New AI models you may not have heard of: 43 models from 22 vendors

The first reason is price. Flagships from OpenAI, Anthropic and Google cost $5 to $30 per million output tokens, while many models in this article cost under a dollar. For classification, translation, data extraction and drafts the quality gap is often invisible, while the bill differs several times over.

The second is speed and specialisation. The diffusion model Mercury 2.5 outputs hundreds of tokens per second, Cohere's Command A+ is built for the strict schemas of enterprise agents, ByteDance's Seed 2.0 Code for frontend work. The third is openness: the weights of MiMo V2.6 Flash, Nemotron 3, Inkling, Laguna S 2.1 and Trinity are published, so what you tested through the API you can later host yourself.

And these models are already used at scale. According to OpenRouter for the week starting October 13, the five most used models by token volume were DeepSeek V4.1 Flash, GLM 5.3 Flash (7.6% of all tokens), MiMo V2.6 Flash (6.1%), Step 5 Preview (5.2%) and Hy4 Preview (5.0%). Not one from OpenAI, Anthropic or Google; Claude Opus 5.5 is only number 7.

MiMo V2.6 Pro is Xiaomi's largest model, over a trillion parameters, built for heavy agent workloads. It natively understands text, images, video and audio, holds about a million tokens of context and calls tools. Its Artificial Analysis index is 46.3, only 1.3 points below GPT-6 Sol, while its output is 5.4 times cheaper: $0.5 / $1 per million tokens. Terminal-Bench 4.0: 34.8%; SciCode: 60.9%.

MiMo V2.6 Flash is an open model with 309B parameters, 15B active per token. Same modalities and context as Pro, but cheaper: $0.2 / $0.3. Index 37.9, and it was OpenRouter's third most popular model last week. API IDs: mimo-v2-6-pro and mimo-v2-6-flash.

Nemotron 3 Ultra is NVIDIA's largest open model: a Transformer–Mamba hybrid mixture-of-experts with 550B parameters, 55B active. It was built for agent pipelines with an emphasis on planning — orchestrating other agents, coding agents and deep research. Text only; on Kumo a 256K-token context and tool calling, at $0.7 / $2.8.

Nemotron 3 Super is the smaller sibling with 120B parameters and 12B active, predicting several tokens at once for faster generation. NVIDIA published not only the weights but the training data and recipes. At $0.09 / $0.5 it is one of the cheapest catalogue models for multi-step agents. Our snapshot has no independent Artificial Analysis scores for either Nemotron yet.

Step 5 Preview is the preview of Shanghai-based StepFun's flagship agent model: 600B parameters, 27B active, a 1M-token context. It targets software development and knowledge work, especially finance, where it iterates with tools over large codebases and document sets. In the Artificial Analysis agent ranking, Step 5 Preview running inside Claude Code scored 51.2 — more than Grok Build with Grok 4.6. Kumo price: $1.2 / $3.1; on OpenRouter it was the fourth most popular model of the week.

Step 3.7 Flash is an economical multimodal model: about 11B active parameters per token, with the reasoning level chosen per request. Index 19.5, price $0.2 / $1.3.

Yes, the Meituan that delivers food in China. LongCat 2.0 is a sparse model with 1.6T parameters, 48B active, built for coding, repository-scale changes and long agent sessions. Context 1M tokens, price $0.3 / $1.4. Its Artificial Analysis index (19.1) is modest next to the leaders, but by parameter count it is one of the largest models in the catalogue.

Mercury 2.5 works differently from every other model here. A normal language model writes one token at a time; a diffusion model sketches many tokens at once and refines them in parallel — roughly how image models develop a picture out of noise. The result: 706 tokens per second measured by Artificial Analysis, the highest speed in the Kumo catalogue and several times that of the flagships.

It supports reasoning, parallel tool calls and schema-bound JSON, and costs $0.05 / $0.2. Do not expect deep reasoning — its index is 12.3 — but for search agents, voice interfaces and fast subagents, where every second of latency shows, it is an excellent candidate.

Canada's Cohere builds models for business, and Command A+ is its flagship for enterprise agents. Its strength is discipline: native tool calling with strict schemas, JSON-schema output, configurable safety modes and optional reasoning. Context 188K tokens, speed 166 tokens per second, price $0.3 / $1.7.

Korea's Upstage built Solar Pro 4 for long office tasks: large document volumes, coding, agent scenarios. Context 512K tokens at $0.1 / $0.4 — one of the cheapest catalogue models with a window that large.

Thinking Machines Lab is the company of Mira Murati, OpenAI's former CTO, and Inkling is its open-weights model: 975B parameters, 41B active. A general model for reasoning, coding, agents and retrieval that natively understands images and audio. Context 512K tokens, speed 221 tokens per second, index 25.0, price $1.2 / $4.7.

Arcee AI's Trinity Large Thinking is an open reasoning model for agent workloads and the second fastest in the catalogue: 374 tokens per second. It works best when its earlier reasoning is passed back in later turns. Price $0.3 / $0.9.

Poolside's Laguna S 2.1 is built specifically for coding agents: 118B parameters, 8B active, a 1M-token context, open weights. At $0.1 / $0.2 it is one of the cheapest coding models with a million-token window.

Tencent's Hy4 Preview is a text model with 770B parameters, 49B active, for coding agents and tool-heavy scenarios. Context 1M tokens, price $1 / $2.9; on OpenRouter it was fifth by popularity for the week.

ByteDance Seed has two models in the catalogue. Seed 2.0 Code is tuned for agentic coding with a frontend focus and meant to run inside agents such as Kilo and OpenCode; Seed 2.1 Turbo can carry development from plan to result, debugging its own mistakes. Both have a 256K-token context; prices $0.6 / $3.5 and $0.6 / $2.9.

AionLabs' Aion 3.5 is a system for role-play and fiction: several specialised GLM-based models compose each reply together for a better-built plot. IBM's Granite 4.2 8B is a small 8B model with three reasoning modes and 12 languages, at $0.07 / $0.3. Amazon's Nova 2 Lite is an inexpensive reasoning model for extracting data from documents and video, answers with citations and automation; 1M-token context.

These vendors are well known, but their lineups hold strong models people often skip. Zhipu's GLM-5.3 is the best model outside the big four on Terminal-Bench 4.0 (41.9%) at $0.8 / $2.6, and GLM 5.3 Flash is OpenRouter's second most popular model. Moonshot AI's Kimi K3 is a flagship with an index of 43.6. DeepSeek V4.1 Flash costs $0.1 / $0.5 with an index of 39.5 — more in the DeepSeek API article.

Alibaba's Qwen has five models — from the 2.4T-parameter Qwen3.8 Max to Qwen3 Coder Next with only 3B parameters active per token. MiniMax M3 handles a million tokens of context more cheaply than its predecessors thanks to sparse attention. Mistral Large 4 is a European alternative with a million-token context and replies up to 256K. Meta is represented by Llama 4 Maverick and the compact open-weights Muse Glimmer 30B.

The chart shows new-vendor models (blue) and, for reference, a few well-known ones (grey). The higher the dot, the smarter the model by the composite index; the further left, the cheaper its output tokens on Kumo.

Two Xiaomi models stand out: MiMo V2.6 Pro delivers close to GPT-6 Sol level for $1 output, and MiMo V2.6 Flash an index of 37.9 for $0.3. Zhipu's GLM 5.3 Flash is another dot in the top-left corner. The other lesser-known models score lower, and you should pick them not for overall level but for a specific property: speed, long context, strict JSON or open weights.

Every model from vendors beyond OpenAI, Anthropic, Google and xAI that Kumo carries today. A dash in the index column means Artificial Analysis has not measured the model yet.

On Claude, GPT and Gemini models Kumo charges below the vendor price list. On most models in this article prices are currently at market level — the same as the vendors and the large platforms, without a discount. We are negotiating lower rates with suppliers and will cut prices as agreements land; the current rate of any model is always visible in advance on the models page.

Even at market prices many of these models cost several times less than the flagships, and you pay from the same balance as for Claude or GPT: by card, SBP or crypto.

Every catalogue model is available on one Kumo key through an OpenAI-compatible API: to compare them on your own task, change the model name. The script below sends the same prompt to four models and prints each answer and its time.

The same key works in Cursor, Cline, OpenCode and any other tool with an "OpenAI Compatible" setting: the address https://api.kumorouter.com/v1 and a model ID from the table above. You can also compare on the models page, and independent measurements for every model are on the benchmarks page.

Model descriptions follow the Kumo catalogue and vendor materials. Scores and speed: Artificial Analysis, snapshot of October 20, 2026. Popularity: the OpenRouter weekly ranking for the week starting October 13. Prices: Kumo price list version 33 on the models page.