Use Your Own API Key in a Browser Extension: Real Cost vs ChatGPT Plus, and How to Do It Safely

Bring your own key (BYOK) means you paste an API key from a model provider, such as OpenAI, Anthropic or Google, into a browser extension, and the extension sends your requests through your own provider account. You pay the provider per token instead of paying the extension for a subscription. Whether that is cheaper depends almost entirely on how much you use it and which model you pick.

This guide gives you a worked cost model with the arithmetic shown, the setup steps, a safety checklist for any extension that asks for a key, and an honest list of cases where a subscription wins. All prices below are per-token list prices copied from the vendors' pages as of November 2026. Vendor prices change, and at least one vendor page I read shows scheduled increases, so treat every number here as a snapshot and check the linked page before you decide.

Three things people mix up

• A chat subscription (ChatGPT Plus, for example) is a flat monthly fee for the vendor's own app.
• An API account is a separate developer account billed by usage. A ChatGPT subscription does not give you API credit, and API credit does not give you a ChatGPT subscription.
• An API key is the secret string that lets software spend that API account's balance. Anyone holding it can spend your money, which is why the safety section matters.

Copilotly supports this in its bring-your-own-key feature: you connect a key for OpenAI, Claude, Gemini, xAI, DeepSeek, Mistral, Groq or OpenRouter, and requests for that provider's models run on your account. Copilotly does not meter them. The rest of this post works out when that is a good deal.

Sometimes. For light to moderate chat on a small or mid-priced model, the API is usually much cheaper than a $20 subscription. For heavy use of a top-priced model, it can cost more. That is the honest answer, and it is the one that vendor-neutral guides tend to skip.

Here is why it flips. A subscription has a fixed price and a usage policy. OpenAI's ChatGPT pricing page describes Plus as "Expanded messages and uploads" with a "Limits apply" link, and its comparison table marks everyday text chats as unlimited for Plus, with a footnote that usage must be reasonable and follow OpenAI's policies (as of November 2026). OpenAI's help center lists Plus at $20 a month, according to the search result for its What is ChatGPT Plus? article; the pricing page itself did not show a dollar figure when I loaded it. So the subscription's cost is predictable and its limits are soft.

The API is the reverse: the limits are hard numbers, and the cost is whatever you use. Three variables drive it:

1. The model. OpenAI's pricing page lists models from $0.05 per million input tokens to $30 or more, a spread of several hundred times.
2. Input size. Every message re-sends the conversation so far, so long chats and attached documents multiply input tokens.
3. Output length. Output tokens cost several times more than input tokens on every price list below.

Here is the model, with every assumption stated so you can change it. A token is a chunk of text, roughly three-quarters of an English word, but exact counts depend on the tokenizer, so these are round planning numbers, not measurements.

The assumptions

• You send 50 chat messages a day, for 30 days, which is 1,500 messages a month.
• Each message averages 1,500 input tokens (your question plus the earlier context re-sent with it) and 500 output tokens (the answer, roughly 375 words).
• That is 2.25 million input tokens and 0.75 million output tokens a month.
• I ignore caching discounts, batch discounts and tool charges, all of which can lower the bill, and I ignore taxes.

The formula

Monthly cost = (input tokens in millions x input price) + (output tokens in millions x output price). Prices are per million tokens.

The results

OpenAI prices are from its API pricing page (standard tier, short context). Gemini prices are from Google's Gemini API pricing page. Both read as of November 2026.

Totals are rounded to the cent. Compared with a $20 monthly subscription, the same 50-a-day habit costs about $2 to $5 on a small model, about $17 on a mid-priced one, and about $34 on a top-priced one. That last row is the case where the API is more expensive.

Break-even: how many messages until the API costs $20?

Divide $20 by the cost of one message. One message is the monthly total divided by 1,500.

• gpt-5-mini: $2.06 / 1,500 = about $0.0014 per message, so $20 buys roughly 14,500 messages a month, about 480 a day.
• gpt-5.4: $16.88 / 1,500 = about $0.01125 per message, so roughly 1,780 messages a month, about 59 a day.
• gpt-5.5: $33.75 / 1,500 = $0.0225 per message, so roughly 890 messages a month, about 30 a day.

Why long chats cost more than you expect

A chat model has no memory between requests. The app re-sends the whole conversation each time. Take a 20-message conversation where each exchange adds about 700 tokens (200 from you, 500 from the model). Message 1 sends 200 input tokens, message 2 sends 900, message 3 sends 1,600, and so on. The sum over 20 messages is 700 x 190 + 200 x 20 = 137,000 input tokens, plus 10,000 output tokens. On gpt-5.4 that is 0.137 x $2.50 + 0.01 x $15.00 = about $0.49 for one conversation. Start a fresh chat for a new topic and you cut this cost directly.

A heavy-document case

Now suppose you paste a long report, about 30,000 tokens, ask for a 1,000-token answer, and do it 10 times a day for 30 days. That is 9 million input tokens and 0.3 million output tokens. On gpt-5-mini: 9 x $0.25 + 0.3 x $2.00 = $2.85. On gpt-5.4: 9 x $2.50 + 0.3 x $15.00 = $27.00. On gpt-5.5: 9 x $5.00 + 0.3 x $30.00 = $54.00. Model choice matters more than anything else on the page.

Two things in the vendor pages I read are worth knowing before you commit to a budget.

First, Google's Gemini pricing page lists some models with dated price changes. As of November 2026 it shows, for example, Gemini 3.8 Flash at $0.75 input and $3.75 output through December 31, 2026, rising to $1.50 and $7.50 from January 1, 2027, and Gemini 4 models with a similar step on December 1, 2026. If you plan a year, do not multiply today's price by twelve.

Second, the same page lists models marked "Free of charge" on a free tier (for example Gemini 3.5 Flash-Lite) and others marked "Not available" on it. Free tiers have their own rate limits, which the Gemini rate-limits page says are viewable in Google AI Studio rather than listed on the page.

Claude prices: not verified. I could not load Anthropic's official pricing page from my environment (the connection failed a certificate check), and third-party price sites I found disagreed with each other on which Claude models are current and what they cost. So this post gives no Claude per-token figures. Open Anthropic's pricing page directly and plug its input and output prices into the formula above. The method is identical. See the Claude model page for how the family behaves in practice.

The steps differ slightly by vendor, but the shape is the same: make an account, add credit or billing, create a key, then cap spending. Do the cap before you paste the key anywhere.

OpenAI

1. Sign in to the developer console and open the API keys page.
2. Add a balance first. Copilotly's feature page notes that the OpenAI API will not answer until the account has a balance.
3. Create a new secret key and name it for the one tool that will use it, for example "Copilotly".
4. Copy it immediately. Some providers show a key only once.

OpenAI's rate-limits guide explains the two spending controls, as of November 2026. A spend alert sends a notification while API traffic continues. A hard spend limit makes affected requests return a 429 error, which enforces a cap. Both can be set per organization or per project. Use the hard limit. A runaway loop or a leaked key then costs you at most the cap.

Google Gemini

Create a key at Google AI Studio (see also the Gemini API key guide). Some models are free of charge with rate limits. Paid tiers are unlocked by linking a billing account, and the rate-limits page lists billing caps per tier ($250 for Tier 1, $2,000 for Tier 2, as of November 2026).

Anthropic (Claude)

Keys are created in the Anthropic console at platform.claude.com/settings/keys and start with sk-ant-. Add prepaid credit and keep the balance small.

One cap that always works: prepaid credit

If a provider lets you prepay, load only what you are willing to lose, for example $10. Without auto-reload, the balance is your hard cap whatever the dashboard settings say.

Name it, scope it, date it

Create one key per tool, so you can revoke the extension's key without breaking anything else. Write the creation date in the key name or a note, and plan to rotate it every few months.

This is the verified flow from the bring-your-own-key feature page. It works on the Free plan.

1. Install the extension from the download page (Chrome, Edge, Firefox, Brave or Opera).
2. Open Settings and choose API Keys. The page is headed "Use any model, without limits".
3. Pick a provider card and follow its short guide to your provider's console.
4. Create the key, add credit or billing if the provider needs it, and copy the key.
5. Paste it and choose Save key. Copilotly checks the key with the provider using a read-only call. When it works, you see "Connected - your next reply uses this key".
6. Pick a model from that provider in the model picker and chat as normal.

What changes once a key is connected

• Requests for that provider's models run on your account and are never metered by Copilotly, on any plan.
• No Copilotly credits are used. You pay the provider's list price, and Copilotly adds no markup.
• It covers more than the main chat. Copilots and the reply assistants in Gmail, LinkedIn, X and WhatsApp use the key too, and so does rewrite anywhere on selected text.

What a key does not cover

Image generation and whole-page translation are Copilotly services with their own daily allowances, so a key does not lift those. A card in settings shows whether the key works and how many requests it has handled. If a provider rejects the key, usually for billing or revocation, Copilotly pauses it and answers with its own models, with a notice. Fix the problem at the provider and re-enable it.

Watch the expensive features

Deep research and long file chats can use a lot of tokens, and the provider bills for all of it. Set the spending cap first, and keep a frontier model for the jobs that need it. The deep research and chat with PDF pages explain what those features do.

Copilotly accepts keys for eight providers. Which to start with depends on how many model families you actually use.

OpenRouter, in one paragraph

OpenRouter is a router: one key and one bill that reaches many providers' models. The trade-off is a second company between you and the model. If you connect both a direct key and OpenRouter, Copilotly uses the direct key first for that provider's models and OpenRouter covers the rest. OpenRouter's quickstart covers creating a key and adding credit.

Which model for which job

The cheap models are fine for rewriting, summaries and routine Q&A. Save the pricey one for contract-length reasoning or code. The ChatGPT vs Claude vs Gemini comparison covers strengths, and the model pages for ChatGPT, Gemini and DeepSeek cover the rest.

It is as safe as the weakest place the key gets stored. An API key is a password with a price attached. Before you paste one into any extension, ask these questions. They apply to Copilotly and to everything else.

1. Where is the key stored?

There are two designs. A local-only extension keeps the key in your browser and calls the provider directly. A server-side design stores the key on the vendor's server and calls the provider from there. Neither is wrong, but they have different risks: local keys are exposed to malware or other extensions on your machine, while server-side keys are exposed to a breach at the vendor.

Copilotly uses the server-side design, and I want to be plain about it. Per the feature page, your key is sent to Copilotly's server and stored there, not only in your browser. It is checked with the provider using a read-only call, encrypted with AES-256-GCM, never sent back to your browser, and only its last four characters are shown. It is used only to answer your own requests and is not shared or used for training. Removing a key deletes it at once. If you are not comfortable with a key on a third-party server, use a key with a tiny spending limit, or choose a local-only extension.

2. What permissions does the extension request?

Open the extension's listing and read the permission list. Reading every site you visit is normal for a sidebar that summarizes pages, but you should know it is happening.

3. Who sees your prompts?

With your own key, the provider sees your prompts under its API terms, and the extension vendor may see them in transit if it proxies requests. Read both privacy policies. Do not send anything you would not paste into an email.

4. Can you cap the damage?

Set a hard spend limit or use prepaid credit, as described above. This is the single most effective control.

5. Can you revoke it in seconds?

Know where the revoke button is before you need it. Removing a key from an extension does not stop a copied key from working. Only the provider can do that.

6. Is the key scoped to this tool alone?

Never reuse your production application key. Create a new one for the extension, named for it.

If you think it leaked

1. Revoke the key in the provider's console.
2. Remove it from the extension.
3. Check the provider's usage dashboard for requests you do not recognise.
4. Create a new key with a lower cap.

Never paste a key into a chat message. If you did, revoke it.

A key is not always the answer. A subscription is usually better when:

• You use the top-priced model for a lot of messages a day. The break-even math above puts that line around 30 messages a day on gpt-5.5 at those assumptions.
• You value the vendor's own app. The ChatGPT subscription includes the app, uploads and other tools. An API key gives you the model, not the product around it.
• You hate variable bills. A flat fee never surprises you. Per-token billing can, which is why caps exist.
• You use it lightly. If you send ten messages a week, neither costs much. Copilotly's free plan may already cover you.

Copilotly's own plans, for comparison

Per Copilotly's published plans, Free has daily limits, and Pro is $6.99 a week, $19.99 a month or $149.99 a year with a 7-day trial for first-time subscribers (see pricing). Pro's annual price works out to about $12.50 a month. Many people use both: Pro for everyday chat and their own key for the occasional frontier-model task, with a low cap on the key.

"Not metered by Copilotly" does not mean unlimited. Your provider still applies its own limits, covered next.

A simple decision rule

Run the key for one week with a low cap and look at the provider's usage page. Multiply by four. If that number is clearly under what you would pay for a subscription, keep the key. If it is close, take the subscription and stop worrying. The sibling post on what ChatGPT Free and Plus actually give you covers the other side of the comparison.

The two worlds limit you differently, and people confuse them.

App caps

Consumer apps cap usage in ways that are often vague. OpenAI's pricing page says Plus has expanded messages with limits that apply, and everyday text chats are marked unlimited subject to reasonable use (as of November 2026). You usually cannot see an exact number until you hit it.

API limits

The API gives you numbers. OpenAI's rate-limits guide describes two separate controls (as of November 2026):

• Rate limits cap how fast you send traffic, in requests per minute, requests per day, tokens per minute, tokens per day and images per minute. They vary by model and apply at organization and project level. Your exact limits are under Settings, Organization, Limits.
• Monthly usage limits are an approved spend ceiling tied to your tier. The page lists Free at $100, Build at $500 (after $5 in credit purchases), Launch at $5,000 (after $100) and Grow at $200,000 (after $500). Tiers upgrade automatically as credit purchases grow.

Google's model is similar. Its rate-limits page, read in November 2026, measures requests per minute, input tokens per minute and requests per day, applies limits per project and not per key, resets daily quotas at midnight Pacific time, and returns a 429 error when you exceed one. Paid tiers rise with cumulative spend, and spend-based checks add a rolling-window ceiling.

What this means for a browser extension

A new account on a low tier can hit a rate limit during a burst, for example a long file chat followed by several rapid questions. The fix is to wait, use a smaller model, or buy a bit more credit to move up a tier. It is not a Copilotly limit. The same 429 error would appear in any tool using your key.

If you want to compare how each vendor's consumer plan limits work, the sibling guides on Gemini usage limits and Claude usage limits cover the app side.

1. Pick a model first. Price differences between models are bigger than the difference between a key and a subscription.
2. Estimate your tokens with the formula, then check one real week on the provider's usage page.
3. Set a hard cap or prepaid balance before pasting a key anywhere.
4. Create a dedicated, named key per tool and know how to revoke it.
5. Re-check prices. They changed during the weeks I researched this and some are scheduled to change again.

If that sounds right for you, connect a key in Copilotly and compare. If it sounds like too much to manage, a subscription or the free plan is a perfectly good answer.