What runs Bloks AI, and where your words go.

The models

Bloks AI runs Claude Haiku 5.5, GPT‑6 Luna and DeepSeek V4.1 Flash, finetuned by Bloks for agentic work: each kind of task is routed to the model that does it best, with instructions built for agents.

• Coding runs on Claude Haiku 5.5, from Anthropic.
• General work, conversation and anything not tagged otherwise runs on GPT‑6 Luna, from OpenAI.
• Automation, the scheduled and background work, runs on DeepSeek V4.1 Flash, an open‑weights model from DeepSeek. It is served by hosting companies (see below), never by DeepSeek’s own service.

You do not pick between them; the work does. Every answer is labelled Bloks AI, and the app does not show which of the three answered a given request. Each model reads up to about a million tokens of context, and one request, attachments included, can be up to 8 MB.

When the models change, this page changes with them, and its date moves.

The allowance

• Shown as a meter, in percent. Bloks AI includes a monthly allowance, on its own or beside Cloud or Team. The Hosted bundles include a little less. Settings, Plan and the Usage page show what is left of the month.
• Monthly, with no rollover. The allowance resets at the start of each billing month, and what is left at the end of one is not carried over. A yearly plan refills every month.
• A daily limit, which you set. By default one day can use at most half the month, so a runaway loop cannot spend the whole month overnight. Raise or lower it in Settings.
• What counts. Each request is counted at its model’s real price, set by the company that serves it: the text it reads, the text it writes, and its reasoning, which counts as writing. Text a model has cached reads far cheaper. Very long conversations cost more per token past a certain length, and some hosts charge more at busy hours. The meter follows the real cost; there is no markup per request.
• What does not count. A request that is refused, or that never reaches a model, costs nothing. A request you stop partway costs what was written before it stopped.
• Rate. Up to 120 requests a minute per licence, which is a stop for a runaway loop rather than a pace anyone normal meets. There is no separate limit on how many agents work at once: they share the allowance.
• At a limit. When the month or the day is used up, Bloks AI requests stop, and the message box says so, until the limit resets or you top up. Nothing switches to a weaker model, nothing slows down quietly, and nothing is charged.
• Top‑ups. A top‑up adds as much as a month of Bloks AI on its own. It never expires while you are subscribed, and it is used after the month’s allowance.
• Automatic top‑up. Off unless you switch it on. When on, it buys one top‑up on your saved card when the balance runs low. Nothing else is ever charged automatically.
• Teams. A team shares one allowance. Its owner can see each member’s use and cap each member’s share of a day.

Where a request goes

1. Bloks. Your app sends the request to Bloks, on Cloudflare, with your licence, so it can be counted against your allowance.
2. Vercel AI Gateway. Bloks forwards it, under Bloks’s own key, to Vercel’s AI Gateway, which picks a company to serve the model.
3. The company that runs the model. Only companies with a zero data retention agreement with Vercel are ever chosen. For Claude Haiku 5.5: Anthropic, Amazon Bedrock, Google Vertex AI or Claude Platform on AWS. For GPT‑6 Luna: OpenAI or Microsoft Azure. For DeepSeek V4.1 Flash: one of the hosts on Vercel’s zero retention list that serves it, such as Fireworks, Together AI, Baseten or Alibaba Cloud.

This path is for Bloks AI only. The relay that connects your phone and a Hosted workspace machine are separate services. A workspace on your own computer keeps its files there, but its Bloks AI requests still take this path.

These companies run models in the United States and other countries. Bloks AI has no setting to keep a request in one region today.

What is kept, and what is trained on

• Bloks keeps one line for each request: when it was made, which model, which kind of work, how many tokens, what it cost, and, on a team, which member. Never what the request or its answer said. Our logs record errors and counts, not contents.
• Vercel AI Gateway keeps nothing: a request is deleted when it completes.
• The companies that run the models keep nothing either. Bloks asks for zero data retention twice, as a setting on the whole gateway and on every request, and a company without such an agreement is never used: a request fails rather than go to one. OpenAI’s zero retention terms keep an exception for safety review, described in its policy.
• Nobody trains on your content. Not Bloks, not Vercel, not the companies that run the models. The finetuning above is Bloks’s routing and instructions; none of it is learned from what you send.

What the model companies can see

• Not who you are. Requests reach Vercel and the model companies from Bloks’s servers, under Bloks’s key, so they do not see your IP address, your email, your licence or your team.
• What you asked. They do see the request itself, the conversation and anything attached, because that is what they answer. If a name or a number is in the conversation, it is in the request.

What Bloks AI does not do today

• Remove or replace names, numbers or secrets before a request is sent.
• Keep a request in a region you choose.
• Show which model answered a given request.

If you need any of these, your own API key with a provider you choose works in Bloks too, under your own agreement with them. The Privacy Policy covers everything else, and privacy@bloks.dev answers questions about this page.