Bring your own API key: costs, limits and a reusable worksheet

Understand provider billing, retries and optional cloud costs. Use a downloadable worksheet to measure the cost of accepted work; BYOK is not always cheaper than a subscription.

The More AI team · Updated September 14, 2026

Subscriptions vs. API pricing

BYOK means using a provider account and API key you control. The provider bills metered usage; an agent may make many requests for one task. A subscription can cost less for heavy use, while metered access can suit lighter or variable work. Compare the same completed tasks and include failed attempts, tool charges and review time.

The desktop app is free. More AI does not add a markup to BYOK model usage. Optional cloud execution and cross-device features are a separate subscription layer, temporarily free during the beta.

Which providers work with More AI

More AI ships with first-class support for Anthropic, OpenAI, Google (Gemini), Mistral, Groq, DeepSeek, Qwen and OpenRouter, plus any custom endpoint that speaks the OpenAI-compatible API — which includes local model servers like Ollama or LM Studio.

You can add keys for several providers at once and switch models per conversation. OpenRouter is a popular single-key option: one key, one bill, and models from every major lab in one list.

Adding and managing keys

Add a key in Settings → Providers. Desktop storage uses the operating-system keychain when available. The current Linux implementation can fall back to reversible obfuscation if no keyring is available; that fallback is not encryption. A supported cloud run may store the selected credential encrypted on More AI infrastructure and make it available to that execution environment. Do not assume that local and cloud credentials share the same boundary.

Keeping costs under control

More AI shows per-model prices (where the provider reports them) right in the model picker, and an optional budget filter hides models above a price ceiling you set. The context-window meter keeps long conversations honest — you can see exactly when a chat is getting expensive to continue.

A practical pattern: use a fast, cheap model (a “mini” or “flash” tier) for everyday questions and switch to a frontier model only for hard tasks. With BYOK the switch is two clicks, not a plan upgrade.

The searchable model picker showing models from several providers

Local inference: Ollama and LM Studio

Point a custom provider at a local OpenAI-compatible server and model inference can run without an internet connection. With sync, online tools, master telemetry and AI observability (including replay) off, prompts and inference stay on the machine — useful for sensitive documents or isolated environments.

Download a worksheet before comparing plans

The workbook below records one row per attempt, including failed attempts. Its three filled rows and token prices are fictional arithmetic examples, not current provider prices or measured More AI results. Remove those examples before recording your own work. Blue cells are inputs; formulas are kept separately.

Record the app version, exact model ID, date, provider usage reference and current pricing URL. Use one currency. Non-cached input and cached input are separate buckets; do not count the same tokens twice. Include reasoning in the bucket your provider bills, rather than adding it a second time.

Agent cost workbook (.xlsx) — Editable inputs, formulas, retry costs and a completeness check.

Measure the cost of an accepted task

A task is accepted only after the output passes your checks. For a code change, define the test and review the diff; for a document, verify its figures and open the file. Mark exactly one final accepted attempt per distinct task. Failed attempts remain in total spend.

For a simple token-only bill: cost = (non-cached input × its rate + cached input × its rate + billed output × its rate) / 1,000,000. Add other billable items separately. The worksheet divides all recorded attempt costs by accepted tasks, and leaves that figure blank when none has been accepted.

Choose one bounded task and define the acceptance checks before starting.
Record each attempt, even a stopped or failed one. Reconcile the row with the provider’s usage record.
Fill missing fields; a blank price or token count is unknown, not zero.
Compare accepted output, elapsed time and manual corrections, then add separate cloud, tool or hardware costs where relevant.

Use the provider bill as the final check

A model-picker price filter is not a spending cap. Configure any budget or alert available at the provider and verify how it is enforced. A long context, repeated tool loop, cache-write charge or search tool can change the bill. Stop and investigate a run whose cost or progress is unclear.

This page provides a measurement method and a verified spreadsheet, not a benchmark of DeepSeek, Qwen, GLM or More AI. The worksheet cannot establish which model is best until you enter comparable real runs.

Frequently asked questions

Is BYOK cheaper than a subscription?

Not necessarily. Compare provider-billed input, output, caching, retries and tool charges for accepted tasks, plus any separate infrastructure subscription. An inexpensive token rate does not guarantee an inexpensive successful result.

Can I use more than one provider at once?

Yes. Add keys for several providers and switch models per conversation. OpenRouter is a popular single-key option that exposes models from every major lab under one bill.

Can I run More AI with no API cost at all?

Yes — point a custom provider at a local OpenAI-compatible server such as Ollama or LM Studio. Model inference then has no per-token charge and can run offline; turn off sync, online tools, master telemetry and AI observability (including replay) when prompts and inference must remain on your machine.

Where do I get an API key?

From your provider’s dashboard — for example console.anthropic.com, platform.openai.com or openrouter.ai. Create a key, copy it, and paste it into Settings → Providers; More AI fetches that provider’s live model list right away.

Are my API keys safe?

All guides