# MiniMax docs: AI Agent Readiness Score 76.2% (C)

**76.2% · 61/80 · AI Agent Readiness Score · 21/30 reading points · 40/50 agent surface points**

MiniMax received 7 PASS votes and passed 4 of five agent surface checks. The clearest finding came from the find the exact rate limits task.

- Tested: 2026-08-20
- Published: 2026-09-21
- Battery: v1
- Scoring: reading 30 pts · surface 50 pts
- Docs: https://platform.minimax.io/docs/

Three AI models, GPT 5.6 Sol, Claude Opus 5, and DeepSeek v4 Flash, each read MiniMax’s public documentation independently and attempted five first-hour developer jobs: make the first model request, find the exact rate limits, recover from rate-limit errors, authenticate an API request, list video tasks with pagination.

No accounts, API calls, or code execution were used. Every verdict came from public pages and every published quotation passed a live verification check. Preflight substitutions: T3: MiniMax errors 1002 and 2045 replace generic HTTP 429. T4: API-key authentication replaces webhooks. T5: Filtering and pagination replace an official SDK.

## Freshness

[How rechecks work](https://docsforagents.com/methodology/#freshness)

- Category: [AI models & inference](https://docsforagents.com/grades/?category=ai-models)
- Tested: 2026-08-20
- Quotes verified: 2026-08-20
- Surface rechecked: 2026-09-21

No change since the test.

5 of 5 quoted passages still appear on the live pages.

## Agent surface checks · 40/50

| Check | Verdict | Points |
| --- | --- | --- |
| llms.txt | PASS | 10 |
| llms-full.txt | PASS | 10 |
| Markdown mirror | PASS | 10 |
| MCP server | PASS | 10 |
| Docs AI | FAIL | 0 |

## The Reading Test · 21/30

| Task | GPT 5.6 Sol | Opus 5 | DeepSeek v4F | Consensus |
| --- | --- | --- | --- | --- |
| Make the first model request | PARTIAL | PASS | PASS | PASS |
| Find the exact rate limits | PARTIAL | PARTIAL | PASS | PARTIAL |
| Recover from rate-limit errors | PARTIAL | PARTIAL | FAIL | PARTIAL |
| Authenticate an API request | PASS | PASS | PASS | PASS |
| List video tasks with pagination | PASS | PARTIAL | PARTIAL | PARTIAL |

Docs platform: Mintlify (unscored) · verified 2026-08-20

## What to fix first

These 4 fixes could add up to 18 points to the AI Agent Readiness Score. The list ranks each fix by the points it would add. [How the ranking works](https://docsforagents.com/methodology/#what-to-fix-first)

1. **+10 points · Docs AI · check failed**

   **Found:** The live Mintlify configuration disables Ask AI controls.

   **Fix:** Add an assistant to the docs site that answers questions from the docs and links to its sources.

2. **+4 points · Recover from rate-limit errors · PARTIAL**

   **Found:** The Error Codes page gives 1002 and 2045 one-line solutions with no retry timing or backoff. DeepSeek also found the page does not mention HTTP 429.

   **Fix:** Add retry timing, a backoff example, and the HTTP status for 1002 and 2045 to the Error Codes page.

   **Evidence:** [platform.minimax.io/docs/api-reference/errorcode](https://platform.minimax.io/docs/api-reference/errorcode)

3. **+2 points · Find the exact rate limits · PARTIAL**

   **Found:** The Rate Limits page omits TPM or concurrency for most modalities. Opus 5 also found free and paid tiers only in the video table.

   **Fix:** For each modality, state which limit types apply and which tier each value covers on the Rate Limits page.

   **Evidence:** [platform.minimax.io/docs/guides/rate-limits](https://platform.minimax.io/docs/guides/rate-limits)

4. **+2 points · List video tasks with pagination · PARTIAL**

   **Found:** The video task list uses page_num paging, not a cursor. Opus 5 also found no default or maximum page_size and no sort order.

   **Fix:** State the default and maximum page_size and the sort order, and note that paging uses page_num, not a cursor.

   **Evidence:** [platform.minimax.io/docs/api-reference/video-generation-v2-list](https://platform.minimax.io/docs/api-reference/video-generation-v2-list)

## What the docs get right

- **Authenticate an API request: 3 PASS votes.** The pay-as-you-go path documents key creation, .env or environment storage, Anthropic variables, and SDK transmission.
- **Make the first model request: 2 PASS votes.** Prerequisites gives the register link, the API Keys console link, and the two exports; Integrate via SDK then gives pip install anthropic and a runnable client.messages.create call with model="MiniMax-M3", so the path from zero to a first M3 request is one unbroken chain with no guessing.
- **4 of 5 agent surface checks.** Present: llms.txt, llms-full.txt, markdown mirrors, an MCP server. Missing: docs AI.

## Make the first model request

**PASS**

PASS consensus from 2 PASS, 1 PARTIAL.

The request path is clear, but its sample contradicts the documented thinking default; the linked key console did not render. Prerequisites gives the register link, the API Keys console link, and the two exports; Integrate via SDK then gives pip install anthropic and a runnable client.messages.create call with model="MiniMax-M3", so the path from zero to a first M3 request is one unbroken chain with no guessing. From zero to first API call in three steps: register, create API key, set environment variables, then use the Anthropic SDK with model='MiniMax-M3'.

## Find the exact rate limits

**PARTIAL**

PARTIAL consensus from 1 PASS, 2 PARTIAL.

The table gives modality values, but Token Plan throttling is dynamic and most modality tables omit TPM or concurrency. Numbers exist for all five modalities (MiniMax-M3 200 RPM / 10,000,000 TPM, video Hailuo 5 free and 20 paid RPM, H3 2 and 15 CONN, speech T2A 60 RPM, image-01 10 RPM, music 120 RPM and 20 CONN) but TPM is published only for LLM, CONN only for video V2 and music, and Free tier / Paid tier columns appear only in the Video table, so the tier that applies to every other modality has to be guessed. Every model category has a documented limit (RPM, TPM, or CONN). Values are internally consistent and no contradictory numbers appear across pages.

## Recover from rate-limit errors

**PARTIAL**

PARTIAL consensus from 2 PARTIAL, 1 FAIL.

Error 1002 says retry later, while 2045 says smooth request growth; neither gives retry timing or backoff. Error Codes names both causes (1002 "rate limit", 2045 "rate growth limit") and gives 2045 a real request-shaping correction, but its 1002 remedy is only "Please retry your requests later." with no backoff schedule and no Retry-After header documented on any page, so the sole timing signal is the Token Plan FAQ line about a roughly one-minute reset, which an agent has to find on an unrelated billing page. Error codes 1002 and 2045 are identified with a one-line message and one-line solution. No retry-after header, no exponential-backoff guidance, no request-shaping strategy, and no mention of the HTTP 429 wrapper that the OpenAPI spec defines.

## Authenticate an API request

**PASS**

PASS consensus from 3 PASS.

The pay-as-you-go path documents key creation, .env or environment storage, Anthropic variables, and SDK transmission. The full path is documented end to end: Prerequisites distinguishes the pay-as-you-go API Key from the Token Plan Subscription Key and states they are not interchangeable, both it and Token Plan Quick Start recommend storing the key as an environment variable or in a .env or config file, and the Messages API security schemes state both accepted headers and which one wins when both are sent. Two key types (API Key and Subscription Key) are clearly distinguished with separate creation paths. Every interface (Anthropic, OpenAI, AI SDK) shows exactly which environment variables to set and how to send the key.

## List video tasks with pagination

**PARTIAL**

PARTIAL consensus from 1 PASS, 2 PARTIAL.

The endpoint documents one-based pages, page size, filters, total, and seven-day scope; it uses no cursor. Seven-day retention is stated twice and reinforced by the total field description, and all four filters carry explicit enums or types, but pagination is offset-based (page_num, page_size) rather than cursor-based, and the spec gives no default page_size, no maximum page_size, and no sort-order guarantee, so an agent traversing every page must guess both its step size and whether page boundaries stay stable while new tasks are created. Filters (status, task_ids, model, task_type) and 7-day retention are fully documented. Pagination uses page_num/page_size (offset-based) rather than cursor-based pagination. The agent can traverse every page by incrementing page_num, but cursor mechanics are absent because the API does not use them.

## The receipt

> The rate limits applied to your account depend on the model and interface you use.

The table gives modality values, but Token Plan throttling is dynamic and most modality tables omit TPM or concurrency.

- [platform.minimax.io/docs/guides/rate-limits](https://platform.minimax.io/docs/guides/rate-limits)

## Agent surface notes

Initialize returned a valid MCP response with MiniMax API Docs server information.

The live Mintlify configuration disables Ask AI controls.

## Method note

This is a reading test of public documentation, not an execution test. No accounts were created and no API calls were run. The AI Agent Readiness Score counts fifteen reading votes at PASS 2, PARTIAL 1, and FAIL 0, for 30 possible points. Five agent surface checks add 10 points each. The total is 80. Consensus chips show each row majority and do not affect scoring. The panel split on 4 of five tasks. Quotes shown here were re-fetched and confirmed verbatim on 2026-08-20.

Methodology: https://docsforagents.com/methodology/

Canonical URL: https://docsforagents.com/reports/minimax-docs-ai-agent-readiness/