What we added — and why it matters
TokenRate just tripled its provider coverage. The catalogue went from 11 providers to 29 — eighteen new model makers added in a single update, lifting the number of live, daily-priced models past 260. Every one of them now has a full deep-dive page, pricing pulled from the same daily feed as everything else, and a slot in the side-by-side price comparison tool.
The point isn't size for its own sake. The frontier labs — Anthropic, OpenAI, Google — get the headlines, but a large and growing share of real production value sits with open-weight and specialist models that cost a fraction as much. You can't pick the cheapest model that clears your quality bar if it was never on the table. This update puts eighteen more makers on the table. Browse them all in the providers directory, or read on for what each one brings.
Every provider named below links to its own TokenRate page, where you'll find the full model lineup, current input and output prices per million tokens, context windows, and tiers. The official site for each is linked in the Primary sources block at the foot of this article.
The point isn't size for its own sake. The frontier labs — Anthropic, OpenAI, Google — get the headlines, but a large and growing share of real production value sits with open-weight and specialist models that cost a fraction as much. You can't pick the cheapest model that clears your quality bar if it was never on the table. This update puts eighteen more makers on the table. Browse them all in the providers directory, or read on for what each one brings.
Every provider named below links to its own TokenRate page, where you'll find the full model lineup, current input and output prices per million tokens, context windows, and tiers. The official site for each is linked in the Primary sources block at the foot of this article.
The open-weight wave from China
Five of the eighteen are Chinese labs shipping open-weight models that punch well above their price class. Zhipu AI — a Tsinghua University spin-out, also known as Z.ai — builds the GLM series (GLM-4.5, 4.6, 4.7 and GLM-5): strong agentic and coding models with vision variants that are popular for self-hosting at a fraction of frontier API prices. Moonshot AI makes Kimi, whose K2 release is a large Mixture-of-Experts model tuned for tool use, coding, and very long context — one of the leading open alternatives to closed frontier models.
MiniMax, out of Shanghai, pairs very long context windows — north of a million tokens — with aggressive pricing and solid agentic performance. Tencent brings the Hunyuan family, including Mixture-of-Experts variants that power its own products and are available open-weight for general chat, reasoning, and multilingual work. And Baidu adds ERNIE 4.5, China's longest-running large-model line, now with big MoE and vision-language variants. If you've wondered whether the open Chinese models really are as cheap as people say, you can now check the live numbers directly instead of taking anyone's word for it.
MiniMax, out of Shanghai, pairs very long context windows — north of a million tokens — with aggressive pricing and solid agentic performance. Tencent brings the Hunyuan family, including Mixture-of-Experts variants that power its own products and are available open-weight for general chat, reasoning, and multilingual work. And Baidu adds ERNIE 4.5, China's longest-running large-model line, now with big MoE and vision-language variants. If you've wondered whether the open Chinese models really are as cheap as people say, you can now check the live numbers directly instead of taking anyone's word for it.
Search, agents, and computer use
Three of the new additions are built for doing things, not just answering them. Perplexity contributes the Sonar family — models with web search built in that return grounded, citation-backed answers, including Sonar Pro and the Sonar Reasoning and Deep Research tiers. ByteDance, TikTok's parent company, ships the Seed family of general-purpose models plus UI-TARS, a GUI-agent model that drives a computer straight from screenshots.
Nous Research rounds out the group with Hermes — steerable, instruction-tuned fine-tunes of Llama known for strong function-calling and minimal refusals, a long-time favourite of the open-source agent community. If you're building anything that plans, calls tools, or browses, these three are worth pricing against the usual flagships before you default to a frontier model for every step.
Nous Research rounds out the group with Hermes — steerable, instruction-tuned fine-tunes of Llama known for strong function-calling and minimal refusals, a long-time favourite of the open-source agent community. If you're building anything that plans, calls tools, or browses, these three are worth pricing against the usual flagships before you default to a frontier model for every step.
Enterprise, RAG, and writing specialists
Five more target the enterprise stack, where governance, document length, and cost-per-seat matter as much as raw benchmark scores. IBM brings the open-weight Granite family — small, commercially-licensed, transparency-focused models tuned for coding, RAG, and tool use inside watsonx. AI21 Labs, an Israeli lab, adds Jamba, a hybrid Mamba-Transformer architecture built for efficient very-long-context inference, with 256K-token windows aimed squarely at document and RAG workloads.
Writer contributes the Palmyra family, tuned for business writing and domain-specific knowledge work, with Palmyra X5 offering a 1M-token window for whole-document tasks. Arcee AI builds small, efficient models — the Virtuoso, Coder, and Trinity families — using model-merging and distillation for strong quality-per-dollar on coding and on-prem deployment. And South Korea's Upstage adds the Solar family: compact models that punch above their parameter count via depth-up-scaling, with Solar Pro strong on document understanding and multilingual text.
Writer contributes the Palmyra family, tuned for business writing and domain-specific knowledge work, with Palmyra X5 offering a 1M-token window for whole-document tasks. Arcee AI builds small, efficient models — the Virtuoso, Coder, and Trinity families — using model-merging and distillation for strong quality-per-dollar on coding and on-prem deployment. And South Korea's Upstage adds the Solar family: compact models that punch above their parameter count via depth-up-scaling, with Solar Pro strong on document understanding and multilingual text.
Hardware, multimodal, and the fully-open lab
The last five are a more eclectic bunch. NVIDIA — yes, the chip company — publishes the open Nemotron family: models post-trained for reasoning and agentic workflows, including Llama-Nemotron variants, designed to run efficiently on its own hardware. Reka AI builds compact multimodal models, the Reka Flash and Edge series, that understand text, images, audio, and video and are efficient enough for edge deployment. Liquid AI, an MIT spin-out, takes a different path entirely with its Liquid Foundation Models — a non-transformer architecture designed to be tiny, fast, and friendly to on-device inference.
Allen Institute for AI adds OLMo, notable for being fully open: weights, training data, and code are released together, which makes it a reference point for reproducible, transparent research. Finally, Inflection AI brings the Inflection-3 models behind Pi — tuned for empathetic, safe, conversational assistants. Different shapes, different goals, all now priced side by side with everything else in the catalogue.
Allen Institute for AI adds OLMo, notable for being fully open: weights, training data, and code are released together, which makes it a reference point for reproducible, transparent research. Finally, Inflection AI brings the Inflection-3 models behind Pi — tuned for empathetic, safe, conversational assistants. Different shapes, different goals, all now priced side by side with everything else in the catalogue.
How to compare all 29 providers
With 29 providers in the catalogue, the workflow is the same as it has always been: start from your real token mix, not a marketing benchmark. Drop two or more models into the side-by-side price comparison tool to see input, output, and blended cost lined up. Convert a flat monthly budget into tokens with the token-to-USD tool, or model a specific request volume on the API cost estimator. Every figure is driven by the same live feed that refreshes daily, so the prices you compare are the prices you'll actually pay.
The bet behind all of this hasn't changed: the right model is rarely the most expensive one — it's the cheapest one that clears your quality bar. Eighteen new providers means eighteen more chances that the cheapest model good enough for your job is one you simply hadn't priced yet. Open the full providers directory to explore every lineup, or run your numbers through the calculator to see what a switch could save.
The bet behind all of this hasn't changed: the right model is rarely the most expensive one — it's the cheapest one that clears your quality bar. Eighteen new providers means eighteen more chances that the cheapest model good enough for your job is one you simply hadn't priced yet. Open the full providers directory to explore every lineup, or run your numbers through the calculator to see what a switch could save.