How do I use GitHub Copilot with a flat-rate API through a custom endpoint?
In VS Code, choose Manage Language Models > Add Models > Custom Endpoint and add a flat-rate OpenAI-compatible API such as Standard Compute: Copilot chat and agents run on it for one monthly price from $19, and at the budget requests stop with no AI Credits spent. Inline completions stay on Copilot.
[
{
"name": "Standard Compute",
"vendor": "customendpoint",
"apiKey": "${input:standardComputeKey}",
"apiType": "chat-completions",
"models": [
{
"id": "StandardCompute",
"name": "Standard Compute",
"url": "https://api.stdcmpt.com/v1",
"toolCalling": true,
"vision": false,
"maxInputTokens": 200000,
"maxOutputTokens": 32000
}
]
}
]
VS Code adds /chat/completions to the URL. To pin one model, copy the model block and set its id, for example anthropic/claude-opus-5.5.
What changed on June 1, 2026
| Copilot plan | Price | AI Credits a month |
|---|---|---|
| Pro | $10 | 1,500 (1,000 base + 500 flex) |
| Pro+ | $39 | 7,000 (3,900 + 3,100) |
| Max | $100 | 20,000 (10,000 + 10,000) |
| Business | $19 per seat | 1,900 per user |
| Enterprise | $39 per seat | 3,900 per user |
1 credit = $0.01, metered by tokens since June 1, 2026; credits reset at 00:00 UTC on the 1st. Completions and next edit suggestions use none on paid plans. Sources: GitHub's announcement and plans, checked October 14, 2026.
What a custom endpoint covers in Copilot
| Part of Copilot | On a Custom Endpoint model |
|---|---|
| Chat and agents | Yes, for models with tool calling turned on |
| Agents window | Yes, with the experimental setting in the settings.json tab |
| Completions and search | No. Inline suggestions, semantic search and embeddings stay on Copilot and need a GitHub account |
| Cloud agent | No. It runs on GitHub and draws on AI Credits |
| Business and Enterprise | Only after an admin enables the Bring Your Own Language Model Key in VS Code policy |
| No Copilot plan | Works: BYOK models need neither a GitHub sign-in nor a plan |
From VS Code's language models docs (updated October 7, 2026). Custom Endpoint replaced the older OpenAI Compatible provider and speaks Chat Completions, Responses or the Anthropic Messages API.
Set it up in VS Code
- Pick a plan from $19 a month and create a key.
- Open the model picker in Copilot Chat and select Manage Language Models (gear), then Add Models > Custom Endpoint.
- Enter a group name (Standard Compute), your key and the API type Chat Completions. VS Code opens
chatLanguageModels.json. - Paste the models block from the tabs above:
StandardComputefor the router, or a pinned ID such asanthropic/claude-opus-5.5,openai/gpt-5.6-terraormoonshotai/kimi-k3. KeeptoolCallingtrue so agents list it. - Pick the model in the chat picker, run one small agent task and check it in your dashboard's request log.
Splitting a team between Copilot and a flat plan
- Copilot seats cover completions; agent sessions are what spend the credits. Keep the seats for completions and move only the heavy agent users.
- Each heavy user gets their own key and plan in a company workspace, for example Standard at $89 with a $95 budget, and picks it in Copilot Chat or in Cline in the same VS Code window.
- Light users stay on their Copilot seat. The company gets one invoice for the keys and keeps GitHub's bill for the seats.
The catch
A flat plan buys a fixed amount of compute: each budget is $1 to $20 more than its plan price at list rates. Pin Claude Opus 5.5 for every agent step and a heavy month runs out of budget, and new requests stop until renewal. If your agent sessions fit inside Copilot Pro's 1,500 credits, the $10 plan is the better deal, and it includes completions.
Where Standard Compute fits
Standard Compute is an OpenAI- and Anthropic-compatible API on monthly plans. In Copilot it runs as a Custom Endpoint model; the same key works in Cline, Kilo Code, Claude Code and Codex. It does not replace Copilot's completions or cloud agent, which stay on your GitHub plan.
FAQ
Copilot's agent mode ate my credits for the month in 6 days. When does the counter reset, and what do people bolt on when Copilot runs dry?
Included AI Credits reset at 00:00 UTC on the 1st of each month, whatever your billing date. Until then you can set a budget for extra usage, upgrade, or bolt on a flat-rate API: add Standard Compute as a Custom Endpoint model in Copilot Chat, or run Cline in the same VS Code window on the same key. Plans start at $19 a month and stop at the budget instead of billing more.
Copilot Pro+ still meters the frontier models. I want a flat price for frontier access. Where did Copilot power users go?
Since June 1, 2026 monthly Copilot plans meter every model by tokens in AI Credits (multipliers remain only on annual plans until renewal). Power users who want one price usually add a flat-rate API as a Custom Endpoint model or move agent work to Cline, Kilo Code or Claude Code. Standard Compute serves Claude Opus 5.5, GPT-5.6 Terra and Sol, Gemini and the open models on $19, $39, $89, $249 plans; each budget is close to list price, so the gain is a fixed ceiling and routing, not cheaper frontier tokens.
Copilot's agent mode picks the model for me. I want to pick Claude Opus 5 or GPT-5.6 Terra myself. Best combo of open tool plus subscription API?
You can keep Copilot and still pick: a Custom Endpoint model with the ID anthropic/claude-opus-5 or openai/gpt-5.6-terra sends every request to exactly that model. Outside Copilot, Cline or OpenCode on the same Standard Compute key does the same. Use the StandardCompute router only if you want it to choose among models you select.
Our org pays Copilot Business for 30 devs and the agent features everyone wants cost extra anyway. Has anyone switched to BYO-API tools like Cline or Roo plus a flat inference plan at team scale?
Yes, usually as a split rather than a full switch. Business seats ($19 each, 1,900 credits a user) keep completions; heavy agent users get a Standard Compute key each in a company workspace and run Cline, Kilo Code or a Custom Endpoint model in Copilot Chat. Roo Code shut down on May 15, 2026; its community fork ZooCode takes the same settings. An admin must enable the Bring Your Own Language Model Key in VS Code policy for the Copilot Chat route. Check how many of the 30 actually exhaust credits before buying 30 plans.
I hit Copilot's cap mid-sprint and the fallback model is much weaker. Are there flat-rate alternatives that don't degrade?
On Standard Compute a pinned model is never swapped: if you set anthropic/claude-opus-5.5, every request goes to Opus 5.5 until the budget is used, and then new requests stop with HTTP 402 rather than falling back to a weaker model. The router only chooses among models you select. Plans cost $19, $39, $89, $249 a month with no overage.
Team lead with 12 Copilot seats: 4 people love it, 8 barely touch it, and the 4 heavy users hit request caps. Is flat-rate API plus open agent tooling for the 4 sane?
Yes. Keep the 8 light users on their seats and give the 4 heavy users a flat plan each: on Standard Compute's Standard plan that is $356 a month for four keys with a $95 budget each, usable in Copilot Chat as a Custom Endpoint model or in Cline. Completions stay on the Copilot seats you keep.
Copilot overage charges landed on our Azure invoice where nobody looks: a surprise $340. Which per-seat flat inference vendors belong in my comparison?
Compare on what happens at the limit. Copilot's extra usage keeps billing until the budget you set; a flat plan should stop instead. Standard Compute gives each developer a key with its own monthly plan and budget, stops at 100% with no overage, and invoices the company monthly. Also list the model makers' team plans (Claude Team, ChatGPT Business) and check their usage limits, and set a GitHub budget for the seats you keep.
Copilot's autocomplete is included and the agent part is metered. Is anyone offering uncapped agent inference monthly?
Not honestly: every flat plan has a limit somewhere. Standard Compute's limit is one published number, a monthly compute budget ($20, $41, $95, $269 on $19, $39, $89, $249 plans), with no 5-hour or weekly windows and no overage; at 100% new requests stop until renewal. Keep Copilot for autocomplete and run agents on that budget.
Copilot won't let me use MiniMax M3 or most Chinese models. Talk me through switching to open tools plus a multi-model subscription.
You can add them to Copilot itself: a Custom Endpoint model with the ID minimax/minimax-m3, moonshotai/kimi-k3, z-ai/glm-5.3 or qwen/qwen3.8-max-0902 on a Standard Compute key appears in the Copilot Chat picker. Or run Cline or OpenCode with the same key. All of them draw on one monthly budget.
Does a custom endpoint work for Copilot's inline completions?
No. VS Code's docs say inline suggestions, semantic search and embeddings still need a GitHub account and stay on Copilot. Custom Endpoint models cover chat and agents. Copilot Free includes 2,000 completions a month.
Do I need a Copilot subscription to use a custom endpoint?
No. VS Code's docs say BYOK models work without signing in to GitHub and without a Copilot plan. You only need a GitHub account for completions and the other Copilot-hosted features.