Scaled Cognition is an AI model lab building foundation models that aim to eliminate hallucinations in enterprise settings, starting with customer service. Its flagship model, APT (Agentic Pretrained Transformer), is positioned as delivering human-like conversational capabilities with policy-adherent behavior, and is available for VPC and self-hosted deployment so enterprises can run it without ongoing reliance on third-party model providers. The pitch is aimed at operators running high-stakes workflows where a wrong answer has real cost, with early production deployments inside Fortune 500 companies across financial services, healthcare, telecom, and insurance.
This launch matters because the company just closed a $100 million Series A led by Khosla Ventures , capital being directed at expanding the research team and accelerating enterprise rollouts. Scaled Cognition frames its approach as engineering reliability into the model architecture rather than wrapping a frontier LLM in guardrails, and claims APT is smaller, faster, and cheaper to run than the frontier models it benchmarks against. Around the model sits a platform layer with agentic tooling, simulation and evaluation frameworks, and live agent monitoring, with Genesys both deploying APT inside Genesys Cloud for virtual agent capabilities and investing in the company.
The team has direct history in this space. Co-founders are CTO Dan Klein, a UC Berkeley professor and natural language processing researcher, and CEO Dan Roth, who previously built and sold an early agentic AI company to Microsoft (Semantic Machines, where Roth went on to lead conversational AI as a CVP). Customer experience is the wedge, but the stated ambition reaches further into the broader business process outsourcing market, including IT support, HR, and finance functions that enterprises are starting to bring back in-house with AI.
Comments (14)
Sign in to join the discussion.
P
Priya KothariJun 29, 2026
"LLMs lie. We build models that tell the truth." is a banger opening line but also a legal team's nightmare. Hope your eval suite is as bold as your copy.
T
Tomás BregmanJun 29, 2026
Reliability is the actual moat in this category and nobody wants to do the unsexy work. Curious what "super-reliable" means on paper though, like is there a public benchmark or are we vibing.
L
Lina OseiJun 29, 2026
ok wait, you cannot just say "models that tell the truth" and then link a landing page with zero demo video. where is the receipts tab
M
Marek DvorakJun 29, 2026
Before I tweet "interesting" I'd want to know retention on the enterprise pilots and what churn looks like once the novelty wears off. Truthful models sound great until procurement asks for a SOC2 binder.
R
Renata FigueiredoJun 29, 2026
Tagline rewrite, on the house: "The model that says I don't know." You're welcome.
A
Adaeze NwosuJun 29, 2026
stealing this for my own startup, sorry not sorry
K
Kenji HoshiyamaJun 29, 2026
Staff eng brain itching here: are you doing this with a new pretraining objective, post-training tricks, or retrieval gymnastics? "Reliability" is a word that hides at least three different research bets.
S
Sasha BerglundJun 29, 2026
The tweet structure is doing a lot of work: punchy claim, funding flex, then the why. Clean. Whoever drafted this earned their espresso.
Y
Yusuf DemirJun 29, 2026
Every "we solved hallucinations" pitch I've seen ages like milk in a hot car. Talk to me after a public red team.
C
Cam EluwaJun 29, 2026
to be fair they didn't say solved, they said reliable. but yeah the bar is on the floor and people still trip
H
Hanna LiedtkeJun 29, 2026
Would love to chat about what verticals you're piloting in first. On deadline this week if you have ten minutes.
D
Dimitri VolkovJun 29, 2026
watching this drop while I'm 3 weeks out from my own reliability layer launch and feeling deeply normal about it. cool raise though.
M
Mei-Lin TsaoJun 29, 2026
naive question but if the model tells the truth, what happens when the user's prompt contains a false premise? does it just argue with them all day
B
Bilal QureshiJun 29, 2026
this is genuinely the hardest UX problem in the category and a 22 year old just dropped it in a reply, incredible
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Scaled Cognition
Super-Reliable Intelligence
Super-Reliable Intelligence
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Scaled Cognition is an AI model lab building foundation models that aim to eliminate hallucinations in enterprise settings, starting with customer service. Its flagship model, APT (Agentic Pretrained Transformer), is positioned as delivering human-like conversational capabilities with policy-adherent behavior, and is available for VPC and self-hosted deployment so enterprises can run it without ongoing reliance on third-party model providers. The pitch is aimed at operators running high-stakes workflows where a wrong answer has real cost, with early production deployments inside Fortune 500 companies across financial services, healthcare, telecom, and insurance. This launch matters because the company just closed a $100 million Series A led by Khosla Ventures , capital being directed at expanding the research team and accelerating enterprise rollouts. Scaled Cognition frames its approach as engineering reliability into the model architecture rather than wrapping a frontier LLM in guardrails, and claims APT is smaller, faster, and cheaper to run than the frontier models it benchmarks against. Around the model sits a platform layer with agentic tooling, simulation and evaluation frameworks, and live agent monitoring, with Genesys both deploying APT inside Genesys Cloud for virtual agent capabilities and investing in the company. The team has direct history in this space. Co-founders are CTO Dan Klein, a UC Berkeley professor and natural language processing researcher, and CEO Dan Roth, who previously built and sold an early agentic AI company to Microsoft (Semantic Machines, where Roth went on to lead conversational AI as a CVP). Customer experience is the wedge, but the stated ambition reaches further into the broader business process outsourcing market, including IT support, HR, and finance functions that enterprises are starting to bring back in-house with AI.
Comments (14)
"LLMs lie. We build models that tell the truth." is a banger opening line but also a legal team's nightmare. Hope your eval suite is as bold as your copy.
Reliability is the actual moat in this category and nobody wants to do the unsexy work. Curious what "super-reliable" means on paper though, like is there a public benchmark or are we vibing.
ok wait, you cannot just say "models that tell the truth" and then link a landing page with zero demo video. where is the receipts tab
Before I tweet "interesting" I'd want to know retention on the enterprise pilots and what churn looks like once the novelty wears off. Truthful models sound great until procurement asks for a SOC2 binder.
Tagline rewrite, on the house: "The model that says I don't know." You're welcome.
stealing this for my own startup, sorry not sorry
Staff eng brain itching here: are you doing this with a new pretraining objective, post-training tricks, or retrieval gymnastics? "Reliability" is a word that hides at least three different research bets.
The tweet structure is doing a lot of work: punchy claim, funding flex, then the why. Clean. Whoever drafted this earned their espresso.
Every "we solved hallucinations" pitch I've seen ages like milk in a hot car. Talk to me after a public red team.
to be fair they didn't say solved, they said reliable. but yeah the bar is on the floor and people still trip
Would love to chat about what verticals you're piloting in first. On deadline this week if you have ten minutes.
watching this drop while I'm 3 weeks out from my own reliability layer launch and feeling deeply normal about it. cool raise though.
naive question but if the model tells the truth, what happens when the user's prompt contains a false premise? does it just argue with them all day
this is genuinely the hardest UX problem in the category and a 22 year old just dropped it in a reply, incredible