A startup recently valued at $200 million is attracting multibillion-dollar funding offers after launching a generation of AI tools that it claims are more efficient than products from OpenAI and Anthropic.

TypeSafe AI was founded by former OpenAI researcher Diogo Almeida.

The company last week released an AI model called "Jev," positioning it as a cheaper alternative for performing certain tasks rather than relying on traditional large language models (LLMs) such as ChatGPT and Claude.

Jev is aimed almost entirely at software developers, and since the company ended its so-called "stealth mode" and made a public appearance, it has quickly drawn widespread attention and gone viral on social media.

Its product introduction video posted on social platforms received 40 million views in less than a week.

According to people familiar with the matter, potential investors have already proposed funding plans to TypeSafe AI, and the related offers could value the company at $10 billion or even higher.

This enthusiasm reflects a growing concern in the corporate world: the cost of deploying AI is rising rapidly.

TypeSafe is betting on the idea that many routine tasks currently handled by expensive general-purpose models can actually be completed by cheaper, more specialized systems.

If this idea gains market acceptance, it could pressure the business models of frontier AI companies such as OpenAI and Anthropic.

Does AI need to be more specialized?

Almeida said he came up with the idea for Jev four years ago while working at OpenAI.

At the time, he began to wonder: are chatbots really suitable for all the tasks AI might be asked to perform?

He recalled: "Suppose an AI-based economic revolution really happens, then of all the requests calling AI, how many are for human consumption... and how many are actually for computer consumption?"

TypeSafe's core view is that large language models were originally designed to communicate with humans, but are not suitable for many of the programmatic, repetitive, and high-frequency tasks that companies around the world are trying to automate with AI.

Almeida said: "ChatGPT is much smarter than me. But strangely, the work with huge economic incentives and the greatest value for automation is not being automated right now. Because AI currently performs poorly in these areas."

How does Jev work?

Jev does not generate sentences, explain text, or process images.

Instead, its goal is to complete decisions quickly and at low cost inside software applications.

The "classification" tasks it can perform include: determining whether a request should be approved, rejected, or manually reviewed; automatically assigning customer service tickets; conducting insurance risk assessments; and evaluating credit risk.

The model is mainly aimed at developers and is designed to be embedded in the backend of software rather than directly providing a consumer-facing chat interface.

A developer once connected Jev to a joke website called "AskJev," a name paying tribute to the discontinued search engine Ask Jeeves, and the website then quickly spread on social media.

Cost advantage

James Hardiman, a general partner at Silicon Valley venture capital firm DCVC, said TypeSafe's latest funding round was led by DCVC.

He said that because Jev greatly reduces computing costs, the company is already profitable, and its model usage cost is "orders of magnitude lower" than leading large language models.

Hardiman noted that an AI tool like Jev, used for very ordinary tasks, entered the market just as public concern over the potential risks of advanced AI was increasing.

He said: "Just a week ago, everyone was still talking about AI doomsday... In some ways, Jev makes that discussion more rational."

Reducing the cost of AI responses is key to Jev's market appeal.

The name Jev comes from the "Jevons paradox," an economic concept that when a resource becomes cheaper or more efficient, total consumption may actually rise because new use cases keep emerging.

Differences from large language models

Large language models usually require long chains of reasoning, which consumes a large amount of computing resources.

TypeSafe says Jev takes a different approach: it uses probability calculations to quickly choose a result from a limited set of answers.

Its underlying method is closer to early machine learning systems.

Such systems are usually more deterministic and can produce fixed outputs for inputs, so they do not suffer from the "hallucination" problem common in large language models.

Because it requires less computation, Jev also consumes fewer "tokens," the units of computation used by models to process text, thereby reducing costs.

TypeSafe says its cost per query is about one-hundredth that of large language models, and its processing speed is also faster.

The company's pricing is roughly: Jev: about 4.2 cents per million tokens; large language models: possibly several dollars per million tokens.

Market reaction

AI development platform Vercel said that within 24 hours of Jev's launch, interest generated by its paid developer accounts exceeded that of any previous model release, including leading models launched by OpenAI and Anthropic.

Another AI model aggregation platform, OpenRouter, said the number of tokens processed by Jev increased more than threefold over the weekend.

OpenAI co-founder Andrej Karpathy, who recently joined Anthropic, said on social media that Jev appears to have captured latent market demand for "simple, cheap, fast decision-making models."

He said frontier AI companies have long focused on pursuing higher levels of intelligence, and this area has therefore been "underinvested."

Question: Is Jev truly revolutionary?

However, TypeSafe's secrecy about Jev's training method, as well as its similarity to existing technologies, has led some industry insiders to question how innovative it really is.

Anastasios Angelopoulos, co-founder and CEO of AI model evaluation platform Arena, said: "I don't know what the difference is between these models and traditional 'zero-shot classifiers,' which are actually a fairly mature technology."

At present, Meta, Google, and Hugging Face all already offer AI tools capable of classifying data they have not been explicitly trained on.

TypeSafe has not disclosed exactly how Jev is trained.

The company says the model was trained using open-weight models and computer-generated "synthetic data," but the relevant details have been kept strictly confidential.

Funding and future plans

Almeida said TypeSafe completed a $40 million seed round more than a year ago, when the company reached a valuation of $200 million.

He declined to disclose the new funding round or specific investor information, but confirmed: "Investors are beating down our door."

Almeida said: "We are a group of missionaries, a spark of a revolution... Therefore, in the process of scaling the company, how to stay true to our original aspirations and not lose our way is something we must think carefully about."