30 teams, one weekend, 4 billion tokens

We sponsored basedhack in Vilnius. Every team ran through condense, and the proxy logged the whole weekend.

Last weekend we were a tech partner at basedhack — an invite-only AI hackathon in Vilnius, presented by basedcollective and Google Cloud. Thirty teams shipped in two days. We handed every builder a condense key, and for two days the proxy was the compaction layer under the whole room. Here is what that looked like from inside, plus where Adeline runs now, who just backed us, and the one builder we knighted for it.

The weekend in numbers

Aggregated across every hackathon account, measured on the live proxy. No per-user breakdown, nothing identifiable — just what real agent traffic looked like with twenty-plus builders hammering it at once.

tokens compacted

Conversation history compressed before it ever reached a provider, summed across the weekend.

teams hacking

Invite-only, two days, one venue. Twenty-plus builders ran their agents through condense.

from a single builder

One builder accounted for a quarter of the weekend's compaction on their own. Agents at full tilt add up fast — enough, it turned out, to earn a title.

The bill the room never paid

Those 4 billion tokens are tokens that never had to be sent, paid for, or waited on. What that is worth depends entirely on how you would have paid for them otherwise.

$19k

saved at list rates

The raw counterfactual — the same traffic priced with no prompt caching upstream. The most the weekend could have cost.

$3k

saved against a warm cache

The conservative floor — assuming a perfectly warm cache upstream, the cheapest the same traffic could possibly have been.

Real bills land between the two. Either way, a room of thirty teams ran two days of agents and a four-figure-to-five-figure slice of provider spend simply never happened.

Adeline ran on our own GPUs

All of that compaction runs on Adeline, our model built for it — and this was the first event Adeline v1 served entirely from our own hardware. v1 used to run on rented serverless GPUs; it now lives on dedicated cards inside our cluster, which is faster to reach and ours to scale. For the weekend we warmed the pool up to three cards so a room full of agents compacting at once never queued.

That headroom is the point. Compaction stays fast even under load, and at hackathon pace — where nobody wants to watch their context window summarize — the latency is what people actually felt.

Joining the AWS Global Startup Program

The weekend also marked the start of something we are glad to be part of: condense is now in the AWS Global Startup Program. That backing is what underwrites the reserved GPU capacity Adeline now runs on — room to keep the warm pool warm, bring v1 back online, and grow the card count without the bill dictating the model. The serving story above exists because of it.

We knighted Sir Condenser

That single builder from the numbers up top — the one who compacted a billion tokens alone, a quarter of the whole weekend — had more than earned a title. So we did the only reasonable thing: a saber, a bent knee, and a new name. Arise, Sir Condenser.

A kneeling basedhack builder is knighted with a saber in front of a 'solo-ution' slide. — The knighting of Sir Condenser — one billion tokens compacted, basedhack Vilnius.

Thank you

basedhack was put together by basedcollective with Google Cloud, and it was a good room to be in. We also got to collaborate with Requesty over the weekend to let some of their requests flow through condense. Thanks to the fellow tech partners — ElevenLabs, AXIOMETA, Cursor, openmail, and Hostinger — and to the ecosystem partners who made it happen: Go Vilnius, bek, Inovo VC, Firstpick, and Brite. And to the thirty teams who pointed their agents at us for two days and trusted the proxy with their context: you were the load test we could not have run ourselves.

See you next week.

Run your agent on the cheaper bill.

Get access

30 teams, one weekend, 4 billion tokens.