Product Marketing Manager - Enterprise

Image Benchmarks: see the capabilities of every model. Brian Thomas ·8/21/2026 Unlike choosing a text model, where OpenRouter, Inc has a vast array of LLM benchmarks, picking an image model can feel arbitrary. Output samples tend to be curated eye candy and LLM-as-a-judge evals can't yet capture the details a human would notice instantly. While arena scores help, they evaluate which outputs people prefer rather than what a model can actually do. Today OpenRouter, Inc is launching Visual Image Benchmarks to help you quickly evaluate the capabilities of all the image models OpenRouter, Inc offer (39 as of August '26). OpenRouter, Inc has selected a set of challenging prompts designed to differentiate the capabilities of models, and show every result in a grid with sorting for both price and generation time. Prompts designed to test the boundaries of image model capabilities. Each challenge is designed to differentiate the capabilities of image models. OpenRouter, Inc has initially grouped them into seven families: * Improbable scenes. A wine glass filled level with the rim, umbrellas that are closed. Training data is full of the ordinary version of both. * Counting. Three fingers, specific numbers of cards and dice. * Text. One long exact string on a poster, and several languages in the same frame. * Spatial relations. Occlusion and mirror reflections. * Negation. A zebra with no stripes, a Times Square with no advertising. * Editing. Minimal diffs, object removal, person removal, all from a reference image. * Consistency. Holding a product or four reference subjects steady across a new scene. The prompts are written so you can instantly evaluate them visually. For example, can a model follow the instruction to fully fill a wine glass? More visual evals and additional modalities coming soon. It's similarly challenging to evaluate video and audio models to understand their quality and capabilities. OpenRouter, Inc intend to expand this tool out across modalities as well as keep it up to date with all the new image models OpenRouter, Inc add. Check out its image benchmarks today! If you have any prompts that could challenge the next round of models, share it with OpenRouter, Inc in #feedback on its Discord. Generating images via the OpenRouter API and Chat. Once you've evaluated models via these benchmarks, try them on your own prompts through the image generation API or Chat to see how they perform on your own content.

Ox Alpha: a stealth 1M-context coding and agent model on OpenRouter, free while it collects feedback. Published August 20, 2026 4 min read OpenRouter, stealth models, coding agents, long context, reasoning OpenRouter listed a stealth model called Ox Alpha on August 20 with a 1M context and text, image and video input. The API confirms both exactly: 1,048,576 tokens and text+image+video->text, with 131,072 max output, one of 421 models on the router. Two things the announcement leaves out are in the same response. Reasoning is mandatory and cannot be disabled, and the default effort is max, which sits awkwardly beside the word efficient. And the price is zero on both input and output. Links & resources. | Resource | Link | | The announcement | @OpenRouter, August 20 | | The model | openrouter.ai/stealth/ox-alpha | | Model ID | stealth/ox-alpha | | Price | Free while in stealth | OpenRouter put a stealth model called Ox Alpha on the router on August 20, described as "a frontier model built for efficient coding, sustained agentic work, and real-world production use," with a 1M token context window and text, image and video input. Nobody has said whose model it is. That is the point of a stealth listing: you get to use it, the lab gets feedback, and the name stays off until launch. The claims check out against the models endpoint, which is the useful thing about OpenRouter, because everything a listing asserts is queryable. Context length is 1,048,576, so the 1M is exact rather than marketing rounding. Modality is text+image+video->text, matching the three input types in the tweet. Max completion tokens is 131,072. It is one of 421 models on the router today. Two things the tweet does not mention are in the same response, and both change how you would use it. Reasoning is mandatory and defaults to max. The API's reasoning block for this model reads: "reasoning": {"mandatory": true, "default_enabled": true, "supported_efforts": ["max", "high", "low"], "default_effort": "max"} Mandatory. You cannot turn reasoning off on Ox Alpha, and if you send a request without specifying effort, you get max, the most expensive and slowest of the three settings. Set that against the word in the announcement: "efficient coding." A model that always reasons, and reasons at maximum by default, is not the shape most people picture when they read efficient. It may well be efficient per unit of work completed, which is the meaningful measure for agentic tasks and probably what OpenRouter means. It is not efficient per request, and anyone benchmarking it against a non-reasoning model without pinning reasoning_effort: "low" is comparing two different things. Note also that the three efforts are max, high and low. There is no medium and no minimal, so the step from low to high is a wide one and there is no way down from low. It is free, and the listing does not say so. Pricing in the API is "prompt": "0" and "completion": "0". Zero on both. OpenRouter's own model card confirms it, printing "Free" in the pricing panel next to "1.05M" for context. The tweet says "Try it now and share feedback to improve the model," which implies the arrangement without stating it. Free is the consideration you are paying with feedback for, and it is worth being explicit about what that means: stealth models on OpenRouter exist so an unreleased lab model can be exercised on real traffic. Your prompts are the product being bought. If you are sending anything you would not want an unnamed lab reading, this is the wrong model to test it on. There is also a small tell in the metadata. The listing's description calls it "a reasoning model designed for coding, sustained agentic work, and production workloads," where the tweet calls it "a frontier model built for efficient coding." Same sentence, one word swapped, and the API's version is the accurate one. What is knowable about whose it is. Not much, and I am not going to guess. hugging_face_id is null, knowledge_cutoff is null, and the tokenizer is listed as "Other," which is what OpenRouter shows when it cannot map a model to a known family. The expiration_date field reads 2098-12-31, which is a placeholder rather than a plan. The supported parameters are ordinary: tools, tool_choice, response_format, temperature, top_k, top_p, max_tokens, reasoning, reasoning_effort, include_reasoning. Default temperature 1, default top_p 0.95. Nothing there fingerprints a lab. The one structural hint is the combination itself. A 1M context with video input, mandatory reasoning at three effort levels, and 131K max output is a short list of labs. Video input in particular narrows it considerably. But a short list is not an answer, and stealth listings have surprised people before. What to do with this. Try it, because it costs nothing and 1M context with video input is not a combination you can test for free anywhere else. Pin reasoning_effort explicitly on every call, or your latency and token numbers will be from the max setting and you will draw the wrong conclusion. And treat the free tier as what it is. Feedback is the price, and the model card tells you the model is unreleased and unnamed. Test it on work you would be comfortable having read.

Payments firm Stripe to buy marketplace OpenRouter in AI push. 20 Aug 2026 01:36AM (Updated: 20 Aug 2026 03:43AM) Add CNA as a trusted source to help Google better understand and surface our content in search results. Aug 19: Payments firm Stripe said on Wednesday it has agreed to buy OpenRouter, which helps businesses route and optimize token usage, at an undisclosed value, its latest bet on the fast-growing AI industry. The deal comes as soaring bills drive companies toward cheaper models and routing tools like OpenRouter, and deepens a year-long push by the fintech into AI. The push has included launching products like token billing to track AI model consumption. AI marketplace platforms like OpenRouter allow developers to send queries to dozens of AI models through a single interface, making them a popular testing ground for new systems. "Tokens are the central currency for companies building with AI, and it's clear that the real-world economic potential will depend on making good use of scarce compute resources," Stripe CEO Patrick Collison said. Bloomberg News reported earlier this month that Stripe had agreed to buy OpenRouter for more than $7 billion, citing people familiar with the matter. Stripe declined to comment on the purchase price, while OpenRouter did not immediately respond to Reuters request for comment. OpenRouter, which was founded in 2023, processes more than 10 trillion tokens per day from more than 400 AI models for a community of over 10 million developers and companies. A Deloitte report in March found that most surveyed companies with annual revenue of at least $500 million expect to consume more than 10 billion tokens per month by 2028. OpenRouter has raised capital from heavyweights such as Menlo Ventures and Andreessen Horowitz. In May, the firm raised $113 million in a funding round led by Alphabet's independent growth fund CapitalG. Stripe, which was valued at $159 billion in a tender offer earlier this year, has made a joint offer with private equity firm Advent International to buy PayPal for more than $53 billion, Reuters reported last month. Businesses running on Stripe generated $1.9 trillion in total volume last year, a 34 per cent jump from 2024.