OpenClaw can burn through a plan faster than a normal chat app. A repo-maintenance agent reads files, calls tools, retries commands, keeps context, and may run for minutes. That turns plan selection into a routing problem, not a brand preference.
The recommendations use official vendor docs and OpenClaw provider docs checked on July 31, 2026. Fire Pass isn't in the buying table because Fireworks documents it as invite-only, limited to non-production coding use, and restricted to enabled models on the Fire Pass page. If you already have it, treat it as bonus capacity, not the base of a new OpenClaw budget.[1]
The credible buying set right now is:
- Z.AI GLM Coding Plan for the lowest public monthly sticker price among OpenClaw-documented plans compared here, if you accept GLM-first routing and best-effort OpenClaw scheduling.
- MiniMax Token Plan for the practical default starter when you want a shared quota pool without Z.AI's model-weighted burn.
- Qwen Cloud Coding Plan for the broadest single subscription model menu (verify exact model availability and quota before relying on it).
- OpenAI as the premium escape hatch, through Codex auth or direct API billing.[2][3][4][5]
None of these are unlimited production SLA lanes. MiniMax individual interactive limits, Qwen no-batch/script constraints, and Z.AI fair-use plus best-effort OpenClaw scheduling all sit next to sticker prices. Run
openclaw models listbefore routing production traffic.
What matters
An OpenClaw plan has to survive agent traffic. Five details matter more than benchmark drama:
- Does OpenClaw document the route?
- Does the plan count tokens, prompts, requests, or subscription quota?
- Does the limit reset gradually, weekly, monthly, or not at all?
- Is the plan limited to personal or interactive coding-tool use?
- Where does metered pricing start?
Current starting routes to verify:
- MiniMax Token Plan:
minimax-portal/MiniMax-M3for OAuth orminimax/MiniMax-M3for API-key setup. - Qwen Cloud Coding Plan:
qwen/qwen3.5-plus, the bundled default in OpenClaw docs. - Z.AI GLM Coding Plan:
zai/glm-5.2, with GLM-4.7 as the recommended fallback. - OpenAI:
openai/gpt-5.6-solfor fresh Codex-backed setup, oropenai/gpt-5.6for direct API-key billing. Useopenai/gpt-5.5only when the account doesn't expose GPT-5.6.
Always run openclaw models list after setup. Vendor plan pages and OpenClaw provider catalogs don't always move at the same speed.[6][7][8][5]
馃挕 Key insight: OpenClaw plan selection is routing, not brand picking. Each lane fails differently: rolling quota, hard request stop, queue delay, or rising metered spend.
MiniMax: practical starter lane
MiniMax Token Plan lists Plus at $20/month, Max at $50/month, and Ultra at $120/month. That is not the lowest sticker price in this comparison. Z.AI Lite starts at $18/month. MiniMax still earns the practical-starter slot when you want a shared subscription quota across text and multimodal tools without Z.AI's model-weighted burn and secondary OpenClaw scheduling. All three MiniMax tiers use 5-hour rolling and weekly quota windows, cover API Platform models through a Subscription Key, and share quota across supported text, image, speech, and music resources. Purchased Credits are priced at 1,000 credits = $1, and Token Plan quota is used before eligible purchased Credits.[2]
OpenClaw now defaults its MiniMax provider to MiniMax M3. OAuth setups use minimax-portal/<model>, while API-key setups use minimax/<model>. M2.7 and M2.7 high-speed routes still exist, but M3 is the clean default.[6]
Fit and failure mode
Best fit: solo OpenClaw users who want a simple practical starter and are happy making MiniMax M3 the routine lane.
Watch out for: MiniMax says Token Plan is meant for individual, interactive developer use and recommends pay-as-you-go for production. During peak traffic, it may apply dynamic rate limiting, and the same subscription quota is shared across tools.[9]
Qwen Cloud: broadest bundle
Qwen Cloud Coding Plan lists $50/month with 6,000 requests per 5 hours, 45,000 per week, and 90,000 per month. Its current recommended model list includes qwen3.7-plus, kimi-k2.5, glm-5, and MiniMax-M2.5; the broader allowlist also includes qwen3.6-plus, qwen3.5-plus, qwen3-max-2026-01-23, qwen3-coder-next, qwen3-coder-plus, and glm-4.7.[3]
OpenClaw's bundled Qwen provider still documents qwen/qwen3.5-plus as the default route, and says availability can vary by endpoint and billing plan even when a model appears in the bundled catalog.[7]
Two setup details are easy to get wrong:
- Use a plan-specific
sk-sp-...key. - Use a coding endpoint such as
coding-intl.dashscope.aliyuncs.com/v1for the global OpenAI-compatible path.[10][7]
Best fit: users who want one subscription lane that can expose Qwen, Kimi, GLM, and MiniMax-family options.
Watch out for: Qwen Coding Plan is for interactive coding tools, not scripts or batch calls. When request quota runs out, Qwen says calls fail directly rather than falling back to pay-as-you-go.[10]
Z.AI: lowest sticker price, GLM-first lane
Z.AI's GLM Coding Plan starts at $18/month, the lowest public monthly sticker price among OpenClaw-documented plans compared in this table, and supports GLM-5.2, GLM-5-Turbo, and GLM-4.7. As verified on July 31, 2026, its published allowance is credits, not prompt counts: Lite has 2,000 credits per 5 hours and 10,000 per week, Pro has 12,000 and 60,000, and Max has 28,000 and 140,000.[4][11]
Credit burn is model-weighted: Z.AI divides the weighted input, cached-input, and output tokens by 10,000. Current multipliers are 6.9/1.7/24 for GLM-5.2, 5.7/1.5/21 for GLM-5-Turbo, and 4.6/1.2/16 for GLM-4.7. Model usage costs 50% of the standard credit rate off-peak; peak hours are Monday through Friday, 14:00 to 18:00 Singapore time. Route routine work to GLM-4.7 when its capability is sufficient, then recheck this dated policy before planning a sustained agent budget.[4][11]
OpenClaw scheduling
Z.AI's OpenClaw guide is unusually explicit: OpenClaw traffic uses secondary scheduling and best-effort delivery, while coding-agent tasks get priority under load. Heavy load can trigger dynamic queueing and fair-use limits.[8]
Best fit: users who specifically want GLM and are comfortable routing simple tasks away from GLM-5.2 to preserve quota.
Watch out for: this isn't an unlimited low-latency lane. It's a GLM subscription with scheduling and model-weighted quota behavior.
OpenAI: premium escape hatch
GPT-5.6 is generally available through the OpenAI API, not a limited preview.[12] OpenClaw's OpenAI docs point fresh Codex-backed setups at openai/gpt-5.6-sol. Direct API-key setups use openai/gpt-5.6, which currently resolves to the Sol tier. Accounts without GPT-5.6 access can select openai/gpt-5.5 explicitly.[5]
OpenAI's GPT-5.6 Sol model page lists standard short-context rates at $5.00 per 1M input tokens, $0.50 cached input, and $30.00 output. Requests above 272K input tokens cost 2x for input and 1.5x for output, which yields $10.00 input, $1.00 cached input, and $45.00 output per 1M tokens. Treat those as calculator values for direct API use, separate from Codex subscription access.[12]
Fit and spend risk
Best fit: hard reasoning, stubborn debugging, architectural decisions, and high-impact changes.
Watch out for: output-heavy and long-context agent runs make the direct API bill grow quickly. OpenAI shouldn't be the default route for every file read, directory scan, or retry.
Head-to-Head
Among plans documented by OpenClaw in this comparison, Z.AI has the lowest current public monthly sticker price at $18, but OpenClaw traffic runs through best-effort scheduling and higher models burn quota faster. MiniMax costs $20 to start and is the practical default when you want a shared quota pool without that weighted-burn tradeoff, though peak-hour shaping still matters. Qwen Cloud is the broadest one-provider model bundle, but it can hard-stop at quota and depends on the exact model allowlist. OpenAI is the premium fallback, but direct API spend grows fastest under long agent loops.
| Option | Quota or price shape | OpenClaw route | Main restriction | Failure mode |
|---|---|---|---|---|
| MiniMax Token Plan | $20/month entry, 5-hour and weekly quota windows, shared subscription quota | minimax-portal/MiniMax-M3 or minimax/MiniMax-M3 | Individual interactive developer use; peak-hour shaping can apply | Quota pressure recovers gradually, or peak traffic slows work |
| Qwen Cloud Coding Plan | $50/month, 5-hour, weekly, and monthly request quotas | qwen/qwen3.5-plus default in OpenClaw docs | Plan-specific key and coding endpoint; no pay-as-you-go fallback after quota | Requests hard-stop until quota resets |
| Z.AI GLM Coding Plan | $18/month entry; Lite 2,000/10,000 credits per 5-hour/weekly window, with model-weighted burn | zai/glm-5.2, with GLM-4.7 fallback | OpenClaw traffic uses secondary scheduling and fair-use limits | Queue delay rises under load |
| OpenAI | Direct API token rates; GPT-5.6 Sol short context $5/$0.50/$30 per 1M input/cached/output, above 272K input $10/$1/$45 | openai/gpt-5.6-sol for Codex auth or openai/gpt-5.6 for API keys | Generally available API; access still depends on account or product availability, metered billing, and any separate Codex subscription path[12] | Spend rises continuously, especially on output-heavy long-context loops |
The short comparison hides the real operational difference: each option fails differently.
If your primary lane hard-stops at quota, prewire fallback before production traffic hits the cap. Queueing providers need visible delay and retry policy; metered lanes need an escalation reason every time the router opens them.
Practical routing plan
Most users don't need four active providers. Start with one subscription lane and one premium escape hatch:
- Pick Z.AI if lowest monthly sticker price among these OpenClaw-documented plans matters most and GLM is acceptable.
- Pick MiniMax if you want the practical default starter with a shared quota pool.
- Pick Qwen Cloud if you want the broadest subscription model menu.
- Keep OpenAI for hard work and rescue paths.
Add a second subscription only when you can name the exact failure mode: Qwen quota hard stop, MiniMax peak-hour shaping, Z.AI queueing, or missing model coverage.
Guardrails
OpenClaw routing needs policy, not model ids alone. Route routine reads, search, summaries, and small edits to the subscription lane; trim or summarize before escalation; limit automatic retries per task class; and log why each fallback model was used.
The common mistake is treating "fixed price" as unlimited capacity. These plans still have scopes, queues, rolling windows, exact model allowlists, and weighted burn rates. Read the operating limit before routing agent loops through it.
鈿狅笍 Common mistake: Opening the premium API lane for every file read or retry. Log escalation reasons, cap automatic retries per task class, and trim context before you burn metered tokens on routine repo scans.
Recommendation
For most OpenClaw users:
- Start with Z.AI Lite if the lowest current public monthly sticker price among OpenClaw-documented plans here matters most and GLM is acceptable.[8][4]
- Start with MiniMax Plus if you want the practical default starter with a shared quota pool.[2][6]
- Choose Qwen Cloud Coding Plan if you want the broadest bundled provider route.[3][7]
- Keep OpenAI as the premium fallback, not the default lane.[13][12]
Best stack: one good subscription lane, an optional second lane only after you see a real failure mode, and one premium escape hatch. The routing policy matters as much as the plan. A well-priced agent with bad retry logic will still waste quota; a strong model with no routing discipline will still waste money.