A Slack ping that says "summarize last night's CI failures and open a PR" looks like one request. Inside OpenClaw it can become file reads, test retries, a browser check, and extra model calls when a tool result is ugly. OpenClaw is a self-hosted personal assistant. Its Gateway wires hosted models to messaging channels, files, browser actions, scheduled jobs, and other tools.[1]
That traffic pattern makes the best plan a billing and capacity choice. A ChatGPT subscription, a provider coding plan, and a direct API key are different products. They can expose the same model name while using different credentials, quotas, and invoices.
Official provider and OpenClaw docs were checked on August 13, 2026. Current recommendation:
- Already pay for ChatGPT Plus or higher with Codex access? Start with OpenAI Codex OAuth. It spends subscription quota, not OpenAI Platform API credit.[2][3]
- Buying the cheapest public coding plan? Z.AI Lite starts at $18/month, but OpenClaw traffic is best-effort and quota burn is model-weighted.[4][5]
- Want a MiniMax-first lane with a shared rolling bar? MiniMax Token Plan Plus is $20/month.[6]
- Want the broadest published model allowlist in one coding plan? Qwen Cloud is $50/month and covers selected Qwen, Kimi, GLM, and MiniMax models.[7]
- Use direct OpenAI API billing only when you want metered spend. It isn't included with a ChatGPT subscription.[3][8]
馃挕 Key insight: Choose credential type before model.
openai/gpt-5.6-solcan ride Codex subscription OAuth or an ordered OpenAI API-key profile. The model ID doesn't tell you who pays.[3]
Compare separate billing products
Keep one sample turn in mind: 20,000 fresh input tokens, 10,000 cached input tokens, and 4,000 output tokens. That's a plausible CI-summary pass with tool results still in context. The table prices plans. The figure below prices that turn, so you can see why leftover quota from one vendor isn't a leftover request from another.
| Option | Current price or meter | OpenClaw auth and route | Capacity boundary | What happens at the boundary |
|---|---|---|---|---|
| ChatGPT Plus with Codex | $20/month; Codex included in the ChatGPT plan | OpenAI OAuth, usually openai/gpt-5.6-sol when the account catalog exposes it | Subscription usage window and ChatGPT credit balance; no Platform API credit included | Wait for the window, buy extra ChatGPT credits, or rotate to another configured OpenAI auth profile |
| MiniMax Token Plan Plus | $20/month | Coding Plan OAuth: minimax-portal/MiniMax-M3; API key: minimax/MiniMax-M3 | Shared usage bar with rolling 5-hour and weekly quota windows | Capacity returns through rolling windows; eligible purchased credits can cover overflow if enabled |
| Qwen Cloud Coding Plan | $50/month | Plan key through qwen/*; qwen/qwen3.7-plus is a current recommended option | 6,000 requests per 5 hours, 45,000 weekly, 90,000 monthly | Calls fail until quota returns; no automatic pay-as-you-go fallback |
| Z.AI Lite | $18/month | Coding Plan key through zai/*; zai/glm-5.3 is current | 2,000 credits per 5 hours and 10,000 weekly; model-weighted burn | Dynamic queueing, fair-use limits, or quota exhaustion |
| OpenAI API | No required monthly plan; per-token billing | OpenAI API-key profile through openai/* | Project budget, rate limits, and available credit | Requests continue while allowed and metered spend grows |
Prices and limits come from each provider's current plan or pricing pages.[2][9][6][7][4][10] Don't compare request counts with credits, tokens, or a shared usage bar. Qwen says a simple coding prompt may consume 5 to 10 requests and a complex task can consume 30 or more. Z.AI applies different credit multipliers by model and token type. MiniMax says its console usage bar is source of truth because actual deduction follows resource consumption.[7][4][11]
The reset clocks also disagree. MiniMax uses rolling 5-hour and weekly windows, and unused included quota doesn't carry into the next billing cycle. Qwen's 5-hour quota is a sliding window (each request returns 5 hours after it was used), weekly quota resets Monday 00:00 UTC+8, and monthly quota resets on the subscription anniversary at 00:00 UTC+8. Z.AI refreshes 5-hour credits 5 hours after consumption and resets weekly credits every 7 days from signup.[6][7][4]
OpenAI: use subscription access before adding a meter
OpenClaw supports two distinct OpenAI paths:
- Codex OAuth uses a ChatGPT/Codex subscription and its quota. Codex isn't a separate SKU. It's included in ChatGPT Free, Go, Plus, Pro, Business, and Enterprise plans. Plus is $20/month and is the usual paid starting point.[2]
- Platform API-key auth uses direct API billing.
Both use the canonical openai/* provider route. Fresh subscription setup selects openai/gpt-5.6-sol only when no primary model is already configured. OpenClaw currently labels GPT-5.6 a limited preview and says exact Sol, Terra, and Luna availability depends on the signed-in account. If GPT-5.6 is absent, choose openai/gpt-5.5 explicitly.[3]
On the sample CI-summary turn, GPT-5.6 Sol standard API rates are $5 per 1M input tokens, $0.50 per 1M cached input tokens, and $30 per 1M output tokens. That mix costs $0.225. Requests with more than 272K input tokens use 2x input and 1.5x output for the full request ($10, $1, and $45 per 1M). A hundred similar turns are about $22.50 before long-context multipliers. Those rates don't describe Codex subscription quota.[10]
Plus and Pro users who hit included Codex limits can buy extra ChatGPT credits without upgrading the plan. That still isn't Platform API credit. OpenClaw can also keep serving if you add an API-key profile to auth.order.openai. That's a different invoice.[9][12]
Best fit: existing ChatGPT subscribers who can serve routine OpenClaw work from Codex quota, or users who deliberately accept API spend for harder tasks.
Main risk: adding an OpenAI API-key profile after OAuth can create an automatic paid overflow path. OpenClaw rotates auth profiles inside a provider before moving to configured model fallbacks. Keep the API key out of auth.order.openai unless that spend is intentional.[12]
MiniMax: simple rolling plan
MiniMax Token Plan lists Plus at $20/month, Max at $50/month, and Ultra at $120/month. All tiers use 5-hour rolling and weekly quota windows. Quota is shared across supported text, image, speech, and music resources. Simple work deducts less than long-context, multi-turn, multimodal, or complex agent work; console usage bar is source of truth. Purchased credits cost 1,000 credits per $1 and can cover eligible overflow after included quota.[6][11]
During peak traffic MiniMax may apply dynamic rate limits. FAQ's typical window is weekdays 15:00-17:30, with Plus sized for about 3-4 agents, Max 4-5, and Ultra 6-7. Unused included quota doesn't carry into next billing cycle.[11]
Subscription Keys and normal pay-as-you-go API keys are separate credentials. In OpenClaw, Coding Plan OAuth uses minimax-portal/MiniMax-M3; API-key setup uses minimax/MiniMax-M3. Don't paste a normal API key into a Subscription Key flow or infer plan entitlement from the model route.[11][13]
Best fit: one MiniMax-first lane with capacity that returns gradually rather than at one fixed reset.
Main risk: MiniMax describes Token Plan as individual, interactive developer access and recommends pay-as-you-go for production. Every supported modality draws from the same quota, so image or speech jobs steal capacity from coding turns.[11]
Qwen Cloud: broad allowlist, hard quota stop
Qwen Cloud Coding Plan costs $50/month. Its current recommended set includes qwen3.7-plus, kimi-k2.5, glm-5, and MiniMax-M2.5; the larger published allowlist includes additional Qwen, GLM, and coding models.[7]
Qwen is now an official external OpenClaw plugin, not a bundled provider. Coding Plan setup uses a plan-specific sk-sp-... key and the Coding Plan endpoint. qwen/qwen3.7-plus, qwen/qwen3.6-plus, and qwen/qwen3.5-plus are supported on the plan, while some catalog models are Standard pay-as-you-go only. OpenClaw's example config still shows qwen/qwen3.5-plus as the default, so set qwen/qwen3.7-plus yourself if that's the model you want.[14][15]
Best fit: users who will exercise several supported model families enough to justify one broader subscription.
Main risk: Qwen restricts Coding Plan to personal, interactive coding tools. Scripts and batch calls are excluded. Quota exhaustion fails directly instead of falling back to pay-as-you-go, and peak traffic may reduce concurrency.[15]
Z.AI: lowest entry price, weighted credits
Z.AI Lite starts at $18/month and includes 2,000 credits per 5 hours plus 10,000 weekly. Pro includes 12,000 and 60,000; Max includes 28,000 and 140,000. Current plan models include GLM-5.3, GLM-5-Turbo, and GLM-4.7. Requests for GLM-5.2 or GLM-5.1 route to GLM-5.3. On GLM-5.3 with a 90.9% cache hit, Z.AI estimates weekly token allowance at 43 to 87 million on Lite, 263 to 526 million on Pro, and 613 to 1,226 million on Max, with low end all-peak and high end all off-peak.[4]
Credit burn depends on input, cached input, output, and model multipliers. Official peak multipliers are GLM-5.3 at 6.9/1.7/24, GLM-5-Turbo at 5.7/1.5/21, and GLM-4.7 at 4.6/1.2/16. The published formula is:
1credits = (input_tokens 脳 input_mult
2 + cached_input_tokens 脳 cached_mult
3 + output_tokens 脳 output_mult) / 10,000Plug in the sample CI-summary turn on GLM-5.3 at peak: (20,000 脳 6.9 + 10,000 脳 1.7 + 4,000 脳 24) / 10,000 = 25.1 credits. Lite's 2,000-credit 5-hour window holds about 80 of those turns. Off-peak model use costs 50% of standard credits, so the same turn is 12.55 credits. Z.AI defines peak as Monday through Friday, 14:00 to 18:00 Singapore time.[4][16]
OpenClaw uses an external Z.AI plugin and the Coding Plan default zai/glm-5.3. A zai/glm-5.2 request still works; Z.AI routes it to GLM-5.3.[17] Z.AI says OpenClaw traffic receives secondary, best-effort scheduling while coding-agent tasks take priority under load.[5]
馃幆 Production tip: If you pick Z.AI, run fat OpenClaw jobs outside Monday-Friday 14:00-18:00 Singapore time. Off-peak model use costs half the standard credits, and OpenClaw already sits behind coding-agent traffic when the cluster is busy.
Best fit: price-sensitive users who want GLM and can schedule around a weighted quota.
Main risk: cheapest sticker price doesn't guarantee lowest cost per completed task or low latency. Validate your own workload before upgrading a tier.
Set up without crossing billing paths
The setup decision is credential-first. Each branch below can lead to a similar model name but a different invoice or quota window.
For a ChatGPT/Codex subscription, use OpenClaw's own OAuth store. It no longer imports an existing ~/.codex login:
1openclaw onboard --auth-choice openai
2openclaw config set agents.defaults.model.primary openai/gpt-5.6-sol
3openclaw models list --provider openaiFor MiniMax Coding Plan OAuth:
1openclaw onboard --auth-choice minimax-global-oauth
2openclaw models list --provider minimax-portal
3openclaw models set minimax-portal/MiniMax-M3For Qwen Cloud Coding Plan, install the official provider plugin first:
1openclaw plugins install @openclaw/qwen-provider
2openclaw gateway restart
3openclaw onboard --auth-choice qwen-api-key
4openclaw models set qwen/qwen3.7-plusFor the global Z.AI Coding Plan:
1openclaw plugins install @openclaw/zai-provider
2openclaw gateway restart
3openclaw onboard --auth-choice zai-coding-global
4openclaw models set zai/glm-5.3These commands follow current OpenClaw provider docs.[3][13][14][5] After setup, inspect resolved model, fallbacks, and auth status without sending a paid probe:
1openclaw models status --json
2openclaw config get agents.defaults.model --json
3openclaw models auth listopenclaw models list proves catalog visibility, not successful execution. OpenClaw's --probe option makes live provider requests and can consume quota, so reserve it for deliberate troubleshooting.
Capacity and billing fail differently
OpenClaw failover has two stages: auth-profile rotation inside the current provider, then agents.defaults.model.fallbacks. Pure overload failures can retry the chain up to 10 times before tool execution or output begins. An explicitly selected session model stays strict instead of silently walking the default fallback list.[12]
That behavior makes billing guardrails part of routing policy. Don't configure automatic metered backup for a hard-stop subscription unless uninterrupted service is worth the variable bill. For personal workflows, a visible failure is often safer than silent paid overflow. Model Gateways, Routing, and Fallbacks shows how to make those transitions observable. Agent Failure & Recovery covers bounded retries so a 429 doesn't become an unbounded loop.
Match the plan to the job
Start with one lane. Add another only after logs show a specific gap: unavailable model, exhausted quota, unacceptable queue delay, or a task class that routinely fails.
Useful starting policy:
- Send summaries, search, small edits, and scheduled housekeeping to the primary plan.
- Bound retries and context growth before changing provider.
- Send stubborn debugging or high-impact decisions to a metered lane on purpose. Don't wait for a quota failure to "discover" that a production incident needed a stronger model.
- Record selected model, credential profile, fallback reason, latency, and cost or quota signal.
This is a workload hypothesis, not a universal quality ranking. Test representative tasks before buying a higher tier. LLM Cost Engineering covers the token and retry math behind that evaluation.
Secure the gateway before optimizing price
OpenClaw's Gateway is designed for one trusted operator, not mutually hostile users. Main-session tools run on the host unless sandboxing is enabled, and sandboxing is opt-in. Keep direct messages in pairing mode, restrict channel and group allowlists, enable sandboxing where practical, and require approval for dangerous execution. Run openclaw security audit after connecting channels or adding plugins.[18]
Treat prompts, tool results, files, and conversation context as data sent to the selected hosted provider. A self-hosted Gateway doesn't make hosted inference local. Provider privacy terms also differ by product. OpenAI states that API data and data from its business products aren't used for training by default, but that statement shouldn't be generalized to personal ChatGPT plans or other providers.[19] Code Generation & Sandboxing is the deeper lesson on keeping generated code and tool execution inside a real boundary.
鈿狅笍 Common mistake: Storing several provider keys, enabling automatic auth rotation, and assuming fallback is free. Limit secret access, name every paid backup, and review the resolved auth order before leaving OpenClaw unattended.
A starting plan
Use the plan you already have before buying another:
- Existing ChatGPT/Codex subscriber: start with OpenAI OAuth and watch subscription quota. Buy extra ChatGPT credits only if you want more of that same meter, not as a substitute for an API key.
- Lowest entry price: test Z.AI Lite on real tasks before committing to GLM-weighted credits and best-effort scheduling.
- MiniMax-first workflow: choose MiniMax Plus for rolling quota and a straightforward M3 route.
- Multi-model exploration: choose Qwen Cloud only if its current allowlist and hard request quotas fit your workload.
- Direct OpenAI API: add it only as an intentional metered lane with a budget and escalation rule.
There's no honest universal winner without task traces. One primary credential, bounded retries, and visible fallback reasons matter more than a longer provider list. If your OpenClaw setup is becoming a repeatable engineering workflow, AI Coding Workflow with Agents shows how to structure planning, execution, and verification around it.