How does condense actually shrink the bill?

Your agent re-sends the whole conversation to the model on every turn, so by mid-session you are paying again and again for text the model has already seen. On our own traffic 67.7% of the bill is re-reading old conversation, and only 0.2% is new text you just typed. condense shortens that history on its way to the model. Helene 1 handles the newest turn, dropping the parts that are not pulling their weight and passing the rest through word for word, fast enough that you will not notice it. Adeline 1 goes back over older work that is already finished and rewrites it to about a tenth of its length. What always survives: your instructions, what you asked for, and the answers you got back. Code, patches and error messages are never touched.

Will condense ruin my cache hit rate?

No. condense is built around the cache rather than in spite of it. We split each request into three parts. The oldest part is shortened once, then locked and never touched again, so the provider keeps serving it from cache cheaply for the rest of your session. The middle is the part still settling, which is due to be rewritten anyway. The newest part is this turn's text, which Helene trims before it is ever cached at all. Because the locked part only ever grows at the end, shortening what comes after it cannot spoil what is already cached. Our measured cache hit rate is 98.6%, and the published benchmark prices every run from real provider bills, so a broken cache would show up as a bigger bill rather than hide.

How much will I actually save?

It scales with how long your sessions get, because the saving comes from the history an agent re-sends every turn. Our published live run removed 37.3% of the bill on chains averaging ~179k tokens; sessions peaking past 800k save 66–72%. Dollar-weighted across all our own spend it comes to 66%. Chains under roughly 4k are too short to be worth compressing. The benchmark is public, so you can run it on your own sessions before you pay us anything.

How does condense compare to headroom, caveman or RTK?

We published a head-to-head against headroom: minmax-bench replays the same real session through both, reads back the real provider bill, and prices both the same way. condense removed 37.3% of the bill against headroom's best mode at 13.8%, close to three times as much. Compared only over the shorter sessions headroom managed, condense is at 28.0%, still about twice their best. The difference is built in: headroom's default setting keeps its cache clean but never touches the piled-up history that is most of the bill, while its more aggressive mode spoils the cache and ends at 1.8%. RTK, caveman and ponytail are not really competitors. They go after the model's replies and tool output, which is a different part of the bill, and they work alongside condense rather than replacing it.

What counts as a token saved?

Only the context we remove. If condense strips 4,000 tokens from a request before it leaves for Anthropic or OpenAI, that is 4,000 tokens saved: tokens the provider never saw and never billed you for. Whatever passes through untouched costs you nothing here.

Do I still pay Anthropic or OpenAI?

Yes. condense sits in front of your provider and uses your own API key, so the provider still bills you directly, just for far fewer tokens. That is where the saving lands. What you pay us for doing the compacting is either a flat monthly tier or a metered rate on the tokens we save.

Every tier is a weekly pool of tokens saved. Free gives you 25M a week. Starter is $5/month for 500M, Pro is $20/month for 2.5B, Max is $50/month for 10B. The pool refills every week, and when it is empty your requests keep flowing to your provider uncompressed until it refills. You lose the savings, not the service, and nothing extra is charged. Calling the API directly instead of subscribing is metered on the same tokens saved: $0.90 per million for Adeline 1, $0.40 for Helene 1. See the full pricing page for team plans and how to size a tier.

Does condense store my prompts or my provider key?

Your upstream provider key is never stored. Only a sha256 fingerprint reaches the usage ledger, enough to attribute requests without holding the secret. We never train on your data. Content is deleted on a 7-day window, all storage and processing stays in the EU, and zero data retention is available on request. Read the security page for the full posture.

Helene 1.1: the rate you ask for

Helene 1.1 removes more of your context for the same answer quality, and it removes the amount you asked for.

Helene 1.1 is now available on condense.

It is a straightforward improvement on Helene 1. Compress the same amount and it answers more accurately. Ask it for a compression rate and it delivers that rate. Both hold across every model and language we tested.

+7.1pts

over bear-2 at matched compression

+12.5pts

over kompress-v2 at matched compression

50.5%

removed when asked for 50%

The numbers below come from the public Lost in Compression benchmark, run against bear-2 and kompress-v2 over ten languages and four answering models.

Accuracy at matched compression

Compressors do not return the rate you request, so comparing them at the same requested rate compares different amounts of surviving text. We compare at matched compression instead: each competitor setting paired against the Helene setting that leaves the same number of tokens behind.

Higher and further left is better: more accuracy on fewer surviving tokens. Every point is the mean of four answering models (claude-haiku-4-5, claude-sonnet-5, gpt-5.6-luna, gemini-3.6-flash), n = 3,000 per model per cell, all ten languages, as-scored.

Helene sits above both competitors at every matched point, and only Helene runs below 0.48 surviving text at all. Against bear-2 the margin is +7.09 accuracy at the ~0.50 band and +3.73 at ~0.75; against kompress-v2 it is +12.54 at ~0.50 and +12.01 at ~0.75.

The residual gaps in surviving text run both ways, and the large ones run against us. kompress-v2 holds 13.40pp more text than Helene at the ~0.50 band and 5.46pp more at ~0.75, and scores lower at both. Against bear-2 the tilt is ours and small, 1.02pp at ~0.50 and 3.28pp at ~0.75, and correcting for it does not change the ranking.

It holds across models

A win on one scoring model could be a quirk of that model, so we handed the identical compressed text to four unrelated ones and scored it again.

Every model scored the identical compressed text, and every row is the mean of the ~0.50 and ~0.75 keep bands, n = 3,000 per model per band, all ten languages, as-scored. The figures on the right are Helene’s margin over each competitor.

Helene is ahead on all four, by 3.88 to 7.08 points against bear-2 and 9.43 to 17.67 against kompress-v2.

The rate problem

Ask Helene to remove half your tokens and it removes half, in every language. That is what makes a token budget in front of it work: a compressor is only integrable if you can say “give me half” and get half.

How much text actually survives at one fixed setting, one dot per language.

At one fixed setting, how much text Helene actually removes varies 1.2× across the ten languages. kompress-v2 varies 1.7×, bear-2 23.6×.

Asked to remove 50% of the tokens, Helene removes 50.5% and kompress-v2 removes 21.6%.¹ bear-2 takes an aggressiveness setting rather than a rate.²

How we measured

The benchmark is Lost in Compression (Mantas Lukauskas), built on Belebele and FLORES-200 passages (facebook/belebele, CC-BY-SA-4.0): 10 languages, 300 items each, 3,000 paired items.³ It is third-party and public. Its harness, items, prompts, scoring and pairing are unmodified. The answering calls are not.⁴

bear-2 was called live through The Token Company’s API. kompress-v2 is the benchmark’s own published compression cache, byte-identical to our arm. Helene ran the 1.1 checkpoint with a per-document budget matched to the competitor’s spend. The rate results need no judge and no harness: three API calls and a token count in o200k_base reproduce them.

Over all cached outputs Helene emits 0.000% novel tokens and preserves source order on 100% of documents, in every language. It deletes at word granularity, so what comes back reads as an extract rather than an excerpt. It never rewrites.

Try Helene 1.1

Helene 1.1 is serving on condense now. Point your agent at the proxy and it compresses on the way through: same model, same API, fewer tokens on the bill.

Run your agent on the cheaper bill.

Notes

kompress-v2 exposes a keep threshold rather than a rate; the threshold-to-keep-rate table published on its model card bottoms out at 0.705 surviving tokens.
The Token Company document aggressiveness as controlling how much content is removed, banded by use case: Light 0.05 to 0.15, Moderate 0.15 to 0.4, Aggressive 0.4 to 0.9. No mapping from the setting to how much text actually survives is published.
Compression is not uniform across the ten: kompress-v2 keeps 0.911 of the Chinese text at the setting where it averages 0.6267, against 0.543 to 0.721 on the other nine languages.
gpt-5.6-luna ran at max_tokens=16 rather than the benchmark’s 8, because Bedrock (mantle) rejects a cap of 8.