo-V2.6 series, a sparse MoE with 309B total and 15B active parameters, 1M-token context, native omni-modal, released and open-sourced alongside Pro under MIT

The high-speed serving tier of Xiaomi MiMo-V2.6-Pro, sharing the same checkpoint while advertising up to 20x output speed at roughly ten times the standard price, built for real-time interaction and latency-sensitive production workloads

A family of general-purpose computer-use agent models released by H Company on September 28, 2026, including two sizes, 27B and 35B-A3B, capable of completing tasks across desktop, web, mobile, and API.

A MoE large model released by Ant Bailing (InclusionAI) on September 30, 2026, with approximately 560B total parameters and about 25B activated per token, and a maximum context window of 1M; continuously optimized around general-purpose agents, search, daily office work, and software development, with enhanced capabilities in scenarios such as healthcare, finance, and materials science; a two-week free trial is available (the service length during the trial period is 256K), after which it becomes paid and is planned to be open-sourced

A decision model from Liquid AI that takes text or images plus questions and returns a probability for each answer in one forward pass without generating tokens, with text decisions in 200 to 300 ms and billing on input tokens only

Mistral released this frontier model in Research Public Preview on October 6, 2026. It has 1T total parameters with 49B active, native multimodality, a 512K context window and up to 256K output tokens. It scores 38 on the Artificial Analysis Intelligence Index, level with GPT-6 Luna at max effort and the highest score from outside the US and China, plus 50 on the Cyber Index, with 1T weights planned for release at the end of October

Google DeepMind released this multimodal embedding model in October 2026. Built on the Gemma 4 architecture with 740M parameters and open weights under Apache 2.0, it maps text, code, images, video frames and audio into a single embedding space, so cross-modal queries such as searching video with words work directly, at a size that still fits on-device and in-browser deployment

Microsoft released this decision model on October 9, 2026. Post-trained from Qwen3.5-9B, it turns text and images into structured answers such as probabilities, options and scores, and is available on Microsoft Foundry and OpenRouter. Microsoft says it records the highest accuracy across about 36 blind benchmarks covering roughly 150,000 questions and runs 4.5x faster than the runner-up, at $0.042 per million input tokens with free output and a 32K context.

Cloudflare released this open-weight multimodal decision model on October 9, 2026. A 30B-total / 3B-active MoE post-trained from Qwen3-Omni-30B-A3B-Instruct, it scores text, images, audio and video in a single call and returns a calibrated probability for every option rather than generating text, supporting true/false, multiple choice and ordinal score questions. Weights ship under Apache-2.0; on Workers AI it runs as @cf/cloudflare/clef-omni with a 64K context, about 130 ms median latency for text-only decisions, $0.15 per million input tokens and no output charge.

Nace AI open-sourced this 8.95B dense decision model. It reads a text or JSON state plus typed questions and returns a probability for every option without generating any text, supporting choice, yes/no and ordinal score questions over a POST /v1/systemone endpoint. It scores 58.08 on the public Decision Index 0.3.1, inside the tie band of the closed-source Jev 1.13.0 and the highest result for any model under 10B parameters, with 93.4% accuracy on 32K to 128K token documents at 2.0 s median latency. Weights are on Hugging Face, a 9.5 GB Q8_0 variant runs on Apple silicon and CPU, and hosting on OpenRouter costs $0.04 per million input tokens.

Meta flagship model line released on 2026/10/12. Muse Wonder and Muse Wonder Full score 52 and 56 on the Artificial Analysis Intelligence Index, with 1M context, built for personal and professional agent use.

Open-weight model from Liquid AI released on 2026/10/13, 24B total parameters with 2B active per token, 131K context, reasoning first, built for fast private agents on a laptop or your own hardware

30B-A4B MoE trained from scratch in Korea, 256K context, Apache 2.0

Liquid AI's 2026/10/27 release of six open-weight 2B-class models, all branched from one tri-hybrid backbone and covering text, vision, audio, omni and decision work, with the text model at 1.73B parameters, a 256K context and about 1GB of memory at 4-bit

Google open-source multimodal model family April 2026, 2.3B to large, strong reasoning and Agent ability

Meta multimodal model for smart glasses and wearables, powering Meta Glasses 2026

SenseTime daily-refresh model (RiRiXin) for enterprise multimodal AI applications