A $25 pack at the standard rate, which is the list price plus 25%. Claude Sonnet 5.5 lists at $2 per million input tokens and $10 per million output tokens on October 1, 2026. With three input tokens for every output token, $25 buys about 5,000,000 tokens.
Buy AI in packs.Keep it or gift it as long as you want.
A Brass pack is prepaid intelligence you can gift. The recipient gets an OpenAI-compatible API key, pastes it into Cursor, Claude Code, a bot, or any tool that already speaks OpenAI, and Brass meters the pack until it hits zero. Credits never expire. A simple web chat is there too if they want it.
Paid credits never expire. For people 18 and over in the United States.
You have reached your usage limit. Your limit resets at 4:00 PM.
A typical message from a flat AI plan.
A pack stays in your stash until you spend it.
Most AI plans sell an allowance that refills every few hours or every week, and whatever you did not use is gone. A Brass pack is a dollar balance. Nothing resets it. The pack you buy this year is the same pack next year if you have not spent it.
Paid packs carry no expiry date and no fees. If Brass ever gives promotional credits, each one shows its expiry date. Monthly packs are planned for early 2027. If you cancel one, you keep every credit you paid for.
If Brass ever closes, we will email you at least 30 days ahead and pay out every sealed pack and unspent balance, as the Terms promise.
The same $25 buys a different number of tokens on each model.
Brass charges the model’s list price plus 25%, and both numbers print on every receipt. Pick a model to see what a $25 pack buys today.
5,000,000
tokens on Claude Sonnet 5.5 for $25
List prices published by each model maker on October 1, 2026. Assumes three input tokens for every output token, so the blended list price for Claude Sonnet 5.5 is $4.00 per million tokens and the Brass price is $5.00 per million. Rounded to the nearest 100,000 tokens.
Each model splits text into tokens its own way, so a token on one model is not a token on another. The market board compares every model in one unit, dollars per million words.
Packs come in $10, $25, $50, $100 and $250. Monthly packs of $20, $50 or $100 are planned for early 2027.
A pack is counted in dollars, so it buys whichever model does the work for less.
Epoch AI measured the price of a fixed level of AI performance and found it has fallen about 47% a quarter since 2023, which is about 13-fold a year. The newest top models are not the cheap ones. GPT-4 listed at $30 in and $60 out per million tokens in March 2023, and GPT-6 Astra and Claude Fable 5.1 list at $10 in and $50 out today. Each year, smaller and cheaper models reach the scores that the previous top models set, and a pack of dollars goes to whichever of the eight models Brass offers does the job for less.
Brass reads the model makers’ own pricing pages every hour. A price cut reaches every pack once Brass confirms it on the maker’s page, usually within hours. A price increase applies to new requests from the day it takes effect. Increases happen too, as in April 2026, when GPT-5.5 listed at twice the price of GPT-5.4. The dollars in your pack do not change either way.
Using GPQA Diamond scores from Epoch AI (a test of graduate-level science questions) and each developer’s published prices, we tracked the cheapest model that scores at least as well as the original GPT-4. In {THEN_MONTH} that model was {THEN_MODEL} itself, at ${THEN_PRICE} per million tokens with three input tokens for every output token. By {NOW_MONTH} it was {NOW_MODEL}, at ${NOW_PRICE} per million. At list price, $25 would have bought about {THEN_TOKENS} tokens of that work in {THEN_MONTH} and about {NOW_TOKENS} by {NOW_MONTH}. Newer models often write more tokens per answer, so the cost of one correct answer on this test fell about 70-fold over the same period. {NOW_MODEL} is not in Brass yet; the cheapest of the eight models Brass offers, DeepSeek V4.1 Flash, gives about 38,100,000 tokens for $25 at the Brass rate.
Using GPQA Diamond scores from Epoch AI (a test of graduate-level science questions) and each developer’s published prices, we tracked the cheapest model that scores at least as well as the original GPT-4. In March 2023 that model was GPT-4 itself, at $37.50 per million tokens with three input tokens for every output token. By July 2026 it was Qwen3.7 Flash, at $0.055 per million. At list price, $25 would have bought about 667,000 tokens of that work in March 2023 and about 455,000,000 by July 2026. Newer models often write more tokens per answer, so the cost of one correct answer on this test fell about 70-fold over the same period. Qwen3.7 Flash is not in Brass yet; the cheapest of the eight models Brass offers, DeepSeek V4.1 Flash, gives about 38,100,000 tokens for $25 at the Brass rate.
Token figures in this history are 25 divided by the list price per million tokens, rounded to three significant figures; the picker above adds Brass’s 25%.
Cheapest list price for GPT-4-level results on GPQA Diamond, in US dollars per million tokens, three input tokens per output token, log scale.
The second line is the launch list price of OpenAI’s newest flagship model at each date, on the same blend.
Price data is being compiled.
Cheapest list price for GPT-4-level results on GPQA Diamond, in US dollars per million tokens, three input tokens per output token, log scale.
Buy a pack for anyone 18 or over in the US, or send one you already own. It travels as a link, an email or a printed card with a QR code, with a note if you want one. The note we suggest is “Open when you hit your limit.” If nobody claims the pack within 90 days, it comes back to you, still sealed.
A sealed pack can be kept, sent or opened. A pack you received can be sent on once, from the first midnight Eastern 14 days after it was bought. Once it is opened, the credits stay in the owner’s stash. Unopened, unclaimed packs can be refunded to the original card within 30 days.
Giving it on a later day? Buy it as a gift and keep the link until then, since its 90 days start the day you buy it, or buy it for yourself and send it from your stash on the day. A pack you buy for yourself can be sent at once when your bank asks you to confirm the payment and you do, or you pay with Apple Pay; otherwise from the first midnight Eastern 14 days after you buy it. The buy page says which suits your account.
Gifts sent or claimed from Nov 23 to Dec 31 carry HOLIDAY 2026 on the token’s rim.
Still shopping for the holidays? A Brass pack goes by email or as a card you print at home, so it is there the day you give it.
Gift intelligence they can use where they already work.
Send a sealed pack to a parent, a partner, an adult child, a teammate or an employee. They claim it, copy an OpenAI-compatible API key, and paste that key into the agents and tools they already use. Brass brokers the models and meters the prepaid balance. When the pack hits zero, the key stops.
Monthly packs are planned for early 2027, so you can send someone a pack that arrives on the same day each month. You will see how much they spent and never what they asked. A household plan for up to five people, with a spending cap for each person, is planned after that.
Prefer a browser? Brass still has a simple web chat with a model picker. Every spend gets a receipt with the list price, our cut and the total.
An example receipt. Under every answer it reads as one line: Claude Sonnet 5.5 ·1,240 in, 310 out ·list $0.0056 ·Brass cut $0.0014 ·total $0.0070 ·$24.99 left. Figures use the October 1, 2026 list price.
Buy or gift your first pack.
Packs come in $10, $25, $50, $100 and $250. Gift prepaid intelligence someone can spend through an API key in their own tools, or keep a pack for yourself. The first 1,000 packs carry founding serial numbers, 000001 to 001000.