About anex.sh

Last updated: October 20, 2026

anex.sh is a batch inference service that runs open-weights LLMs over whole datasets at a price approved before any work starts, for teams that need to process thousands to millions of texts, images, audio clips or videos.

What anex.sh does

Batch inference

You submit a JSONL file with one prompt per row and get back one result file with a response per row. Before anything runs, we price the job and hold it until you approve the estimate, and an approved job never costs more than that estimate.

Media understanding

Rows can carry images, audio and video. Models such as nemotron-3-nano-omni-30b read up to 40 minutes of audio per clip, and qwen3.8-27b reads up to 120 seconds of video, so a whole archive of recordings or product photos becomes structured data in one job.

Embeddings

Embed a whole corpus in batch, then embed live queries with the same model on the realtime endpoint, so both sit in one vector space. The VL (vision-language) embedding models take text, images and video.

Typed decisions

The laya decision model answers named questions about each row (a choice, a yes or no, a score) with a probability per option, at $0.005 per 1M input tokens. There is no generated text to parse, which makes it the cheapest way here to route, triage or tag large volumes of text.

Realtime and OpenAI-compatible APIs

The same models answer immediately on a realtime API and on an OpenAI-compatible API at oai.anex.sh, so existing OpenAI SDK code works by changing the base URL. One organization and one set of API keys cover batch and realtime.

Custom models and EU data residency

On request we serve an open-weights model we do not list yet, or your own fine-tune, to your organization only. EU data routing keeps an organization's uploads, prompts and results in the EEA (European Economic Area), at the same prices.

What makes anex.sh different

4 to 6 times cheaper than Claude batch on typical jobs

At list prices on 20 October 2026, captioning 1M images costs $268 on Qwen 3.8 27B at anex.sh and $1,075 with Anthropic's Message Batches on Claude Sonnet 5. Describing 100,000 one-minute videos costs $242 against $1,435, and 10M text tokens $2.70 against $15.

Every row has a 100-token prompt and 150 output tokens, with reasoning off. Images and video frames are 400 x 400 pixels; Sonnet 5 does not take video, so each video goes in as 60 frames at one per second. The text job is 10,000 rows of 1,000 input and 100 output tokens. Sonnet 5 batch is $1 / $5 per 1M input / output tokens.

The price is fixed before the job runs

Every job stops at a cost estimate, and the approved estimate is a hard ceiling on the bill. On aggregators like OpenRouter the billed rate can drift from the listed one: on 1 August 2026 we measured a model listed at about $0.87 per 1M output tokens billing about $3.00, 3.4 times the list price (details).

You know exactly which weights answer

Every model name maps to one published checkpoint at a stated precision, listed in the models table, and we never fine-tune, distill or quantize it ourselves. On OpenRouter the same model name can route to providers running different quantizations, some of which report theirs as unknown.

Webhooks to automate jobs

A signed webhook calls your endpoint at every step of a job, including the moment its results are ready, so there is nothing to poll. Together with auto-approve, a pipeline can submit a job, collect the results and start the next job on its own, while the approved estimate still caps what each job costs. This suits eval runs well: score every new model checkpoint on the same test set as soon as it is ready, and pick up the results without checking in by hand.

Your data stays yours

Jobs run on dedicated machines, uploaded data and results are never used to train models, and a job's data can be deleted with one API call the moment it finishes. Without that call, inputs are purged after 7 days and results after 30.

No contract and no subscription

You buy prepaid credits that do not expire and pay only for the jobs you approve. There is no seat fee, no monthly minimum and no commitment period; the smallest job costs between $0.02 and $1.00 depending on the model.

Who uses anex.sh

• ML and data teams that run model evaluations, benchmarks or synthetic-data generation over a fixed dataset.
• Search teams building AI search over text, image and video collections.
• European companies that need their data processed inside the EEA.
• Marketing and social media analysts who study thousands of TikTok, Instagram or YouTube videos, like our 2,264-post TikTok study.
• E-commerce teams that tag, categorize, describe or embed product catalogs of tens of thousands to millions of items.
• Sales and customer support teams that transcribe, summarize and categorize call recordings and tickets.

The team behind anex.sh

anex.sh started as ANEX, our open-source engine for integrating GPU containers into Kubernetes clusters, and that engine still runs the GPU fleet today. We built the batch service because we kept paying API prices for dataset work and knew, from running our own GPUs, how much cheaper the same tokens could be.

We are a small team of AI and infrastructure engineers in Prague, with more than 30 years of combined experience building production ML systems. anex.sh is operated by MITON AI Spice s.r.o.

Radek Bartyzal, CEO

Radek has 10 years of experience building production AI and machine learning systems, including a recommender system that serves millions of users in 14 countries. His work covers training neural networks and large language models (LLMs), quantitative trading and ranking algorithms. LinkedIn

Antonín (Tonda) Hoskovec, Head of AI

Antonín holds a PhD in quantum physics and has 15 years of experience building production AI and machine learning systems. He has built search engines, image recognition, quantitative trading systems and neural network training. LinkedIn

Jan Skarupa, DevOps

Jan has 10 years of experience in DevOps and machine learning operations (MLOps), managing large Kubernetes clusters. LinkedIn

How anex.sh works

Getting started

Create an account in the console, buy credits and create an API key. From there you submit jobs from the browser, the batch HTTP API or the OpenAI SDKs; the docs walk through the first job.

Turnaround

A job is delivered within 24 hours, and most come back in less than one hour. If you need a tighter guarantee for a particular workload, write to us and we will agree it.

Talking to us

Support is by email at support@anex.sh, and we reply within two business days. You talk to the engineers who wrote the code, in English or Czech. For invoicing, custom models or larger projects, write to hello@anex.sh.

Billing

Credits are prepaid, and every top-up comes with an invoice.

Key facts

Frequently asked questions

Which models can I run?

Open-weights models from the Qwen, Gemma, Nemotron and Laya families, for text generation, media understanding, embeddings and typed decisions. The models table names each exact checkpoint and precision, and we can serve another open-weights model on request.

Do I need to sign a contract or talk to sales?

No. You can sign up, buy credits and run a job without talking to anyone. Write to hello@anex.sh only if you want a custom model or a delivery guarantee.

Do you train on my data?

No. Uploaded datasets and results are never used to train models, and we do not sell customer data. What we store and for how long is in the privacy policy.

Is the infrastructure open source?

The engine that runs our GPU fleet, ANEX, is open source and free on GitHub. The batch service built on top of it is a hosted product.