--- id: PX-RESEARCH-TQLVNIM7 type: research title: MiniMax M Plan Capabilities and APIs --- ## Evidence boundary Observed 2026-11-09; official global-service documentation only. Scope: full capability/API-family inventory, including Explore-specific restrictions. No credentials inspected, authenticated requests made, packages installed, or generation allowance consumed. Published API existence ≠ M Plan entitlement ≠ account access ≠ successful execution. Sources establish capability-level inclusion, not an exhaustive endpoint/model entitlement matrix. ## Plan capabilities Sources: [M Plan introduction](https://platform.minimax.io/docs/m-plan/intro.md), [FAQ](https://platform.minimax.io/docs/m-plan/faq.md), [usage rules](https://platform.minimax.io/docs/m-plan/usage-rules.md). | Capability | Go | Explore | Build | | --- | --- | --- | --- | | Relative included usage | 1× | 3× | 7.5× | | Text model | M3.1 Flash | M3.1 Flash | M3.1 Flash | | Image/audio generation | Included | Included | Included | | Search/multimodal understanding | Included | Included | Included | | H3 video | Not included | Included | Included | | Non-video windows | 5 hours + weekly | 5 hours + weekly | 5 hours + weekly | | Video window | No included video | Weekly only | Weekly only | All tools and included modalities share subscription usage; video consumes weekly allowance without consuming the non-video 5-hour allowance. Windows start with first use, reset independently after 5 hours/7 days; unused allowance does not accumulate. Consumption varies with model, context, reasoning effort, task complexity, and modality; relative usage is not a fixed request/token guarantee. M Plan targets individual interactive use; vendor recommends pay-as-you-go for production. Legacy Token Plan subscriptions retain existing benefits while auto-renewal remains enabled; do not apply the current M Plan matrix to them without checking subscription identity. ## M Plan Explore **No Explore-only endpoint prohibition is documented in the reviewed API/setup sources.** This establishes absence of a published restriction in those sources, not proven parity across every public API. - Versus Go: 3× relative included usage; adds included H3 video. - Versus Build: same named capability/model families; 3× versus 7.5× Go allowance, so Explore has 40% of Build's relative usage. - No documented Explore-specific reduction in context length, output length, image resolution, video duration, voice count, or API protocol support was found. - Exact numeric Explore RPM/TPM, concurrent-task ceilings, and per-endpoint/model entitlements remain unpublished in reviewed M Plan sources. - Dynamic traffic throttling remains separate from allowance exhaustion; available allowance does not guarantee immediate admission. - Public API model variants and ancillary operations remain unverified for Explore, including H3 Highspeed/Lite/Max, Context-IR, regeneration, ASR, voice cloning/design, music, and file management. - Lower quota can block an otherwise included capability; endpoint permission and remaining quota are different questions. Conclusion: Explore explicitly includes text, image, audio, search, understanding, and H3 video; upgrading to Build is evidenced as more usage, not a documented blanket unlock of otherwise inaccessible API families. ## Authentication and billing Sources: [FAQ](https://platform.minimax.io/docs/m-plan/faq.md), [general setup](https://platform.minimax.io/docs/m-plan/other-tools.md), [pricing](https://platform.minimax.io/docs/pricing/overview.md). | Credential | Resource charged | Acquisition | | --- | --- | --- | | M Plan API key, also called Subscription Key | Included quota, then eligible purchased Credits | Plan Details | | Standard API Key | Pay-as-you-go account balance | API Platform API Keys | Both use API authentication; subscription ownership does not convert a standard API Key into subscription billing. An M Plan key does not automatically charge account balance when quota runs out, but can automatically consume purchased Credits. Included allowance is used before Credits; earliest-expiring Credits are consumed first; Credits expire one year after purchase. Credits can also be bought without an M Plan subscription; request success alone therefore does not prove included subscription entitlement. FAQ says priced endpoints debit included quota at their corresponding list price, but also explicitly directs resources outside M Plan coverage to pay-as-you-go keys. Neither a published PAYG price nor a Bearer-auth schema proves inclusion or exclusion. Separate Audio/Video Packages are not M Plan; their rate limits, voice slots, and model coverage cannot be imported into Explore. Global and mainland-China service regions differ; official CLI supports explicit region selection when automatic key-region detection fails. ## API inventory Global HTTPS origin: `https://api.minimax.io` unless explicitly stated otherwise. Routes below identify published interfaces, not blanket Subscription Key permission. References: [API overview](https://platform.minimax.io/docs/api-reference/api-overview.md), [complete index with OpenAPI/AsyncAPI schemas](https://platform.minimax.io/docs/llms.txt). ### Text, reasoning, model discovery Sources: [Anthropic compatibility](https://platform.minimax.io/docs/api-reference/text-anthropic-api.md), [OpenAI compatibility](https://platform.minimax.io/docs/api-reference/text-openai-api.md), [Responses](https://platform.minimax.io/docs/api-reference/responses-create.md). | Operation | Published route | | --- | --- | | Anthropic Messages | `POST /anthropic/v1/messages` | | Anthropic token estimate | `POST /anthropic/v1/messages/count_tokens` | | OpenAI Chat Completions | `POST /v1/chat/completions` | | OpenAI Responses | `POST /v1/responses` | | Responses token estimate | `POST /v1/responses/input_tokens` | | OpenAI model list/detail | `GET /v1/models`; `GET /v1/models/{model_id}` | | Anthropic model list/detail | `GET /anthropic/v1/models`; `GET /anthropic/v1/models/{model_id}` | | Native text interface | `POST /v1/text/chatcompletion_v2` | M Plan setup explicitly selects `MiniMax-M3.1-Flash`; `MiniMax-M3` appearing in platform API references does not establish its included entitlement. `MiniMax-M3.1-Flash-Preview` is retired; vendor says requests using that alias are served by `MiniMax-M3.1-Flash`. Platform references advertise 1,000,000 total input/output tokens; a client example configured for 512K is not evidence of an Explore-specific cap. Anthropic SDK base URL is `https://api.minimax.io/anthropic`; OpenAI SDK base URL is `https://api.minimax.io/v1`. Streaming, client tool calls, interleaved reasoning, and multimodal input are documented; protocol compatibility is partial rather than full upstream parity. M3.1 Flash always thinks; disabling thinking or requesting effort `none` returns 400. Effort accepts `low`, `medium`, `high`, `xhigh`, `max`; omitted effort defaults to `max`. Thinking counts toward output limits; insufficient output budget can produce no visible answer. Preserve reasoning and tool-call blocks unchanged across tool turns. Anthropic parameters including temperature, top-p, top-k, stop sequences, MCP servers, and context management are ignored; OpenAI sampling/penalty options also have documented omissions. Model discovery and token estimation are published APIs; their M Plan eligibility and billing are not exhaustively specified. ### Search and understanding Sources: [server tools](https://platform.minimax.io/docs/guides/server-tools.md), [M Plan MCP](https://platform.minimax.io/docs/guides/token-plan-mcp-guide.md), [newer search-only MCP guide](https://platform.minimax.io/docs/m-plan/mcp-guide.md), [CLI](https://platform.minimax.io/docs/m-plan/minimax-cli.md). - M Plan MCP documents `web_search(query)` and `understand_image(prompt, image_source)` using a Subscription Key. - Image understanding accepts local paths or HTTP(S) URLs; documented formats: JPEG, PNG, WebP. - Newer M Plan MCP page documents only `web_search`; neither page establishes removal of understanding from all installed versions. - Official CLI separately exposes search and vision; documentation recommends CLI over MCP for simpler setup. - Model-hosted search uses Anthropic Messages or OpenAI Responses, not a documented Chat Completions server-tool route. - Anthropic search type: `web_search_20250305`; Responses search type: `web_search`. - Anthropic search may return `pause_turn`; continuation must preserve returned content rather than falsely report completion. - Pricing explicitly says M Plan API-vlm usage debits included quota at its PAYG price, then eligible Credits. - No stable standalone search/understanding REST contract was established from reviewed M Plan pages; MCP tool names must not be invented into REST routes. ### Images Sources: [text-to-image](https://platform.minimax.io/docs/api-reference/image-generation-t2i.md), [image-to-image](https://platform.minimax.io/docs/api-reference/image-generation-i2i.md). Both operations use `POST /v1/image_generation`. Public references document `image-01`; image-to-image schema also enumerates `image-01-live`, without establishing M Plan variant eligibility. Text prompt limit: 1,500 characters; batch size: 1–9 images; aspect ratio or explicit supported dimensions; URL or Base64 output. Reference-image mode documents character/portrait subject references, not arbitrary masked/inpainting edits. Reference inputs accept public URLs or Base64 data URLs; JPG/JPEG/PNG, less than 10 MB. Image generation is included at capability level; text-to-image is demonstrated by official M Plan CLI, but every image-to-image option/model remains unverified under subscription billing. ### Video Sources: [H3 creation](https://platform.minimax.io/docs/api-reference/video-generation-v2-create.md), [query](https://platform.minimax.io/docs/api-reference/video-generation-v2-query.md), [cancel/delete](https://platform.minimax.io/docs/api-reference/video-generation-v2-delete.md), [legacy APIs](https://platform.minimax.io/docs/llms.txt). | Operation | Published route | | --- | --- | | H3 generation | `POST /v2/video_generation` | | Query one task | `GET /v2/query/video_generation/{task_id}` | | List tasks | `GET /v2/query/video_generation` | | Cancel queued task/delete terminal record | `DELETE /v2/video_generation/{task_id}` | | Context interpretation | `POST /v2/h3_context_ir` | | 768P → 2K regeneration | `POST /v2/video_regeneration` | | Legacy task query | `GET /v1/query/video_generation` | | Legacy download/file metadata | `GET /v1/files/retrieve` | | Template Video Agent create/query | `POST /v1/video_template_generation`; `GET /v1/query/video_template_generation` | H3 API accepts text/image/video/audio content; public variants include H3, Highspeed, Lite, Max. Tasks support text-to-video, image-to-video, first/last frames, and multimodal references, subject to model-specific constraints. Single-task query/list cover the last seven days. Cancellation of queued work differs from deletion of a completed/failed record; deletion is not evidence of cancellation/refund. Index also retains legacy text-to-video, image-to-video, first/last-frame, and subject-reference creation APIs; these are separate from H3 V2, not evidence of M Plan inclusion. Explore/Build include H3 at family level; variant, regeneration, Context-IR, template, and legacy-model coverage remain unresolved. PAYG pricing includes input-media charges as well as output video; regeneration can bill original inputs again. ### Speech and voice Sources: [HTTP TTS](https://platform.minimax.io/docs/api-reference/speech-t2a-http.md), [WebSocket TTS](https://platform.minimax.io/docs/api-reference/speech-t2a-websocket.md), [bidirectional TTS](https://platform.minimax.io/docs/api-reference/speech-t2a-websocket-bidi.md), [ASR](https://platform.minimax.io/docs/api-reference/speech-to-text.md), [overview](https://platform.minimax.io/docs/api-reference/api-overview.md). | Operation | Published route | | --- | --- | | HTTP TTS, optional streaming | `POST /v1/t2a_v2` | | Sentence-fed WebSocket TTS | `wss://api.minimax.io/ws/v1/t2a_v2` | | Incremental bidirectional TTS | `wss://api.minimax.io/ws/v1/t2a_v2_bidi` | | Async long TTS create/query | `POST /v1/t2a_async_v2`; `GET /v1/query/t2a_async_query_v2` | | Speech-to-text | `POST /v1/speech_to_text` | | Upload clone/prompt audio | `POST /v1/files/upload` with operation-specific purpose | | Voice clone/design | `POST /v1/voice_clone`; `POST /v1/voice_design` | | List/delete voices | `POST /v1/get_voice`; `POST /v1/delete_voice` | Public TTS families include speech-2.8-hd/turbo, speech-2.6-hd/turbo, speech-02-hd/turbo; per-model subscription eligibility is not enumerated. Synchronous TTS supports up to 10,000 characters; async guide distinguishes 100,000-character direct/file inputs from up to 1,000,000 aggregate characters via ZIP. ASR documents streaming, diarization, and subtitle export; audio-generation inclusion does not establish transcription entitlement. Bidirectional TTS supports task cancellation without closing the session; request-final audio, sentence completion, and session completion are distinct signals. Cloning requires individual/enterprise verification; cloned/designed voices unused for synthesis within seven days are deleted. PAYG clone/design activation fees apply on first synthesis use, with separately priced previews; corresponding M Plan treatment remains unknown. Speech synthesis is demonstrated by M Plan CLI; ASR, all transport variants, voice creation/management, and ancillary charges require separate entitlement evidence. ### Music and lyrics Sources: [music](https://platform.minimax.io/docs/api-reference/music-generation.md), [lyrics](https://platform.minimax.io/docs/api-reference/lyrics-generation.md), [cover preprocessing](https://platform.minimax.io/docs/api-reference/music-cover-preprocess.md), [availability/pricing](https://platform.minimax.io/docs/pricing/overview.md). | Operation | Published route | | --- | --- | | Song generation | `POST /v1/music_generation` | | Lyrics generation/edit/continuation | `POST /v1/lyrics_generation` | | Cover preprocessing | `POST /v1/music_cover_preprocess` | Catalog includes music-3.0 and older music APIs; indexed references are not evidence of new-user availability. Vendor states paid music/lyrics APIs closed to new users from August 20, 2026; existing paying users retain access, while free music-generation APIs are discontinued. No explicit M Plan music/lyrics/cover grant was found; generic audio inclusion must not be interpreted as one. ### Files and usage Sources: [file APIs](https://platform.minimax.io/docs/api-reference/file-management-upload.md), [FAQ usage endpoint](https://platform.minimax.io/docs/m-plan/faq.md), [CLI](https://platform.minimax.io/docs/m-plan/minimax-cli.md). | Operation | Published route | | --- | --- | | Upload | `POST /v1/files/upload` | | List | `GET /v1/files/list` | | Metadata | `GET /v1/files/retrieve` | | Download content | `GET /v1/files/retrieve_content` | | Delete | `POST /v1/files/delete` | | M Plan remaining usage | `GET https://www.minimax.io/v1/token_plan/remains` | Usage endpoint takes Bearer M Plan API key; no response schema or stability guarantee was established from the FAQ example. Official CLI exposes quota inspection without generation. File operations support downstream workflows; that role alone does not establish permission for every file purpose or operation under M Plan. Async TTS overview says returned file download URLs expire after nine hours. Official CLI supports text, image, video, speech, vision, and search; availability depends on plan and CLI version. ## Conclusions - M Plan is a multimodal subscription/API credential path, not merely a coding-chat allowance. - Explore's documented differences are quota and H3 inclusion versus Go; no separate Explore API blacklist was found. - Capability discovery must keep documented support, account entitlement, quota, rate admission, and execution outcome distinct. - Subscription-backed success can consume prepaid Credits; strict subscription-only spending cannot be promised without an enforceable Credits boundary. - Public model catalog, PAYG pricing, and generic authentication examples are insufficient evidence for complete M Plan API access. ## Unresolved questions - Exact Explore endpoint/model matrix, including H3 variants, ASR, voice activation, music, files, model discovery, and token estimation? - Exact included usage quantities, Explore RPM/TPM/concurrency ceilings, and any tier-specific media constraints not published in reviewed sources? - Can automatic Credits overflow be disabled or bounded before a request without changing credentials to PAYG? - Does remaining-usage API distinguish included quota, Credits, both reset windows, video consumption, and dynamic request limits with stable fields? - Which MCP/CLI versions expose understanding versus search-only behavior, and which underlying public APIs do they call? - What usage is charged for token estimation, failed generation, cancellation, retries, and ancillary operations?