←
The Token Company (YC W26)34 posts
The Token Company (YC W26) avatar

The Token Company (YC W26)

@thetokenco

Building custom ML models for compressing raw LLM inputs

◷ Joined January 2026▧ 7 Photos and videos
The Token Company (YC W26) avatar
The Token Company (YC W26)@thetokenco·Jun 9

bear-2-safety is a fully extractive compressor, it retains surviving tokens in their original order without any paraphrasing. To implement bear-2-safety, just run one compression model call in front of Llama-Guard, ShieldGemma, or any other safety classifier.

Post attachment
The Token Company (YC W26) avatar
The Token Company (YC W26)@thetokenco·Jun 9

Does compression reduce safety classification accuracy? Across five safety benchmarks: OpenAI Moderation, WildGuard, Aegis, BeaverTails, and PKU-QA, F1 retains quality across SoTA classifiers.

Post attachment
The Token Company (YC W26) avatar
The Token Company (YC W26)@thetokenco·Jun 9

Our core approach targets asymmetric retention. The vast majority of input tokens don’t influence safety classification pipelines, so we focus our token allocation on problematic spans. Potentially harmful content is maximally retained for safety classifiers, while benign filler gets compressed by >50% so we pass less content to more expensive safeguard models.

Post attachment
The Token Company (YC W26) avatar
The Token Company (YC W26)@thetokenco·Jun 9

We’re introducing bear-2-safety for private preview, a compression model to optimize cost and performance for LLM safety moderation pipelines, such as Llama Guard and ShieldGemma. Bear-2-safety selectively deletes tokens before safety inference, lowering costs while maintaining classification accuracy.

Post attachment
The Token Company (YC W26) avatar
The Token Company (YC W26)@thetokenco·May 22

👀👀

🦔Microsoft canceled its internal Claude Code licenses this week after token-based billing made the cost untenable, even for a company with effectively infinite cloud resources. Uber's CTO sent an internal memo warning the company burned through its entire 2026 AI budget in just four months. American AI software prices have jumped 20% to 37%, and GitHub (owned by Microsoft) is dropping flat-rate plans for usage-based billing across its products. My Take The AI subsidy era is ending in real time. The same company that put $13 billion into OpenAI and built the Azure infrastructure powering most of Anthropic's compute just looked at the bill from a competitor's coding tool and decided it was not worth paying. That is not a productivity failure on Anthropic's end. Token-based pricing is forcing every enterprise customer to confront the actual cost of running these models at scale, and the number turns out to be far higher than the flat-rate experiments suggested. This ties directly to my Gemini Flash post yesterday. Anthropic, OpenAI, and Google all raised effective prices in the last six months. Enterprises that built workflows assuming AI costs would keep falling are now watching annual budgets evaporate in months. Two outcomes look likely from here. Either enterprises scale back AI usage to fit budgets, which slows the revenue ramp the labs need to justify their valuations ahead of IPOs, or the labs cut prices and absorb the losses, which makes the unit economics worse at exactly the wrong moment. Both paths land in the same place, the numbers stop working, and somebody has to take the writedown. Hedgie🤗

Post attachment
The Token Company (YC W26) avatar
The Token Company (YC W26)@thetokenco·May 14

Save on your LLM bill with @opencode and @thetokenco

ricky@drf0kMay 13

pushed an update to the @opencode plugin using the @thetokenco compression model for reducing input tokens for every prompt. i've been meaning to ship an update every since plugins got updated! > added compression status to sidebar widget > added opencode command for easier adjustments i set it, forget it, and use it every single day

0:29
The Token Company (YC W26) avatar
⟳ The Token Company (YC W26) retweeted
The Token Company (YC W26)@thetokenco·May 13

The Token Company is now HIPAA compliant! Customers using The Token Company's models to compress LLM inputs can now securely process protected health information. Our compression models reduce bloated context from LLM inputs making LLM models perform better and cheaper under long inputs

Post attachment
Rasmus avatar
⟳ The Token Company (YC W26) retweeted
Rasmus@rasmus_up·May 7

Claude Code making really pretty animations

0:05
The Token Company (YC W26) avatar
The Token Company (YC W26)@thetokenco·Mar 12

Evaluating compressed prompts just got easier. You can now compare LLM outputs before and after compression in our Compression sandbox to directly evaluate your use case.

0:20
ricky avatar
⟳ The Token Company (YC W26) retweeted
ricky@drf0k·Mar 7

i built a plugin that saves you hundreds of dollars in @opencode by using the @thetokenco compression model, it shrinks the amount of tokens in your input query before hitting the models the queries are faster and cheaper, all while maintaining output quality

0:11

The Token Company (@thetokenco) builds LLM input optimization to lower costs, reduce latency, and improve accuracy. Congrats on the launch, @OtsoVeistera! ycombinator.com/launches/Pb3…

0:37
The Token Company (YC W26) avatar
The Token Company (YC W26)@thetokenco·Mar 6

Compression absolutely crushed SEC filings Does anyone actually read these? I don't know, but we let our bear-1.2 LLM compression model do that on FinanceBench (150 real SEC filing questions) Bro got up to 84.7% accuracy (vs 82% baseline) with 20% token reduction. Saves costs, boosts speed, crushes financial analysis Removing bloat makes the model perform better

Post attachment
Dan avatar
⟳ The Token Company (YC W26) retweeted
Dan@aidaniil·Mar 4

One of my batchmates is starting his fundraise earlier because he needs to take his highschool exams right before demo day I've never been aura farmed on so much in my life