TypeSafe AI emerges from stealth with $40M: Diogo Almeida (co-inventor of ChatGPT) wants to replace LLMs with composable models so that machines, not humans, use them
🔎 ChatGPT's co-inventor has just declared war on chat
While the entire industry piles onto the same equation — ever-larger models, trained to speak like humans, resold by the token via API — Diogo Almeida has just broken ranks. On September 16, 2026, his startup TypeSafe AI emerged from stealth with $40M in seed funding led by DCVC (Business Wire via Yahoo Finance, September 15, 2026).
His thesis boils down to one sentence, and it's a disruptive one: models that have become excellent at talking to humans are not the right models for running software. Behind this observation lies a company founded in 2024 in San Francisco with Erik Gafni and Sasha Sheng, and a first product with a strangely austere name: Jev.
What makes the story interesting is the résumé. At OpenAI, Almeida co-invented RLHF and worked on InstructGPT, ChatGPT, and GPT-4. In other words: he helped build the very paradigm he is now attacking. And DCVC put $40M on the table — a seed round in the 99th percentile of all AI seed rounds in history according to Dealroom (sample of 28,473 deals, September 2026).
Here's what you need to know, and what you need to verify before rewriting your architecture.
Key takeaways
- TypeSafe AI emerges from stealth on September 16, 2026 with a $40M seed round led by DCVC — in the 99th percentile of AI seed rounds in history (Dealroom, Sept. 2026).
- Founded in 2024 in San Francisco by Diogo Almeida (ex-OpenAI: RLHF, InstructGPT, ChatGPT, GPT-4), Erik Gafni, and Sasha Sheng.
- First model: Jev, a "composable machine-native intelligence" that returns structured decisions with calibrated confidence scores — not conversational text — with a claimed latency of under 100 ms and the ability to process hundreds of outputs in parallel from a single prompt.
- Published pricing: $0.39 per 1,000 workflows, compared with $3.31 for GPT-5.6 Luna (OpenAI) and $19.49 for Claude Haiku 4.5 (Anthropic) (SiliconANGLE, Sept. 2026).
- Caution required: the spectacular multiples (194x faster, 445x cheaper) are self-reported, measured on workflows designed by TypeSafe itself (TS2.tech, Sept. 2026).
Recommended Tools
| Tool | Primary use | Price (September 2026) | Ideal for |
|---|---|---|---|
| Jev (TypeSafe AI) | Structured decisions for software | $0.39 / 1,000 workflows | classification, triage, agent output review |
| GPT-5.6 Luna (OpenAI) | Conversational and agentic LLM | $3.31 / 1,000 workflows | complex reasoning, text generation |
| Claude Haiku 4.5 (Anthropic) | Fast multitask LLM | $19.49 / 1,000 workflows | conversational production, agent pipelines |
| Local models (Ollama) | Simple local classification | Free (excluding hardware) | sensitive data, zero marginal costs |
Prices collected by SiliconANGLE on September 16, 2026. Startup pricing changes fast: check TypeSafe's website before making any architecture decisions.
Who is Diogo Almeida, and why his résumé changes everything
This isn't yet another ex-OpenAI employee looking for a pretext to raise: he's one of the architects of the very paradigm he wants to replace. At OpenAI, Almeida co-invented RLHF — the technique that made ChatGPT usable — then worked on InstructGPT, ChatGPT, and GPT-4 (TechStartups, September 16, 2026).
His starting question, posted on X on September 15, 2026, deserves to be read exactly as written:
"After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?"
Translated: why have superhuman conversational models failed to produce AGI? His implicit answer stings — because conversation is probably not the right interface for intelligence. You can disagree with that conclusion; you cannot accuse the man of being unfamiliar with the subject.
Meanwhile, the frontier race continues on its side: OpenAI keeps stacking records with GPT-6 Astra on Terminal-Bench, science, and ARC-AGI-3. Two visions now openly coexist in the industry: pushing the limits of models for humans and agents, or building an entirely different category of models for software. My take? Both currents are necessary, but the second is structurally underfunded — hence the real significance of this seed round.
The thesis: intelligence should live in software, not in chat
For TypeSafe, the problem isn't the models' capability — it's their interface. A chatbot optimizes text production for a human; software, by contrast, needs typed, predictable, machine-evaluable decisions.
Almeida puts it bluntly: "most intelligence should eventually live inside software, running quietly in the background" (TechStartups, Sept. 2026).
Concretely, Jev doesn't hold conversations: it produces structured decisions, not conversational text (SiliconANGLE, Sept. 16, 2026). It can process hundreds of outputs in parallel from a single prompt, and every decision arrives with a calibrated confidence score. The software then knows what to do: act autonomously above a threshold, defer to a human below it.
To get there, TypeSafe isn't just fine-tuning an existing LLM: the company is announcing a new training method, RLCD (TechStartups, Sept. 2026). Technical details remain thin at this stage, but the intent is crystal clear: train a model for machine users from the ground up, rather than stripping the conversation out of a chatbot after the fact.
And this is no isolated case: the "one chat LLM for everything" consensus is cracking on several fronts at once. Subquadratic takes on quadratic attention with 12 million tokens of context, Alibaba arrives with the Qwen3.6 model family, and TypeSafe attacks on the usage front. Three different angles of attack, one and the same conclusion: the chatbot monolith is no longer untouchable.
Jev in practice: what a "for machines" model looks like
A model for machines looks like a component: you send an input, Jev returns a typed decision along with a probability, in under 100 ms as advertised. No paragraph to parse, no JSON to wrench out of a model that would rather chat.
Let's take a concrete example. A ticket comes in: "I haven't been able to log in since yesterday." Today, you call a conversational LLM, you wait for a natural-language response, you parse, you hope. With a machine-native approach, Jev directly returns something like facturation.expiration → confiance 0,94. Your router acts, or it doesn't.
The markets targeted by TypeSafe are deliberately prosaic, for that matter (SiliconANGLE, Sept. 2026):
- support request classification — route a ticket without paying for a full LLM call;
- invoice evaluation — extract a verdict and a confidence level;
- security alert triage — filter out the noise before human intervention;
- review of AI agent outputs — a programmatic guardrail on what your agents do.
That last point deserves a closer look. If you run agents — with the tools compared in our selection of LLMs for AI agents like OpenClaw or Hermes — you know that the weak link is validating their outputs. A specialized model that evaluates each agent action with calibrated confidence is exactly the missing piece in most of today's agentic pipelines.
An important nuance on latency: the 100 ms from the press releases is a marketing ceiling. TS2.tech reports measured latencies of 70 to 500 ms depending on the workloads tested (Sept. 2026). That's still one to two orders of magnitude below a typical LLM round trip — but measure on your own data before promising real time to your users.
$0.39 vs $19.49: what are Jev's quantified promises really worth?
On price, the advantage is public and verifiable. On performance, stay cautious: the spectacular figures are self-reported.
Let's start with the solid part. TypeSafe has published its pricing: $0.39 for 1,000 workflows, versus $3.31 for GPT-5.6 Luna and $19.49 for Claude Haiku 4.5 (SiliconANGLE, September 16, 2026). Even using the company's own four reference workflows, the gap ranges from 8.5x to 50x. And that takes no independent benchmark to verify: it's a price list, not a promise.
| Model | Cost for 1,000 workflows | Gap vs Jev |
|---|---|---|
| Jev (TypeSafe) | $0.39 | — |
| GPT-5.6 Luna (OpenAI) | $3.31 | 8.5x |
| Claude Haiku 4.5 (Anthropic) | $19.49 | 50x |
TS2.tech converts this pricing into $0.042 per million tokens, with no surcharge on output tokens (Sept. 2026). For a classification pipeline processing millions of them per day, the difference runs into thousands of euros per month — one million decisions would cost about $390 with Jev, $3,310 with Luna and $19,490 with Haiku 4.5, based on the published rates.
Now for the caveats. The "20-200x faster, 40-400x cheaper" claims and the internal tests showing ~194x and ~445x rest on four reference workflows… designed by TypeSafe's model-capabilities team, with no objective reference answer — the measurement is taken against the average of probabilities from two large external models (TS2.tech, Sept. 2026). The company itself acknowledges a possible bias and places its gains "at the top of the range." The "up to 100x" from the press releases? A marketing ceiling, not a median.
This debate lands right in the middle of the token war raging on the open-source side: tools like Caveman, RTK or Omniroute optimize token spending at the application layer; TypeSafe claims to divide it at the model layer. If the two approaches combine, the economics of LLM pipelines change scale.
One last telling detail: even between giants, the price gap is dizzying — $3.31 versus $19.49 for similar use cases. That's the whole market paradox dissected in our Claude vs ChatGPT comparison. In this context, a newcomer that publishes its prices to the tenth of a dollar at least forces transparency.
$40M in seed: what DCVC is really buying
DCVC isn't buying a startup that wraps the APIs of OpenAI, Anthropic, or Google: the fund is explicitly backing a model-layer thesis — what models must become when their users are machines (TechStartups, Sept. 2026). That's rare, and it's what makes this round interesting beyond the amount itself.
The amount, precisely: $40M in seed, i.e. the 99th percentile of all AI seed rounds in history across a sample of 28,473 deals (Dealroom, Sept. 2026). We're a long way from the standard $2M or $3M of early-stage funding.
James Hardiman, General Partner at DCVC, sums up the bet:
"TypeSafe is approaching one of the biggest remaining challenges in AI: turning increasingly capable models into technology that developers can reliably build into products at scale."
In other words: the real challenge is no longer model capability, but their reliability once integrated into products at scale. That's exactly the pain point of anyone who has ever put an LLM call into production.
The risk is real, and twofold. First, technical: proving that a non-conversational model covers enough use cases to justify a shift in primitives for developers. Then, contextual: with the White House now looking to verify AI models before their release, a model that self-declares its calibrated confidence will sooner or later have to submit to independent evaluations — in-house benchmarks won't be enough. My take: that's healthy. A startup selling reliability must accept being measured by others, not just itself.
Should You Rethink Your LLM Architecture Today?
No — but if your product relies on high-volume classification, routing, or triage, put Jev on your testing list, and keep a frontier model for everything else.
The architecture taking shape is hybrid. A frontier model — to be chosen from our monthly comparison of the best LLMs — for complex reasoning and generation; a machine-native model like Jev for high-volume repetitive decisions. It's exactly the same logic that applies to code: you don't entrust every task to the same model.
Two precautions before migrating anything:
Validate the calibration on your own data. SiliconANGLE explicitly reminds us: Jev's probability scores reduce the risk of hallucination, but don't eliminate it. It's up to you to verify that "confidence 0.97" actually means 97% accuracy on your use cases — that's evaluation work, not a checkbox.
Measure the real cost of your decisions. The right metric isn't the token price, it's the cost per 1,000 correct decisions, p99 latency included. On this front, a well-chosen local model can still win: our guide to the best LLMs to run locally and the installation tutorial with Ollama or LM Studio remain the reference for private processing at near-zero marginal cost. And if you want to prototype your own routing and monitoring layer without depending on anyone, a VPS from Hostinger is more than enough to get started.
Between open model families, post-attention architectures, and machine-native models, one thing is certain: choosing a model in 2026 is no longer a logo choice, it's an architecture choice.
❌ Common Mistakes
Mistake 1: Taking the 194x and 445x multiples at face value
These figures come from internal tests run on workflows designed by TypeSafe itself, with no objective reference baseline (TS2.tech, Sept. 2026). The fix: build your own mini-benchmark — 100 to 500 real samples from your workload, with cost and accuracy measured — before making any commitment.
Mistake 2: Confusing calibrated confidence with accuracy
A score of 0.97 means the model is 97% "confident" in its decision, not that it is correct 97 times out of 100. SiliconANGLE specifically warns about this confusion. The fix: plot the calibration curve (stated confidence vs observed accuracy) on your own data before letting the software act autonomously.
Mistake 3: Trying to migrate everything to a specialized model
Jev is built for high-volume structured decisions — classification, triage, evaluation. It won't write your emails or your code. The fix: keep a hybrid architecture, with a frontier LLM for generation and reasoning, and a machine-native model for volume.
Mistake 4: Confusing marketing latency with measured latency
The "under 100 ms" claims from press releases coexist with measurements of 70 to 500 ms depending on the workload (TS2.tech, Sept. 2026). Both can be true: it all depends on the use case. The fix: measure the p99 on your real requests, not on the vendor's demo.
❓ Frequently Asked Questions
What exactly is TypeSafe AI?
A startup founded in 2024 in San Francisco by Diogo Almeida (ex-OpenAI), Erik Gafni, and Sasha Sheng. It came out of stealth on September 16, 2026, with a $40M seed round led by DCVC. Its specialty: "composable machine-native intelligence" — models designed for machine users, the first of which is named Jev.
Who is Diogo Almeida?
An OpenAI veteran who co-invented RLHF and worked on InstructGPT, ChatGPT, and GPT-4. He left the lab with a question: why did superhuman conversational models not lead to AGI? His answer: we need models for machines, not just for humans.
Will Jev replace ChatGPT and conversational LLMs?
No, and that's not the goal. Jev doesn't hold a dialogue: it returns structured decisions meant to be consumed by software. Chat LLMs keep the human-facing side. On the other hand, for the high-volume LLM calls hidden inside your products, Jev claims to replace them at a fraction of the cost.
How much does Jev cost?
TypeSafe lists $0.39 per 1,000 workflows, versus $3.31 for GPT-5.6 Luna and $19.49 for Claude Haiku 4.5 (SiliconANGLE, September 2026). TS2.tech reads that as $0.042 per million tokens, with no output premium. Startup pricing being what it is: check the current pricing table before costing out a migration.
What is RLCD?
A new training method announced by Almeida at the same time as Jev (TechStartups, September 2026). Public details remain thin, and no verifiable paper has been published at this stage. The idea: train a model for machine users from the start, whereas RLHF optimizes responses for humans.
When will Jev be available?
No general availability date had been announced as of September 16, 2026. The publication of a pricing table suggests a commercial launch is close, but nothing is confirmed. The most reliable approach: watch TypeSafe's announcements and request access if your use case matches the target workloads.
✅ Conclusion
With $40M and the co-inventor of ChatGPT at its helm, TypeSafe AI has just bought the right to prove that a market for machine models exists — public, verifiable pricing, benchmarks still self-reported: the burden of proof now rests with Almeida, and that's exactly how the industry moves forward. Until independent evaluations come out, keep your hybrid pipelines and compare the best LLMs right now before migrating anything.