What models are there?
- Three, all live. Decider 1 answers typed questions over a piece of state — yes or no, a choice between labels, a score on a rubric — each as a calibrated distribution, all in one call. Restyler 1 rewrites machine-written prose so it reads like a person wrote it. Query Fanout 1 takes a question someone would ask an AI assistant and returns the web searches that assistant is likely to run before answering. Same API key, same prepaid credit, billed per token.
What is Decider 1 for?
- Decisions inside software: routing a ticket to the right team, deciding whether an alert needs a human, checking whether an agent's step succeeded, scoring severity against a rubric. You send the state and as many typed questions as you like, and every answer comes back as a probability distribution rather than a bare label, so you can set your own threshold for acting on it. Twenty questions over one state cost about the same as one.
How does Decider 1 compare with TypeSafe Jev?
- On the typed-decisions benchmark — four workflows and twenty question schemas it never saw in training — it scores 0.768 accuracy against Jev's 0.727, with probabilities far closer to the reference answers (KL 0.096 against 1.442), at $0.03 per million input tokens against Jev's $0.042. It speaks the same System One schema, so the typesafe-sdk works against it by changing its base URL. Two differences to know: a request can hold up to 4,096 tokens, where Jev takes far longer states, and a choice question offers at most ten options.
What is Restyler actually for?
- Rewriting machine-written prose so it reads like a person wrote it, without changing what it says. Generated drafts, documentation, product copy assembled by a pipeline. It preserves facts, names, numbers, links and formatting; it is not a summariser and not a general-purpose rewriter.
How good is Restyler, honestly?
- On our held-out benchmark it keeps every fact in 41% of blocks and 73% of facts on average, with zero invented first-person voice and no repetition loops. That means roughly a quarter of blocks lose something, so it is not a substitute for review where accuracy is critical. The full numbers, including where it is weakest, are on the model page.
What is Query Fanout for?
- Generative-engine optimisation. When someone asks an assistant about your category, the assistant runs its own searches before it answers, and the pages that rank for those searches are the pages it can cite. Query Fanout returns up to six of those searches for a question, which is a more useful brief than a keyword list. It is built from searches actually observed from ChatGPT, Claude and Gemini, not from queries invented by prompting a larger model.
Does it tell me what ChatGPT will really search?
- No, and treat anything that claims to with suspicion. Two runs of the same question on the same engine share none of their query strings, so the target is not a fixed list. What the model gives you is the likely directions. Measured on coverage it scores 0.563 against a hand-written rule at 0.530, where one real engine run predicting another scores 0.75 to 0.91. That ceiling is the honest upper bound and we are well below it.
Is this an AI-detector bypass?
- No, and we would rather you did not use it as one. It was trained for naturalness under strict semantic preservation; evasion was never an objective and is not measured. Detector scores move around with every detector release, so anyone selling you a guarantee there is selling you something they cannot deliver.
Why not just use a frontier model?
- You can, and for a one-off it is fine. But asking a large model to do a narrow job costs orders of magnitude more per token, and on restyling it measures worse on preservation: frontier models tend to compress, dropping content while producing text that reads well. On fan-out, a frontier model prompted blind on the same labels scored 0.082 lower on coverage, though it returned two queries where ours returned six and coverage rewards more attempts, so some of that gap is budget rather than skill.
Is the API really OpenAI-compatible?
- For Restyler and Query Fanout, yes: point any OpenAI client at meragpt.com/v1, change the model name, and it works. Decider 1 generates no text, so it is served only at /v1/systemone, in the System One format the typesafe-sdk speaks. Two honest caveats for the other two. Sampling parameters like temperature and top_p are accepted and ignored, because each model's decoding is fixed at whatever measured best for its job: greedy for Restyler, sampled for Query Fanout, where a single greedy answer collapses into six rewordings of one search. And streaming arrives a paragraph at a time rather than a token at a time.
What happens when my credit runs out?
- Requests return a 402 with a message saying so, and nothing breaks silently. Auto top-up, if you enable it, charges your card when the balance falls below a level you pick. You can set a monthly ceiling, above which it stops and emails you instead of carrying on.
Do you train on what I send?
- No. Requests are not used as training data. We record token counts, timing and cost for your usage dashboard and our billing; we do not retain the text of your requests or responses after serving them.
Does credit expire, and can I get it back?
- Credit is valid for 12 months from the day you buy it, and after that we may expire whatever is left — we would email you first. It is not refundable, which is the normal arrangement for prepaid API credit. If something breaks on our side, write to us and we will sort it out.
What happened to the Hermes agent boxes?
- That product ended on 2 September 2026 and existing subscriptions were closed out. meraGPT is now an API platform for small task-specific models. If you were a box customer and still need your data, email us.