Listen to this episode on Spotify or watch it on YouTube.
Welcome back to The Cost of Intelligence, our comedy news show about the money behind AI. Each episode, Ben Sand reads the industry’s invoice and asks one question: are you getting more intelligence per dollar?
This week is a special. Most of the show is one open-source project: gbrain, a free shared memory for AI agents, built by Garry Tan, the head of Y Combinator. His own tests say it cuts reranking costs 60%, agent calls 32% per task, and retrieval 86% against pasting the whole chat. They also say making agents check it first costs 10 to 55% more per turn. Plus: Tan’s plan to rewrite its search engine in Rust, OpenAI’s GPT-6.2 Astra, and gbrain’s new paid plans. Here’s the written version, with every source linked.
The gbrain special
First, the basics, straight from the repo.
gbrain is a memory layer for AI agents. It’s free, under the MIT license, and Garry Tan wrote it; his README says he built it to run his own agents (gbrain on GitHub). It keeps your notes, people and companies as plain files and indexes them. Then Claude Code, Codex and other agents search that memory before they answer, instead of starting from zero. It has more than 33,000 stars on GitHub. Tan says his own gbrain holds more than 155,000 pages. It went public in April.
The disclosures. Tan owns gbrain. A paid hosted version carries his name, and its site says it’s backed by Y Combinator (gbrain.io). So every savings number today is the maker’s own: from Tan’s repos, his changelog and his pull requests, at list prices. We found no measured before-and-after from an outside user. We looked. An audit where the auditee brought the calculator.
A memory tool cuts a bill by sending the model less text. Fewer tokens in, a smaller invoice. Finance calls that math. Engineering calls it a breakthrough.
But Tan’s own tests also measured the cost of using it. A one-line rule told agents to check gbrain first. Three agents obeyed in 30 of 30 turns; without the rule, they checked in 4 to 6 of 30. Following the rule cost 10 to 55% more per turn, and the pass rate didn’t move. It’s a small test: ten tasks, one run each, from his report dated 7 November (the report). Checking with the manager before every answer. Every company knows that cost.
On gbrain’s own BEAM 10M test, a plain file-search agent scored 67.6% and gbrain scored 58.9%. The test can’t tell them apart (gbrain-evals).
Per-dollar verdict: it saves money where the maker measured it, and costs money where the maker also measured that.
The big picture
The tracker at tokensperday.com starts from a July estimate of 360 trillion tokens a day and grows it at its own assumed rate. Grown to 18 November, that’s about 472 trillion a day, or about 57,000 tokens for every person on Earth.
Epoch AI counts about 34 million AI chips sold since 2022, measured in H100 equivalents: five times the count at the end of 2024.
The data centers behind them draw about 30 gigawatts, our estimate from Epoch’s data center tracker.
Last year, AI-focused data centers used about half a percent of the world’s electricity, by our math on International Energy Agency data. That’s about a tenth of a percent of all the energy we use.
Microsoft, Amazon, Alphabet and Meta spent about $202 billion in capital spending last quarter, the bucket that holds data centers and chips, more than double a year ago (company Q3 filings, cash capex).
The news: Tan’s Rust plan
Tan posted on X early Tuesday that a “GBrain backend engine” will be “written in Rust” and “will replace postgres+pgvector or pglite as the backing store.” He says it’s coming “in the next few weeks.”
That’s a plan, with no cost or speed number and no test. In the next few weeks: the roadmap’s favorite unit of time.
Per-dollar verdict: a rewrite is a promise. Call us when there’s a receipt.
The dial
In May, Tan posted that you can “tokenmax $10k/mo” with an agent plus gbrain and get “the AI that everyone will have in 2028 for $100/mo.”
His own repo sells the same dial. It lists gbrain’s search modes, and at 10,000 queries a month on Anthropic’s Sonnet 5.5, the conservative mode’s reading bill is $80 and the tokenmax mode’s is $400: gbrain’s own estimate (search modes). Five times the bill, same project. The gas pedal and the brake, one maker.
Per-dollar verdict: the dial is yours. So is the invoice.
GPT-6.2 Astra
OpenAI released GPT-6.2 Astra and Sol on Tuesday. OpenAI says they’re in ChatGPT Work and Codex for Plus, Pro, Enterprise and education users, and in the API (OpenAI).
Astra’s API costs $8 per million input tokens and $40 per million output tokens, 20% below GPT-6 Astra (OpenAI pricing).
OpenAI cites Agents’ Last Exam, a benchmark from Berkeley and Snorkel AI, and says Astra scores 5.5 points above Claude Fable 5.2 on it. The public leaderboard doesn’t list Astra yet, and there, Fable leads on pass rate.
OpenAI says Astra did it at 5% of Fable’s cost on that test, though the list prices are the same. OpenAI adds that Fable’s results exclude nine tasks that used a fallback model.
On Epoch’s Capabilities Index, Astra and Fable 5.2 both show 174. Epoch’s raw scores put Astra first, by a hair.
Same sticker, same score, and a launch post that says it’s cheaper.
Per-dollar verdict: 20% off the old Astra is a receipt. 5% of Fable is a claim.
Bonus story: gbrain’s hosted plans
gbrain now has a paid hosted version at gbrain.io. Its pricing page lists Personal at $49 a month, Team at $199 a month with $100 of AI credit every month, and Enterprise at a custom price (gbrain pricing). The open-source version stays free to run yourself.
Per-dollar verdict: free if you run it. $49 a month if someone else does.
Department of Token Efficiency
Chainsaw fueled. Spreadsheets open. All three cuts are gbrain’s, and all three tests are Tan’s.
Cut one: the reranker. A reranker re-sorts search results before the model sees them. gbrain switched its default, for brains that hadn’t picked one, to a cheaper one (changelog).
The cost fell from $1.51 to 60 cents per thousand queries, 60% cheaper.
It found the right page among its top results for 298 of 300 questions. The old one found 299.
Sixty percent off, one question lost. A coupon you didn’t have to clip.
The fine print: Tan’s own test, one benchmark, at list prices. An earlier trial that cut the candidate list lost at least three questions and was thrown out (preregistration).
Cut two: the tool menu. Every turn, an agent rereads its menu of tools. The release cut gbrain’s menu and trimmed its search results (pull request).
The menu went from 13,077 tokens to 5,568.
Cost per task fell from 11.8 cents to 8 cents, 32% less. A team running 1,000 tasks a day saves about $38 a day, our estimate.
A menu you can finally read without a nap. (There’s a goblin. Watch the video.)
The fine print: the maker’s test, six models. It finished 454 tasks of 600; before, 441. That gap is within the noise. The saving comes from the whole release: a shorter menu and leaner search results.
Cut three: retrieval. On Anthropic’s Sonnet 5.5, pasting a whole chat history cost 17.3 cents a question. Retrieving only the relevant pieces cost 2.4 cents: 86% less. gbrain got 139 of 150 right; the full paste got 144 (benchmark).
Skip the novel. Read the index.
The fine print: Tan’s own test, at batch prices. It holds for one question per history. Indexing costs 1.5 cents per history. And the pasted history fit the model’s window.
Ask your CFO if token efficiency is right for you.
A word from ourselves
Your measuring tape for every maker’s own test. Your second opinion on every savings claim. Your coupon for the tokens you didn’t spend. The people who keep the landlord honest.
Strong Compute plugs into 27 compute clouds, from Nebius to Modal to Vultr, and 10 AI vendors, from Anthropic to Fireworks AI to Novita AI. Bring your own keys and see exactly where the GPU hours and tokens go.
Because somebody has to read the bill.
Try it at strongcompute.com
Keep your context short and your invoices shorter. See you tomorrow.


