How do I run Kilo Code on a fixed monthly budget instead of per-token credits?
Add a flat-rate OpenAI-compatible API such as Standard Compute to Kilo Code as a custom provider: one monthly price from $19 buys a fixed compute budget across Claude, GPT-5.6, Kimi K3, GLM, Qwen and DeepSeek, and at 100% new requests stop with HTTP 402 instead of adding a charge. Kilo's own credits and Kilo Pass are both paid per token.
Provider ID: standardcompute
Display name: Standard Compute
Provider API: OpenAI Compatible
Base URL: https://api.stdcmpt.com/v1
API key: YOUR_API_KEY
Models: StandardCompute
Token limits for the extension live in the same kilo.json. Left at 0, Kilo compacts far too early. Your Kilo models stay in the picker, so you can switch back for any task.
Every way to pay for Kilo Code, and what stops you
| Option | What you pay | What happens at the limit |
|---|---|---|
| Kilo Gateway credits | The provider's list price per token, no markup, plus a 5% processing fee on each credit purchase | Paid models return HTTP 402 at a zero balance, unless auto top-up buys more |
| Kilo Pass | $19, $49 or $199 a month, paid out as credits. Bonus credits (10% in month 1, 50% in month 2, then up to 40%) come after the paid credits and expire each cycle | Top up at list price, or wait for the renewal |
| Your own provider keys | $0 to Kilo; the provider's per-token price | That provider's rate limits and any spend limit you set there |
| ChatGPT Plus or Pro sign-in | Your ChatGPT plan; supported OpenAI models only | Requests stop with an error at your ChatGPT usage limit, with no paid fallback |
| Standard Compute | $19, $39, $89, $249 a month for a $20, $41, $95, $269 compute budget | New requests get HTTP 402 until the plan renews, with nothing billed on top |
Kilo terms from kilo.ai/pricing, Kilo Pass, usage and billing and ChatGPT Plus/Pro, checked October 14, 2026. Kilo bills Cloud Agents ($1.20 an hour) and Teams seats ($15 per user a month) separately on every route.
What Kilo itself charges for a model
- Kilo Gateway passes through the provider's price with no markup. On October 14, 2026 it lists GPT-5.6 Sol at $4 input and $20 output per 1M tokens ($0.40 cached), the same as OpenAI's API price. The only extra is the 5% fee when you buy credits: a $20 purchase costs $21.
- Requests on your own provider key are billed $0 by Kilo, so a BYOK OpenAI key costs exactly OpenAI's price.
- Kilo's Auto models route per request; Auto Efficient picks the cheapest model it rates accurate enough for each step. Every Auto request is still paid from credits.
- Kilo does not rate-limit paid requests at the gateway. The limits you hit come from the upstream provider or from an organization's per-member daily spend limit; free models are capped at 200 requests an hour per IP.
A real four-month Kilo bill against a flat plan
These are four months of Kilo spend that a developer posted when asking whether to switch. With every request pinned at list price, only the Pro plan covers the two busy months, and it costs more than the credits did. The Standard plan costs less than the credits but runs out in the busy months unless routing moves routine steps to an efficient model: a median coding-agent day costs $3.67 on Claude Sonnet 5.5 and $0.23 on GPT-6 Luna at list prices.
| Month | Kilo usage | Kilo credits, +5% fee | Standard plan | Pro plan |
|---|---|---|---|---|
| Month 1 | $61 | $64.05 | $89, covered | $249, covered |
| Month 2 | $138 | $144.90 | $89, runs out | $249, covered |
| Month 3 | $89 | $93.45 | $89, covered | $249, covered |
| Month 4 | $204 | $214.20 | $89, runs out | $249, covered |
| Four months | $492 | $516.60 | $356 | $996 |
Budgets: Standard $95, Pro $269. "Covered" means the month's usage at list prices fits the budget with every request pinned. Usage on your own provider keys carries no 5% fee. Day costs from the Agent Burn Index.
Set it up in Kilo Code
- Pick a plan from $19 a month and create a key.
- VS Code: open Settings (gear) > Providers, scroll down and click Custom provider. Enter a Provider ID such as standardcompute, set Provider API to OpenAI Compatible, Base URL to
https://api.stdcmpt.com/v1and paste your key. - Kilo fetches the model list from
https://api.stdcmpt.com/v1/models. AddStandardComputefor the router, or a pinned model such asmoonshotai/kimi-k3,qwen/qwen3.8-max-0902,z-ai/glm-5.3oropenai/gpt-5.6-sol, then click Submit. - CLI: put the provider block (tabs above) in
~/.config/kilo/kilo.json, exportSTANDARDCOMPUTE_API_KEY, and setmodeltostandardcompute/StandardCompute. - Give each model a context limit (1000000 for the router), then run one small edit and check it in the dashboard's request log before a long task.
Why queued Kilo tasks hit rate limits
- Several Kilo tasks on one provider key multiply the requests per minute. A direct key on a low usage tier hits its own request and token limits first, and Kilo then retries.
- On Standard Compute each key accepts 300 (Starter), 600 (Economy), 1,000 (Standard) or 2,000 (Pro) requests a minute, and a transient upstream 429 is retried with another provider for the same model before Kilo sees an error.
- Above the key's ceiling you still get HTTP 429, so keep Kilo's retry backoff on. Over a month, the budget is the limit you will meet.
The catch
A flat price buys a fixed amount of compute: each budget is $1 to $20 more than its plan price, counted at list rates. Pin Claude Opus 5.5 to every request and a heavy month runs out of budget; new requests then stop until renewal. If you code in Kilo most days on one frontier model and keep a Kilo Pass streak, Pass bonus credits can buy more tokens per dollar than our plans.
Where Standard Compute fits
Standard Compute is an OpenAI- and Anthropic-compatible API on monthly plans. Kilo Code connects as a custom provider, and the same key works in Claude Code, Codex, OpenCode, Cline and Roo Code. Every plan has the same model catalog, the dashboard shows the budget live, and nothing is billed on top. It does not cover Kilo's Cloud Agents, Code Reviewer or Teams seats, which Kilo bills separately.
FAQ
Kilo burned through $25 of credit on what I thought was a small task. What's the fixed-fare equivalent of per-token pricing?
A flat-rate API plan used as Kilo's provider. With Standard Compute you pay one monthly price (from $19) for a fixed compute budget; a runaway task can use up budget, but it cannot add a charge, because at 100% new requests stop with HTTP 402 until renewal. Kilo Pass is fixed per month too, but it is paid out as metered credits, and a heavy task still drains them at the token rate.
Is it cheaper to run Kilo through its own credits or bring my own flat-rate API if I use it 4 to 5 hours a day?
It depends on the model mix more than the hours. Each Standard Compute budget is only $1 to $20 above its price at list rates, so on one frontier model all day, Kilo credits or a Kilo Pass with a full streak cost about the same or less. The flat plan wins when most steps run on efficient models (a median coding-agent day is $0.23 on GPT-6 Luna against $3.67 on Claude Sonnet 5.5) and when you want the month capped at a known price.
Which flat-price OpenAI-compatible providers do Kilo users run with BYOK for long agent sessions?
Standard Compute is the flat-price API that covers every major model family: add it as a custom provider with base URL https://api.stdcmpt.com/v1, plans from $19 a month, a router with a 1,000,000-token context limit, and no 5-hour or weekly windows. Kilo's ChatGPT Plus or Pro sign-in is also flat, but it covers only OpenAI models and stops at ChatGPT's usage limit. A model maker's coding plan covers only that maker's models.
Does Kilo mark up GPT-5.6 Sol over OpenAI direct?
No. On October 14, 2026 Kilo Gateway lists GPT-5.6 Sol at $4 input and $20 output per 1M tokens, the same as OpenAI's API price; the only extra is the 5% processing fee when you buy credits, and a BYOK OpenAI key in Kilo costs nothing on top. To stop paying per token for Sol, pin openai/gpt-5.6-sol on a Standard Compute key in Kilo: it draws on the monthly budget instead.
Kilo keeps hitting provider rate limits when I queue up multiple tasks. Is agent burst traffic just unsupported?
It is supported, but per-key limits decide it. Kilo's gateway does not rate-limit paid requests, so the 429s come from the upstream provider's limits on your key. Standard Compute allows 300 to 2,000 requests a minute per key by plan and retries transient upstream 429s with another provider for the same model, so queued Kilo tasks see far fewer retries. Keep Kilo's backoff on for bursts above your key's ceiling.
Can I get Kimi K3 in Kilo Code without buying credits from yet another provider?
Yes. On a Standard Compute key, pin moonshotai/kimi-k3 in Kilo's custom provider; it draws on the same monthly budget as Claude, GPT-5.6, GLM, Qwen and DeepSeek. Every plan has the same catalog, so there is no per-model add-on. Kimi K3 lists at $3 input and $15 output per 1M tokens, so a long session on it uses budget faster than an efficient model.
What's the shortest path to Qwen 3.8 as my Kilo default model?
Create a Standard Compute key, add the custom provider in Kilo (OpenAI Compatible, base URL https://api.stdcmpt.com/v1), and pick qwen/qwen3.8-max-0902 from the fetched list; Qwen 3.8 Flash and Qwen 3.8 27B are there too. In the CLI, set "model": "standardcompute/qwen/qwen3.8-max-0902" in kilo.json. The same key later serves any other model in the catalog.
GPT-5.6 Luna or DeepSeek V4 as my everyday Kilo model, and can a subscription pick for me so I stop redoing the math each release?
Select both on Standard Compute's Models page and set Kilo's model to StandardCompute: the router chooses between your selected models for each request, and the plan price stays the same whichever it picks. To compare them yourself, pin openai/gpt-5.6-luna and deepseek/deepseek-v4-flash in turn on the same task and check each run in the request log.
Between Kilo credits, an OpenAI key and an Anthropic key, I don't know where my money goes. What do people consolidate onto?
One API key with one monthly charge. A Standard Compute key serves GPT, Claude, Gemini and the open-weight models in Kilo, Claude Code and Codex alike, so the plan price (from $19) is the whole bill and the dashboard shows usage per request. A company workspace gives each developer's key its own plan, with monthly invoices in the company's name.
Is there a flat-rate option with Gemini 2.5 Pro's long context for Kilo?
Yes. google/gemini-2.5-pro is in the Standard Compute catalog with a 1,048,576-token context window, billed against the monthly budget instead of a Google Cloud invoice. Long prompts use budget in proportion to their size, so a session that resends a huge context each step costs more of the month than a short one.
OpenRouter keeps returning 402 because auto top-up failed on my prepaid card, and my Kilo agent dies silently mid-task. Are there more forgiving billing alternatives?
Pick billing where the stop is planned, not a failed card. Standard Compute is a monthly subscription: the plan's budget is there from the first day of the cycle, the dashboard shows how much is left, and nothing depends on a top-up charge succeeding mid-task. It still returns HTTP 402 at 100% of the budget, so size the plan to your month ($19 to $249) and upgrade early if the meter runs ahead; Pay as you go top-ups are there only if you turn them on.
GLM-5.3 Flash on a cheap provider is fine 90% of the time, but latency spikes and failed responses break my Kilo automation at night. Is flat-rate worth it?
It can be, for the failover more than the price. On Standard Compute, z-ai/glm-5.3-flash and z-ai/glm-5.3 are in every plan's catalog, and transient upstream errors are retried with fallback to another provider for the same model before Kilo sees them. For overnight runs, set Kilo's retries on anyway and check the request log in the morning; a flat plan also means a runaway night costs budget, not an extra charge.
I'm mid-project in Kilo and the per-token burn is forcing a switch. What's the least disruptive way?
Add the flat-rate provider next to your current one instead of replacing it. A custom provider in Kilo leaves your tasks, history and repository untouched, and both providers stay in the model picker, so you can move the next task over and switch back in one click if a model behaves differently.