chatgpt pro now comes with up to $500 in api credits - here's the catch

from the episode: ChatGPT almost merged chat and work. i measured what actually eats your limit | LAB 0032

Lenny's AI Builders - LAB 0032

how to claim the credits, the one prompt i pointed at them, what 12 transcripts cost, and the catch.

ChatGPT almost merged chat and work, then OpenAI walked it back. the same day, OpenAI handed out up to $500 a month in API credits with one catch.

the catch: the credits don't cover Codex. so you don't point Codex at them. you point your own little agent at them.

why it matters

my usage page showed chat costs nothing, but Codex and work share one limit. one work automation that said "no update" five times ate 2.51% of my whole week.

credits are a sorry in advance. they also get you used to API prices. so measure what your work costs, then make the credits do the heavy lifting.

what you get

one prompt you can copy. it writes a small script that runs your files through the OpenAI API with the credits.
my real run. 12 transcripts, the token counts, the time and the cost.
the catch, dated. who gets credits, what they don't cover, and how fast my first grant expired.

the real example: my run

i pointed the script at 12 LAB transcripts. it asked for one headline and three takeaways per file.

tokens: 54,701 in, 2,469 out
time: 16.83 seconds
cost: about seven cents at OpenAI's list price (AI Lenny's estimate in the episode). that's API billing, not your weekly limit.
what came back: the headline was "test for taste not just benchmark scores". the takeaways: run a blind task-specific model test, put five simulated users through launch QA, and find forgotten follow-ups before customers do.

try it

claim the credits. go to chatgpt.com/settings/billing. just above "cancel plan" there's API subscription benefit. press activate and link your API org.
make one API key and put it in an environment variable called OPENAI_API_KEY. never in the script, never in chat.
paste the prompt below into Codex or Claude to get the script.
drop your files in ./transcripts and run it.
check your usage on platform.openai.com and your ChatGPT weekly bar.

copy this prompt

paste this into Codex or Claude:

Lenny’s AI Builders / LAB 0032

Write me a small Python script. It reads every transcript in ./transcripts, sends each one to the OpenAI Responses API with my API key from the OPENAI_API_KEY environment variable, asks for one punchy headline and three takeaways a builder can act on, runs four at a time, saves each answer to ./out, and prints the total input and output tokens. Never print the key.

this is the exact prompt i used on camera.

you'll know it worked when

there's one answer file per transcript in ./out, the script prints the total input and output tokens, and the usage shows up on platform.openai.com - not on your ChatGPT weekly bar.

a bad result: the key shows up in the script, the terminal or a chat. delete that key and make a new one.

limits and dates

who gets it: Pro and Business, up to $500 a month depending on your plan. as filmed 30 October 2026 - check your billing page, offers change.
not for Codex: the credits don't cover Codex through API keys.
first month is prorated: my first grant was $13 and it expired the next day. use it fast.
one run, not a benchmark: 12 of my own transcripts, 30 October 2026. the seven cents is a list-price estimate, not an invoice.

the rest of this LAB

the reasoning effort study. a new study tested which reasoning effort to pick. on tasks the lowest effort already solves, the top effort costs up to 42 times as much. starting low and stepping up only when it fails matched or beat top effort. one author, Zain Hasan, says they spent $25,000 on it. my take: start low, go up when it fails, and check the effort label before you compare models off those charts - the index charts run on max effort. it made me rethink BuilderBench too: i'm going to rerun it on high instead of extra high, and i'll show you if the scores move.
Claude for Legal got a gallery. ready-made artifacts for legal work, like contract review, regulatory horizon scans and cite checks. you open one from the gallery and it's built for that exact job. i have two law degrees (2018 and 2021) and did this stuff by hand for years. lawyers bill in six-minute increments. my take: it kills the billable hour - and thank god for that.
Gemini 4 Carbon is number one. it tops the Artificial Analysis index at 59, one point ahead of Opus 5.5 at 58. but it only leads 4 of 9 evals, and Opus still wins the coding agent index, 62 to 59. you can't use it yet - cyber defenders get it first. my take: use it before you believe it.
you can text your Dot now. ChatGPT's Dot works over iMessage or RCS, as a preview for Pro users in the US. reports say the texts aren't end-to-end encrypted, a company called Linq handles them, and there's no plain SMS. my take: go to them - more people go to their messages app than to ChatGPT.
AI-run content. Eric Siu let AI run his YouTube without his taste on it and got about 1,000 views a video, against 13,000 to 20,000 when he wrote them himself. Enrico Tartarotti says his channel got 3 million long-form views on videos edited with agents. he also sells a review tool. my take: the edit can be AI, the opinion can't.
the chat and work merge. OpenAI's Tibo says it's "not rolling out just yet". he says the plan is to bring chat and work together at some point this year, and to take it slowly. as of 30 October 2026.

sources

next step

watch LAB 0032 on YouTube - the practical is near the end.

no idea what to point your credits at? give your agent the link to this video and ask it for ideas.

want the next one? get the next field notes. optional - these notes stay free to read.