Free Tokens from OpenAI

OpenAI gives you free API tokens every day if you share your API traffic with them. It's not a lot, but you get up to $12.50/day of GPT-6 Astra, plus 10x as many tokens on GPT-5.6 Terra. Use Astra to plan and Terra to build.

The deal

Group	Tiers 1–2	Tiers 3–5	Models
Frontier	250K/day	1M/day	gpt-6-astra, gpt-6-sol, gpt-6-luna, gpt-5.6-sol, gpt-5.5, o3…
Small	2.5M/day	10M/day	gpt-5.6-terra, gpt-5.6-luna, gpt-5.4-mini, o4-mini…

Astra costs $10/M input and $50/M output, so 250K tokens is worth up to $12.50/day (up to $50/day on Tiers 3–5). Full model list.

Turn it on

1. As an org owner, go to Data Controls → Sharing.
2. Enable input/output sharing, ideally on one dedicated project only.
3. Keep a positive account balance. You won't be charged for the free tokens, but the program won't work at $0.

Rules

• Resets at 00:00 UTC, with no rollover. Each group shares one quota.
• The request that crosses the cap is billed in full, so leave some headroom.
• Fine-tuning, evals, and tool use aren't covered.
• You're paying with your data. Don't send secrets, customer data, or proprietary code.
• To check it's working, open the Usage dashboard and group by service tier. Free tokens show up as "data sharing incentive tier."

The hack: Astra plans, Terra builds

1. Astra writes the plan (~20–60K tokens). Ask for a plan, not code, since output tokens are the expensive ones:

You're a staff engineer. A less capable agent will execute your plan
literally and can't ask questions. Goal: <X>. Context: <stack>.
Write numbered small tasks with exact files, function signatures,
acceptance criteria, and pitfalls. No full implementations.

Save it as PLAN.md.

2. Terra does the work. Point your agent at gpt-5.6-terra:

Execute the next unchecked task in PLAN.md exactly. Check it off when
acceptance criteria pass, then stop.

3. Escalate only when Terra gets stuck. Send Astra just the error and the relevant file.

It costs $0, and you get a frontier model doing the thinking every day. Go turn it on.

If you already have a project, have a bunch of files, and you want to give it to Astra in a single prompt to save tokens (compared to using an agent harness like ChatGPT/Codex/etc) I made a free tool called AI Code Prep GUI which will let you select all the files you want included in a GUI (Windows, Mac, Linux, anywhere that can run a python Qt app). Can run without installing using uv:

uvx aicodeprep-gui

I also made a Rust version that is much faster for larger repos/more files. This is the main way I was coding with AI for a long time, until the agent harnesses finally became very good - the app still comes in handy for situations like free tokens or when I want to send a bunch of files to several different models inside their native web chat interfaces.