Jev vs Claude: $0.042 Per Million Tokens (Sept 2026)

ChatGPT co-inventor Diogo Almeida just launched Jev, a fundamentally different kind of frontier AI. It's not a chatbot. It doesn't generate text. And it costs 40–400x less than Claude or GPT-6 Astra while running 20–200x faster.

For founders building D2C, SaaS, or high-frequency decision systems, this is worth understanding. Not because Jev replaces Claude. But because it reveals something about how AI actually moves from research to product: text generation was step one. Decision-making is step two.

What Changed: System One vs System Two

Claude, GPT-6 Astra, and Fable 5.1 are all designed the same way: they take text in, generate text out, one token at a time. This is called System Two thinking—deliberate, slow, thorough. It's why Claude can write essays, debug code, and reason through complex problems.

Jev is the opposite. It's a System One model, built on a training method Almeida calls RLCD (Reinforcement Learning for Calibrated Decisions). It takes messy state in—an email, a log line, a support ticket, a JSON blob—and returns a typed probabilistic decision out.

Not: "Here's my analysis of your support ticket."

Instead: "probability=0.94, confidence=0.87" for the decision you asked for.

TypeSafe trained Jev over two years in stealth, funded by $40 million led by DCVC. The company launched today (September 15, 2026) with early access via waitlist.

Real Specs: Speed and Cost

Jev's core specs are aggressive:

Speed: - Response time: 70–500 milliseconds (p50: ~150ms, p95: ~350ms) - 20–200x faster than Claude, GPT-6 Astra, or Fable 5.1 - Latency matters because if AI returns in 150ms, developers can call it inside the software loop instead of after-the-fact Cost: - Input: $0.042 per million tokens - Output: FREE - 40–400x cheaper than frontier models (Claude Opus runs $5/$25 per million) - Early tester ran 5,000 requests for roughly $2 Accuracy: - No independent benchmark yet (no paper published) - TypeSafe's own testing shows calibration (confidence scores match actual accuracy) - Launches with closed weights on US-hosted API

Use Cases: Where Jev Actually Wins

Jev excels at decision tasks, not reasoning tasks. For founders, this means:

Routing and classification: - Customer support ticket routing (which team? which tier?) - Intent detection (is this refund request or feature request?) - Content moderation decisions (is this review fake or legitimate?) Real-time decisions inside software: - Fraud detection (is this transaction risky? Score + probability) - Product recommendations (which variant should this user see?) - Workflow steering (should we escalate or resolve?) Chaining with LLMs: - Run Jev first to route/classify (cheap, fast, 150ms) - Call Claude only if needed (expensive, slow, necessary for reasoning) - Run Jev again after Claude to tag/score output (safety check) The economic insight: if you're currently calling Claude for every decision, you're paying $5 per million input tokens. If you can do 90% of decisions with Jev first ($0.042 per million), you've just dropped costs by 100x on that 90%.

Cost Breakdown: Jev vs Claude vs GPT-6 Astra

The math: If you route 1 million support tickets per month, Jev costs $42. Claude costs $5,000. Running Jev first, then Claude for complex cases, drops your total cost from $5,000 to ~$1,542.

When Jev Wins vs When Claude Wins

Jev wins when: - You have a pre-defined schema (classification, yes/no, scoring, selection) - Latency is critical (<500ms preferred) - You're making the same decision thousands of times - You can't afford to call Claude for every decision - You need confidence scores to know when to escalate to humans Claude wins when: - You need free-form reasoning (why this decision, not just what) - You need explanation or justification (regulatory, customer-facing) - The decision is genuinely novel (not pattern-matching) - You need error handling outside a schema - You can afford 2–5 second latency Honest tension: Jev can't generate an answer outside the schema you defined. It can only choose the wrong option inside it. Claude can hallucinate entirely new answers. For high-stakes decisions (financial, legal, medical), this schema constraint is actually a feature. For creative work, it's a liability.

The Architecture Question: Why TypeSafe Built This

Diogo's founding question: "Why have superhuman chat models not led to AGI?"

His answer isn't a bigger chat model. It's recognizing that AGI-adjacent progress doesn't come from better text generation. It comes from AI that can sit inside decision loops, at 150ms latency, and cost $0.042 per million tokens.

Frontier models (Claude, GPT-6, Fable 5.1) optimized for reasoning. Jev optimized for throughput and latency. These are different optimization targets.

The implication: the next wave of AI economics isn't "replace humans with AI." It's "embed AI into every decision, call it thousands of times per day, and make it cost-invisible."

When NOT to Use Jev

Don't use Jev if: - You're doing reasoning-heavy work (complex analysis, debugging, planning) - You need explanation for regulatory compliance - Your decision schema isn't pre-defined - You can afford 2–5 second latency - You're evaluating something genuinely novel to the model Jev is a specialized tool. Useful for production workflows. Not a replacement for Claude.

Practical Decision: Jev vs Claude for Your Stack

If you're a founder handling 10K+ decisions/month in your product: 1. Write your decision schema (is this spam? Low/Medium/High risk?) 2. Test Jev on 1,000 examples (~$0.05) 3. If accuracy >90%, use Jev for all decisions 4. If accuracy <90%, use Jev + Claude (Jev routes, Claude decides) Cost impact: - 10K decisions/month with Claude: $50 - Same with Jev: $0.50 - Jev + Claude hybrid (90/10 split): $5

Want to understand how AI models move from research to product? This is the moment. Jev represents a shift from optimizing for reasoning to optimizing for scale, cost, and latency. For founders, that shift changes how you architect every decision layer in your product.