From 6 to 13 November, 20 models for text agents, classification, voice tasks, and image generation appeared in the OpenRouter catalog. For an engineer, this is more than just a longer list in the menu: the new arrivals range from free options to models with a million-token context. Oink, there’s more to choose from now, but you should still check the token price and data-handling terms before connecting.

Text models use the same Chat Completions route — the model identifier changes in the request. Decision models, speech recognition and synthesis, and image generation, on the other hand, have separate API routes. One OpenRouter key simplifies access, but it doesn’t turn every request format into one universal call.

Low-cost candidates for agents

Muse Ember is a compact model from Alexander Wang’s superintelligence lab, designed for repetitive agent tasks: tool calls, coding-agent subtasks, request routing, and data extraction. It accepts text, images, video, audio, and PDFs, has a context window of 1 million tokens, and responds with text. It costs $0,10 per million input tokens and $0,50 per million output tokens; cached input costs $0,004 per million. The model is closed, and is served by its developer and Google Vertex.

The price makes Muse Ember a potential subagent for routine work under a more expensive planner, but its suitability for a specific workflow needs to be checked with your own prompts. Oink. A cheaper call is a useful feature, not a quality certificate.

Nex-N2.6 from Nex AGI is based on GLM-5.3-Flash and has about 321 billion parameters. It is positioned for computer control, programming, and long tasks involving tool use. Its weights in FP8 format are available under Apache 2.0, its context window is 262 thousand tokens, and it costs $0,12 per million input tokens and $0,40 per million output tokens.

According to Nex AGI’s internal tests, Nex-N2.6 scored 83,0 on OSWorld versus 78,0 for the GLM-5.3-Flash baseline; on Terminal-Bench 2.1, the results were 84,1 and 84,3 respectively, and on AutomationBench, 45,7 versus 48,8. The developers also report that repetitive reasoning fell to 1,6%, and that among responses with tool calls, all used only tools declared in the request. These are internal measurements; this article has no independent confirmation. Don’t take your hooves off the keyboard: you still need to test agent behavior on your own scenarios.

Another new text model, ERNIE-Light-DT-1.0 from Baidu, uses Baidu Sparse Attention. According to the company, the mechanism compresses context history using feedback, reducing KV cache size and computation. The model has a context window of 1 million tokens, accepts text, and costs $1,13 per million input tokens and $3,94 per million output tokens. That’s noticeably more expensive on input than Muse Ember and Nex-N2.6; no separate open weights for this version were found.

Free options — with conditions

ERNIE 5.5 Flash Preview from Baidu is available on OpenRouter for free; there is currently no paid tier in the catalog. It is an MoE model with 14 billion active parameters out of 280 and a context window of 1 million tokens. Baidu claims generation speeds of over 200 tokens per second and improvements in programming, agent work, and deep search. Open weights are available for the base ERNIE-5.5-Flash-Base under Baidu’s own license, but not for Preview.

Liquid AI offers LFM3-1.7B: a model with 1,7 billion parameters and a context window of 256 thousand tokens, tuned for agents, tool calling, and structured output. Its weights are available for local deployment under the LFM 1.0 license. OpenRouter has a free option and a paid tier — $0,08 per million input tokens and $0,20 per million output tokens. At the same time, Liquid AI may retain prompts and responses and use them for training. Oink, a free tier doesn’t mean you can skip reading the data-handling rules.

The catalog also includes free Decider V1.3 and Decider Flash V1.3 from Perplexity, and a free version of Inkling 2 Small from Thinking Machines. For Inkling’s paid tier, the listed price is $0,30 per million input tokens and $1 per million output tokens; its weights are published under Apache 2.0. For the free option, the model page warns that prompts and responses are recorded to improve the models, and that confidential and personal data must not be submitted.

Decision models return probabilities, not chat responses

Instead of generating free-form text, a decision model receives a state — for example, text, JSON, or an image — and a set of questions with specified answer types: “yes or no,” a choice from a list, or a rating on a scale. The output contains probabilities for the options, which application code can compare against a threshold. This approach works for routing, moderation, ticket sorting, and criteria-based evaluation, where parsing an arbitrary chat-model response is inconvenient.

Decider V1.3 and Decider Flash V1.3 accept text, JSON, and images, have a context window of 262 thousand tokens, and are available for free. V1.3 introduced a mode locate via the Extractions route: it extracts fields of type text і list, adding a source snippet and a confidence score. Perplexity recommends Flash for large-scale classification and moderation.

HeyGen Decision 1 works with text and images, has a context window of 131 thousand tokens, and costs $0,14 per million input tokens; output is free. Decision 1 Turbo is marked as a preview, has a context window of 262 thousand tokens, and costs $0,09 per million input tokens. Drex Max Finance from Nace.AI is designed for financial documents: it processes invoices, receipts, and statements as images, as well as text and JSON. It costs $0,06 per million input tokens and has a context window of 262 thousand. The weights for the base Drex Max are published under Nace.AI Open RAIL-M, but the company does not report whether the finance version has weights available.

This group also includes Gemini Decision 1 from Google: $0,10 per million input tokens, free output, and a context window of 262 thousand tokens. There are six decision models in the list overall, counting this new arrival alongside the two Deciders, two HeyGens, and Drex Max Finance.

Voice and images use different pricing units

Grok STT 2.0 from SpaceXAI converts speech to text, adds timestamps to every word, can separate speakers, and supports multichannel audio. It costs $0,000028 per second, or approximately $0,10 per hour of audio. Simba 3.5 from Speechify does the opposite — synthesizes speech from text — and supports 48 languages. It costs $15 per million characters.

MiniMax H3 Image generates images from text and edits or composes them based on between 1 and 16 references. One request can produce up to 10 images, with a size of up to 20 megapixels and an aspect ratio from 1:3 to 3:1. The weights are available under MiniMax’s own license, with restrictions: it does not apply in the US, EU, UK, or South Korea; annual revenue above $1 million or launching a service for third parties requires separate written permission from the company.

H3 Image pricing depends on tokens: $2,30 per million input text tokens, $1,50 per million input image tokens, and $3,80 per million output image tokens. H3 Image Highspeed has the same described capabilities, but costs $3 per million input text tokens and $7,60 per million output image tokens. The catalog doesn’t give a fixed price per image — it needs to be calculated based on actual usage in the field usageSeedream 5.5 from ByteDance Seed is priced per image: $0,02 at standard resolution and $0,035 at high resolution; each input reference costs $0,008. The model supports up to 10 references per request, and its weights are closed.

What was already in the catalog and how to connect the models

Among the 20 entries are models that have already been reported on separately. Claude Fable 5.2 from Anthropic has a context window of 1 million tokens and costs $8 per million input tokens and $40 per million output tokens. Anthropic says it generates up to 30% faster than Fable 5.1; for everyday programming, the company recommends Opus 5.5. Gemini 4 Carbon costs $2 and $10, while Gemini 4 Hydrogen costs $1,50 and $7,50 per million input and output tokens respectively; both have a context window of 1 million. These models are closed.

For text models, simply send a request to Chat Completions and specify the desired identifier in the field modelA free option is marked with the suffix :free. Other model types use separate routes and request formats, examples of which are provided on the OpenRouter model pages.

The practical takeaway is simple: Muse Ember and Nex-N2.6 are interesting for their low cost on agent tasks, free models let you start testing without paying, and decision models can eliminate the need to parse a text response in code. Voice and image pricing is calculated in seconds, characters, or tokens — so you can’t compare them just by the price per million text tokens. PigmanOps approves not of a flashy name, but of testing with your own prompts: most of the claims and benchmarks listed here come from the developers, and MiniMax image-generation costs depend on actual usage.