> The full index of these docs is at [llms.txt](https://u22a8.ai/llms.txt).

# A8 model card

A8 on one page: model IDs, pricing, capabilities, endpoints, features, rate limits and how it compares.

*Preview. Applies to a8-1. Last reviewed 2026-10-26.*

**Model ID:** `a8` (Preview)

For judging text against criteria stated in plain words, at the accuracy each request sets.

- **Minimum accuracy:** 0–99, default 90
- **Context window:** TBD (tokens)
- **Input price:** $2 per 1M tokens
- **Latency:** TBD (median)

## Overview

**A8 (Touchstone)** is a decision model with a chat-model interface. It reads a conversation, judges its subject against the criterion in the system message and returns a verdict in the shape your schema gives. Each request sets a minimum accuracy: A8 answers where its evidence meets it and abstains everywhere else. A case it abstains on goes to a person. Tell your assistant their answer, and over MCP your organization’s model learns it in the background.

## How it compares

| Attribute             | Trained classifier               | LLM judge           | Decision model                                    | A8                                                          |
| --------------------- | -------------------------------- | ------------------- | ------------------------------------------------- | ----------------------------------------------------------- |
| The criterion         | Training labels                  | In the prompt       | In the question                                   | In the prompt                                               |
| Output                | A label or a score               | Text or JSON        | A choice or a score, with a probability           | A verdict in JSON                                           |
| Accuracy              | Measured before use              | Measured afterwards | A threshold you tune on labeled data, per version | Set per request, in the model’s tag; A8 finds the threshold |
| Determinism           | Deterministic                    | Stochastic          | Deterministic                                     | Deterministic                                               |
| What it starts from   | A labelled dataset per criterion | A prompt            | A pretrained model, the same for every customer   | A pretrained model, and examples for your own criteria      |
| Learning a new answer | Relabel, retrain and redeploy    | Rewrite the prompt  | Reword the question                               | Tell your assistant; it teaches A8                          |
| Why it answered       | Feature weights                  | A written rationale | The probability                                   | The taught examples behind the verdict                      |

## Specifications

### Model IDs

| Identifier                | Points to                                                       |
| ------------------------- | --------------------------------------------------------------- |
| `a8`                      | The latest version, at your organization’s latest ready horizon |
| `a8:95`                   | The same model, at a minimum accuracy of 95                     |
| `a8-1-1@20260715t140322z` | Version 1.1 at one horizon: your data held at that moment       |
| `a8@20260715t140322z`     | The same model as the full form                                 |
| `a8-1-1@reference`        | Version 1.1 as we publish it, before any teaching               |

Every response names the full identifier that answered, with the minimum accuracy it answered at as its tag. See [Model identifiers](https://u22a8.ai/docs/a8-model-identifiers.md) and [Horizons](https://u22a8.ai/docs/a8-horizons.md).

### Pricing

| Metric                 | Price | Unit                              |
| ---------------------- | ----- | --------------------------------- |
| Input                  | $2    | 1M tokens                         |
| Input, reference model | $2    | 1M tokens                         |
| Output                 | n/a   | A verdict writes no output tokens |
| Reasoning              | $0    | 1M reasoning tokens               |
| Cached input           | $0    | 1M cached input tokens            |
| Teach                  | $0    | 1 teach                           |
| Retract                | $0    | 1 retract                         |
| Fitting a new version  | $0    | 1 fit                             |
| Precedents and recall  | $2    | 1M tokens                         |

Prices of A8 1.1 Preview. See [Pricing](https://u22a8.ai/docs/pricing.md).

### Capabilities

| Feature           | Value                                                          |
| ----------------- | -------------------------------------------------------------- |
| Input → output    | A conversation in text → a verdict: a label, a number or JSON  |
| Context window    | TBD (tokens)                                                   |
| Conversations     | Multi-turn: A8 answers the latest user message                 |
| Minimum accuracy  | `0` to `99` as the model’s tag, `a8:95`; `90` without one      |
| Reasoning         | On by default; effort `none` or disabled thinking turns it off |
| Determinism       | The same request on the same version returns the same verdict  |
| Learning          | Taught examples, over MCP, fitted in the background            |
| Latency           | TBD (median and 95th percentile)                               |
| Languages         | TBD (languages measured)                                       |
| Reference horizon | TBD (the reference model’s horizon)                            |

### Availability

| Feature    | Value                                                                |
| ---------- | -------------------------------------------------------------------- |
| Status     | Preview                                                              |
| Released   | TBD (release date)                                                   |
| Retirement | On or after TBD (date)                                               |
| Data scope | Your organization’s model, trained on what your organization teaches |

## Endpoints

| Endpoint           | Route                                | Support       |
| ------------------ | ------------------------------------ | ------------- |
| Chat Completions   | `POST /v1/chat/completions`          | Supported     |
| Responses          | `POST /v1/responses`                 | Supported     |
| Anthropic Messages | `POST /v1/messages`                  | Supported     |
| Decisions          | `POST /v1/decisions`                 | Supported     |
| Stored responses   | `GET`, `DELETE /v1/responses/{id}`   | Supported     |
| Input items        | `GET /v1/responses/{id}/input_items` | Supported     |
| Models             | `GET /v1/models`                     | Supported     |
| MCP                | `https://api.u22a8.ai/mcp`           | Supported     |
| Batch              | `/v1/batches`                        | Not supported |
| Embeddings         | `/v1/embeddings`                     | Not supported |

Every route is on `https://api.u22a8.ai`. Each interface page lists the fields A8 reads: [Chat Completions](https://u22a8.ai/docs/a8-chat-completions.md), [Responses](https://u22a8.ai/docs/a8-responses.md), [Anthropic Messages](https://u22a8.ai/docs/a8-anthropic-messages.md), [Decisions](https://u22a8.ai/docs/a8-decisions.md) and [MCP](https://u22a8.ai/docs/a8-mcp.md).

## Features

| Feature                                       | Support                                 |
| --------------------------------------------- | --------------------------------------- |
| Structured outputs (`json_schema`)            | Supported                               |
| A forced function call as the verdict’s shape | Supported                               |
| Multi-turn conversations                      | Supported                               |
| Streaming                                     | Supported                               |
| Reasoning on and off                          | Supported, through effort or `thinking` |
| Content filtering                             | Supported, in each API’s own slot       |
| Teaching and retracting                       | Supported, over MCP                     |
| Precedents and recall                         | Supported, with `models:explain`        |
| Pinned versions                               | Supported                               |
| The `json_object` response format             | Not supported                           |
| More than one choice (`n`)                    | Supported, each computed on its own     |
| Stored responses (`previous_response_id`)     | Supported                               |

## Modalities

| Modality | Support          |
| -------- | ---------------- |
| Text     | Input and output |
| Image    | Not supported    |
| Audio    | Not supported    |

## Rate limits

| Tier    | RPM                       | TPM                     |
| ------- | ------------------------- | ----------------------- |
| Preview | TBD (requests per minute) | TBD (tokens per minute) |

Limits apply per organization. The response headers that report them are on [Rate limits](https://u22a8.ai/docs/rate-limits.md).

## Good to know

- A prompt written for another judge works as it is. A8 finds the criterion, the subject and the answer format, such as `GRADE: C`, in a single message.
- The model’s tag sets the minimum accuracy: `a8:95` asks for 95. Words about certainty in a prompt are read as part of the criterion.
- Reasoning is on by default and reads a request however it is laid out. With reasoning off, A8 reads the layout as written: the system message is the criterion and the latest user message the subject. Each interface page names its switch.
- The `system_fingerprint` on each response moves when anything behind a verdict moves. See [Reproducibility](https://u22a8.ai/docs/a8-reproducibility.md).
- “What have I taught you about refunds?” returns your taught examples, word for word. See [Precedents](https://u22a8.ai/docs/a8-precedents.md).
- `GET /v1/models` lists each version your organization can call, with its `status` and whether it is `active`.

## Evidence

[qed-bench](https://u22a8.ai/research/qed-bench) measured the scoring approach A8 is built on. On essay quality (ASAP 2.0), its rank correlation with trained human raters was 0.815; the strongest LLM judge on the panel, Claude Opus 4.7, reached 0.813. For your criteria, the minimum accuracy each request sets is the promise, and [your own tests](https://u22a8.ai/docs/a8-testing.md) check it.

## Next steps

- [Quickstart: a first verdict](https://u22a8.ai/docs/quickstart.md): Get a key, call A8 from an SDK, from curl or from your assistant, and read the verdict it returns.
- [Accuracy and abstention](https://u22a8.ai/docs/a8-abstention.md): How a request sets the minimum accuracy a verdict must meet, what that level promises and the five kinds of abstention.
- [Known limits: a8-1](https://u22a8.ai/docs/known-limits-a8-1.md): Where a8-1 errs or declines, and what to do instead.
