OpenClaw with MiniMax: provider setup and task costs

MiniMax is a bundled first-party plugin in OpenClaw, not a custom endpoint: one command, openclaw onboard --auth-choice minimax-global-api, writes the whole provider block for you. The plugin registers two provider ids with different capabilities and two credentials MiniMax says are not interchangeable, and OpenClaw's own example cost block for MiniMax-M3 carries the higher of MiniMax's two price tiers — which is the one number worth fixing before you run anything. Everything below is source-document review dated September 21, 2026; nothing here was installed, onboarded or called.

Three name traps first, because the search results mix them. The German fire-protection company Minimax at minimax.com sells fire detection and extinguishing systems, not tokens. The minimax algorithm is the game-theory rule the AI company is named after. And OpenClaw itself was renamed twice in January 2026 — Clawdbot to Moltbot on January 27 after Anthropic raised a trademark concern, then Moltbot to OpenClaw on January 30, per CNBC's account — so pages published under the old names are not current documentation. The project is MIT-licensed with no paid tier, per its repository README.

Two MiniMax surfaces, and only one of them is the chat model

The plugin registers minimax (API key) and minimax-portal (OAuth), and their capability lists differ by exactly one entry. That single difference is the most expensive thing on this page to discover late.

From OpenClaw's MiniMax provider page, read September 21, 2026. Whether the OAuth route genuinely cannot search or the table is merely incomplete is not stated on that page.

Two further boundaries decide whether the tool half of the plugin works at all. OpenClaw's MiniMax web-search tool posts to the subscription endpoint /v1/coding_plan/search, prefers MINIMAX_CODE_PLAN_KEY, and warns that ordinary model API keys may not be accepted — so a pay-as-you-go buyer cannot assume search works because chat does. And image generation always calls MiniMax's dedicated /v1/image_generation endpoint from MINIMAX_API_HOST, ignoring models.providers.minimax.baseUrl entirely. Pointing the chat base URL at anything else does not move image traffic with it.

Choose the account and the key before you touch the config

MiniMax sells two credentials. A pay-as-you-go API key meters per token. A Subscription Key carries Token Plan or Credits access and drives the OAuth route; MiniMax's own OpenClaw setup guide says a Token Plan API Key starts with sk-cp, and has you set it as MINIMAX_API_KEY — so the subscription credential is not confined to the OAuth path. MiniMax's Token Plan overview states plainly that "The Subscription Key is not interchangeable with pay-as-you-go API Keys." Regions are a second split: the international platform pairs platform.minimax.io with api.minimax.io, China pairs platform.minimax.cn with api.minimaxi.com, and MiniMax says a 401 straight after OAuth login usually means the region did not auto-match.

From MiniMax's Token Plan pricing page, read September 21, 2026. All three tiers share 5-hour rolling and weekly quota windows, and MiniMax publishes no token or request count per tier — any sentence converting a tier into tokens per month would be invented. No annual price and no "two months free" wording appears on that page. The live checkout page is client-rendered and returned nothing readable here, so treat checkout as the controlling price and confirm before you subscribe. Coverage is stated as "the full MiniMax lineup (M3 / M2.7 / image / speech)" with an exclusion list — "MiniMax H3, voice design, rapid voice cloning, etc. are not currently supported" — which matters because the plugin can call video and voice features that sit in neither list. Credits sell at 1,000 credits = $1, valid 365 days from each purchase date, and MiniMax's Token Plan overview adds that purchased Credits can automatically cover usage inside Token Plan coverage and that "Unused subscription quota does not carry over to the next billing cycle." One disclosure: OpenClaw's docs carry a referral link advertising 10% off the subscription. That is a referral discount, not a MiniMax list price.

# International, pay-as-you-go API key  -> provider id "minimax"
openclaw onboard --auth-choice minimax-global-api
# International, Token Plan over OAuth     -> provider id "minimax-portal"
openclaw onboard --auth-choice minimax-global-oauth
# China endpoints (api.minimaxi.com)
openclaw onboard --auth-choice minimax-cn-api
openclaw onboard --auth-choice minimax-cn-oauth

# Then confirm the catalog actually loaded
openclaw models list --provider minimax
openclaw models list --provider minimax-portal

The one number in the config worth fixing

The global API-key choice writes baseUrl as https://api.minimax.io/anthropic with api: "anthropic-messages" (the China choices use api.minimaxi.com); https://api.minimax.io/v1 with openai-completions is offered as optional. What they also write is a cost block, and OpenClaw's documented example for MiniMax-M3 is { input: 0.6, output: 2.4, cacheRead: 0.12, cacheWrite: 0 }. Those are MiniMax's rates above 512k input tokens.

MiniMax's pay-as-you-go list and OpenClaw's provider page, both read September 21, 2026. The tier boundary is stated as input token length, not total context, and the lower tier is badged "Permanent 50% off". Leave the example as shipped and OpenClaw reports roughly double the real rate on every request under the boundary — which is not a defect so much as a tier the example happened to pick, and OpenClaw tells you to fix it: "Update pricing values in models.json if you need exact cost tracking."

{
  "input": 0.30,
  "output": 1.20,
  "cacheRead": 0.06
}

Note what is absent. MiniMax publishes no prompt-cache write rate for M3, so this block leaves cacheWrite out rather than asserting zero. These values are MiniMax's published rates placed into OpenClaw's documented field names; the combination was not run here. One more behavioural gate to expect on the anthropic-messages path: OpenClaw disables M2.x thinking by default, because M2.x emits reasoning_content in OpenAI-style delta chunks that would leak into visible output, while M3 defaults to adaptive thinking — its two documented levels are off and adaptive — because it returns empty content with thinking disabled.

What a task costs, and what a scheduled one costs

These are arithmetic on published rates, not measured task costs and not a bill ceiling. Assume a task that bills 1,000,000 input tokens and 100,000 output tokens.

MiniMax pay-as-you-go list, USD, read September 21, 2026. On MiniMax-M2.7 that assumed task costs $0.42 uncached ($0.30 + $0.12). Serve 90% of the input from prompt cache and the same shape costs $0.204 ($0.03 + $0.054 + $0.12). The $22 Plus tier is therefore worth about 52 uncached or 107 well-cached tasks of that shape at list price — a rate comparison, not a claim that the plan delivers that many tasks, because it meters by 5-hour and weekly windows whose token counts MiniMax does not publish. Two caveats on that table: M2.7-highspeed lists double the input and output of M2.7 but the same $0.06 cache read, read once and not cross-checked; and cache mechanics — TTL, minimum cacheable prefix, whether caching is automatic at all — are not published there, so the cached figures are rate arithmetic rather than a prediction (prompt caching covers the concept). MiniMax also lists a priority tier for M3 at 1.5× the standard rate ($0.45 / $1.80 / $0.09 at or below 512k input). How OpenClaw would select it is not documented on MiniMax's price list or OpenClaw's provider page — OpenClaw's note about dropping service_tier is written for openai-completions proxy routes, not for the anthropic-messages path the MiniMax plugin prefers, so treat the priority tier as untested here rather than as unreachable.

Persistent work is where OpenClaw diverges from a coding CLI, and it is the line most budgets miss. A job that wakes every 15 minutes runs 96 times a day, about 2,880 times a month. If each run re-reads a 40,000-token context, that is 115.2M input tokens a month: $34.56 at MiniMax-M2.7's $0.30 per million uncached, or $6.91 if every one of those reads hits prompt cache at $0.06. No such job was run here, and whether a cache survives a 15-minute gap was not tested — the point is the order of magnitude between the two columns, which is what decides whether a scheduled assistant is a rounding error or the whole bill.

Kunavo serves no text-to-speech, speech-to-text or embedding model, so none of those MiniMax capabilities has a Kunavo equivalent to compare against.

Kunavo's catalog also carries no MiniMax text model, and its MiniMax video entries are reserved rather than enabled — the Hailuo API guide explains why. This page is therefore not a route to MiniMax-M3 or MiniMax-M2.7 through Kunavo.

First task, and the errors worth recognising

Run openclaw models list --provider minimax (or --provider minimax-portal) before anything paid: it proves the catalog loaded, not that a task works. Then give it one bounded job and read the charge your MiniMax account recorded for it.

For the quota reading itself, OpenClaw polls https://api.minimax.io/v1/token_plan/remains and inverts MiniMax's usage_percent field, because MiniMax reports remaining quota rather than consumed quota. Outside OpenClaw, MiniMax documents its own CLI: npm install -g mmx-cli, then mmx quota. One version caveat worth carrying: MiniMax's vendor-side guide targets OpenClaw v2026.7.1-2, while the release GitHub flags as Latest is v2026.9.5, published September 19, 2026 (gh release list -R openclaw/openclaw, September 21, 2026). The repo cuts tags on parallel version lines — v2026.7.35 carries a later tag timestamp than v2026.9.5 — so "the newest tag" is not the same question as "the latest release". Where the two documents differ, OpenClaw's docs are the fresher source for config shape and MiniMax's page is the authority on which credential is required.

Direct, subscription, or a second provider

OpenClaw's custom-provider documentation sets the boundaries for that last row, and each one carries its own scope. Only openai-completions and anthropic-messages are documented as api types. A model's cost block defaults to { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 } when omitted, so an unconfigured custom provider reports $0 spend — that rule is unqualified. The two compatibility rules are not: on openai-completions against a non-native host, OpenClaw forces supportsDeveloperRole: false and proxy-style routes drop service_tier; on anthropic-messages against anything other than a canonical Anthropic endpoint, it suppresses implicit Anthropic beta headers. Neither set applies to both protocols. The page's examples all show a baseUrl that already carries its own path suffix, but it does not state whether OpenClaw appends one — so include the suffix and verify rather than assume.

For the same assumed shape as above — 1,000,000 uncached input tokens and 100,000 output tokens — here is what a Kunavo-side primary costs at live catalog rates, so the two sides of a fallback pair are comparable in one unit.

Illustrative token arithmetic, not a measured OpenClaw task and not a bill ceiling: it excludes cache operations, tool calls and the machine the assistant runs on. Kunavo's catalog amount is a billing floor rather than a cap — when the upstream reports its charge, the bill is the greater of catalog cost and upstream cost times the applicable markup. The minimum top-up is $10 in prepaid credit, a funding minimum rather than a task fee or a subscription; see billing details. The OpenClaw-side configuration for a Kunavo provider entry is written up at the OpenClaw provider comparison; that is a documented configuration reference, not a runtime test, and whether OpenClaw appends anything to the base URL is not stated in its custom-provider docs, so confirm it in your own install before you trust it. Create a Kunavo account when you want that second entry funded, and OpenClaw pricing covers the hosting and tool layers this page leaves out.

FAQ

Every source linked above was fetched on September 21, 2026: MiniMax's pay-as-you-go price list, Token Plan pricing page, Token Plan overview, rate-limit page and its own OpenClaw setup guide; OpenClaw's MiniMax provider, MiniMax search-tool and custom-provider documentation; and the openclaw/openclaw release list, read with gh release list. The two January 2026 rename dates come from the CNBC account linked above and were cross-checked against contemporaneous reporting of the same two renames. Nothing on this page was runtime tested: no OpenClaw install, no onboarding run, no MiniMax key issued and no request sent. Kunavo token rates come from the live catalog, and every dollar figure here is arithmetic on published rates.