The Token Company
YC W26San Francisco, US · Founded 2025 · 2 employees · 5 known investors
The Token Company builds a compression model called Bear-2 that reduces the size of LLM inputs such as documents, websites, and transcripts before they enter an LLM context window. The product enables developers and AI applications to pass more context with fewer tokens to language models.
Also known as Otsofy · Token Company
Founders & leadership· Y Combinator alumni (W26)
The Token Company was founded in 2025 by Otso Veisterä.
Investors · 5
The YC application· Winter 2026 batch
The Token Company (YC W26) Launch Video
TheTokenCompany
Alumni Q&A
- What is the core problem you are solving? Why is this a big problem? What made you decide to work on it?
- The founders address context bloat and inefficient LLM inputs by preprocessing them to remove unneeded tokens, aiming for faster, cheaper, and more accurate responses. They note this preprocessing is rarely done in practice and chose the problem because it operates at a more fundamental layer that affects every LLM application.
- What is your long-term vision? If you truly succeed, what will be different about the world?
- The founders aim to optimize every LLM request at the token level before it reaches a model, positioning the company as a neutral, cross-model efficiency layer above competing providers. They argue this layer stays defensible as long as multiple labs exist, since labs are incentivized to sell more tokens rather than reduce usage.
Summarized from the founders' answers on their Y Combinator profile.
Company profile
researched Aug 2026The Token Company develops middleware that compresses inputs to large language models before they reach the model. The service is delivered as an API that sits between an application's prompt and its LLM provider: a small machine-learning classifier scores each token in the input and deletes the tokens least likely to affect the model's output. The company states that nothing is summarized, paraphrased or generated, so the remaining text stays verbatim and in its original order, preserving citations, code, numbers and JSON structure, and keeping output deterministic so prompt caches remain valid. Users can tune how much is removed via an aggressiveness parameter. [0][1][2]
The company describes typical compression of 10-50% of tokens with accuracy maintained, sub-50ms inference for processing raw inputs, p95 API latency of about 150ms, and compression of 100,000 tokens in under 100ms. Named models include Bear-2 and a safety-classifier variant, Bear-2-Safety; a third-party profile also references earlier bear-1 and bear-1.2 models used to remove low-signal tokens and compress background knowledge before it enters the context window. Integration is offered as a one-line wrapper around an existing OpenAI or Anthropic client, distributed via a Python package (the-token-company) and a Node client. [0][1][2][5]
The Token Company positions itself for regulated data: it reports SOC 2 Type I attestation (audited by Advantage Partners, with Type II expected in September 2026), HIPAA readiness with a signed BAA, optional account-wide zero data retention, and on-premise or customer-VPC deployment on its Enterprise tier so raw prompts do not leave the customer network. [0][1]
Founding story
Otso Veisterä founded the company to work on context bloat and inefficient LLM inputs, describing the problem as more fundamental than application-layer products because it affects every LLM request, and noting that token-level input optimization was not being done in practice. A third-party profile describes him as a solo founder, 18 years old, with a 3-month FR8 accelerator stint, four months as an EIR at Lifeline Ventures, a hackathon award, a national physics championship and a Z Fellow designation, and no prior full-time engineering roles or company-building experience. [2][7]
Business model
B2B API service. Customers call a compression API with long-form text and receive compressed output for downstream LLM calls; third-party profiles characterize the model as usage-based/transactional SaaS built around the API, with an Enterprise option for on-premise or in-VPC deployment. One third-party profile cites usage-based pricing of $0.05 per 1M tokens. [0][3][5][7]
Revenue is generated primarily through transactional, usage-based access to the compression API; specific pricing plans are not disclosed by the company, though a third-party profile lists $0.05 per 1M tokens. Enterprise deployments add on-premise/VPC options. [3][5][7]
Traction
Published case studies report that in a 268,000-vote blind LLM arena run with customer Pax Historia (193B tokens/month), compressed prompts outperformed uncompressed prompts across models (+4.9% on Sonnet 4.5, +15% on Gemini 3 Flash) and lifted purchase volume by about 5%; a second case study with Helonic (YC F25) reports roughly 47,000 tokens saved per prompt on long-running construction-drawing agents. Company benchmarks cite CoQA accuracy improving from 93.3% to 95.3% with 8.2% fewer tokens using Bear-2, a 30% reduction in safety-classifier cost with Bear-2-Safety at preserved or improved F1, +2.7 percentage points on financial QA with up to 20% fewer tokens, and up to 37% faster end-to-end latency. A third-party review notes that beyond Pax Historia no other customers and no revenue figures have been disclosed. [0][1][2][7]
Latest developments
The company has published research on conversational-context compression with Bear-2 and introduced Bear-2-Safety for cheaper safety classification. It reports SOC 2 Type I attestation with Type II expected in September 2026. Third-party trackers record March 2026 venture funding from Y Combinator and Pioneer Fund, and state that proceeds are intended for new product development, market expansion and service enhancement. [0][1][5]
▸Full profile — market position, technology, go-to-market, geography, history, risks & controversies
Market position
The company frames itself as a neutral, cross-model efficiency layer sitting above competing model providers rather than an application-layer product, arguing that model labs are incentivized to sell more tokens rather than reduce usage. A third-party assessment identifies open-source prompt-compression research such as Microsoft Research's LLMLingua as a direct conceptual competitor and notes the risk that LLM API providers absorb compression natively. Third-party data providers classify it as an AI/cloud-infrastructure company at the angel/seed stage. [2][3][5][7]
The approach is deletion-only and non-generative: unlike summarization or small-LLM-based compression, it does not rewrite, paraphrase or invent text, which the company says makes it faster, cheaper, fully deterministic and safe for provider- and customer-side prompt caching while preserving citations, code, numbers and JSON. It also offers compliance features (SOC 2 Type I, HIPAA/BAA, optional zero data retention) and in-network deployment. [0][1][2]
Technology
A proprietary, non-generative machine-learning compression model (Bear-2, plus a Bear-2-Safety variant for safety classifiers; earlier bear-1 and bear-1.2 models are also referenced) acts as a token-level classifier that scores and deletes low-signal tokens such as boilerplate, filler, redundant phrasing and structural scaffolding. Because the model only deletes and never rewrites, output is fully deterministic and cache-preserving. Reported performance: below 50ms inference, p95 latency around 150ms, 100k tokens compressed in under 100ms, and typical token reductions of 10-50%. [0][1][2][5]
Go-to-market
Self-serve developer adoption via an API key, published documentation, a Python/Node client wrapper and drop-in integration with existing OpenAI and Anthropic clients, supplemented by founder-led outreach (the founder invites production LLM users to email directly), published research and customer case studies, and a Trust Center for compliance review. Enterprise sales cover on-premise/VPC deployment. [0][1][2]
Developers and companies running production LLM workloads with high token volumes and context bloat from chat history, RAG documents or large inputs, including document-processing, chat and RAG pipelines. Cited customer sectors include finance and construction; named users include Pax Historia (consumer LLM application, 193B tokens/month) and Helonic (YC F25, construction AI). Enterprises with regulated data requirements are addressed via HIPAA/BAA, zero-retention and on-premise options. [0][1][2][5][7]
Geography
Headquartered in San Francisco, California, United States, with customer applications reported across North America and Europe. [2][3][5][6]
History
The company was founded in 2025 and is based in San Francisco. It participated in Y Combinator's Winter 2026 batch (primary partner Jared Friedman), where it published a launch post describing its compression API; the founder states he did not apply and joined after YC partners reached out. Team size is listed as 2. [2][5][6][7]
Risks & controversies
A third-party evaluation flags a solo, first-time founder without a co-founder as a structural gap, the absence of a proprietary data moat or network effects, replicability of compression models by well-resourced teams, open-source alternatives such as Microsoft Research's LLMLingua, and the possibility that OpenAI or Anthropic implement compression natively in their APIs; it also notes limited disclosure beyond one named customer and no revenue figures. Reported funding figures differ substantially across third-party data providers. [3][4][5][7]
Compiled by commissioned research from 8 cited public sources — announcements, filings, and press listed under research sources below.
Key figures
latest reportedCompany-reported or press-reported figures, each dated to when it was claimed — not independently audited.
Competitors · 7
by search overlap
Companies competing with The Token Company for the same Google search keywords, organic and paid, via search-intersection analysis.
Timeline · 6
launches, deals, and filingsThe company raised an undisclosed amount of venture funding from new investors Y Combinator and Pioneer Fund, to be used for upcoming product development, expansion into new markets and enhancement of existing services.
The Token Company joined Y Combinator's Winter 2026 batch, with Jared Friedman as primary partner; the founder says YC partners reached out rather than the company applying.
Research release reporting that Bear-2 improved CoQA accuracy from 93.3% to 95.3% while cutting tokens by 8.2%.
Safety-focused compression model reported to cut safety classifier costs by 30% while preserving or improving F1.
Public launch post describing a drop-in API that uses a fast, non-generative ML model to remove low-value tokens from prompts, cutting token counts, latency and cost.
Externally attested SOC 2 Type I (audited by Advantage Partners) with Type II expected in September 2026, plus HIPAA compliance with a signed BAA, optional zero data retention and on-premise/VPC deployment.
Dated company events from announcements, filings, and press; legal rows summarize public dockets and regulator releases.
In the news
▸Research sources · 8
primary sources listed
- The Token Companythetokencompany.com · web
8 public sources were cited for this profile; the first-party ones are listed here.
Frequently asked questions
- What does The Token Company do?
- The Token Company sells compression middleware that strips low-value tokens from LLM prompts to cut cost and latency.
- Who founded The Token Company?
- The Token Company was founded by Otso Veisterä in 2025.
- Who are The Token Company's investors?
- The Token Company's investors include Failup Ventures, Inception Fund, Wave Ventures, Y Combinator, Visionaries Club.
- Where is The Token Company headquartered?
- The Token Company is headquartered in San Francisco, US.