Cajal
A two-person startup using AI to formally verify math proofs and now binaries.
On the record
Checkpoint: Six month window for Cajal to publish a competitive IMO or Putnam level benchmark
The call
Cajal is a tiny, still repositioning entrant in a field where better funded rivals already have published benchmarks and shipped products.
Cajal builds software that uses artificial intelligence to verify that computer code works correctly and securely. The problem: the biggest companies it wants to sell to, like OpenAI and DeepMind, are smart enough to build these tools themselves rather than buy from Cajal, and competitors like Certora already have real customers and published results showing their tools work. The one thing that could save Cajal is landing an exclusive deal with a major client like the U.S. military (who has strict rules about code safety) or convincing a frontier lab to stick with buying rather than building, but its recent shift in focus from science tools to security tools suggests the company is still figuring out what it actually does best.
Why
01
AI pressure: OpenAI, DeepMind, Anthropic building in-house proof search. The frontier labs Cajal wants as customers have every incentive to build their own proof tools instead of buying. Could AI do the core job? Yes, likely.
02
Bigger players: Certora (AutoProver) and Harmonic (Aristotle). Certora shipped AutoProver in July 2026 and Harmonic already has IMO gold level results and a $1.45B valuation. Could a giant copy it? Yes, easily.
03
What protects it. Its open source Talos Wasm interpreter in Lean 4 could open a niche if paired with one regulated customer.
What would change this call
Cajal publishes an IMO or PutnamBench-class benchmark result within 6 months. Would show Cajal can match Harmonic's gold-medal result or Logical Intelligence's 99.4% score.
Cajal announces a named customer in a regulated field like aerospace or medical devices within 9 months. Would show real commercial traction beyond a two-person team.
Cajal repositions its product again after the Feb to June 2026 pivot from math AI to binary verification. Would show continued struggle to find a stable product-market fit.
Cajal announces a named customer in a regulated field like aerospace or medical devices within 9 months. Would show real commercial traction beyond a two-person team.
Cajal repositions its product again after the Feb to June 2026 pivot from math AI to binary verification. Would show continued struggle to find a stable product-market fit.
Key numbers
- Cajal team size: 2 people
- Harmonic Series C: $120M at $1.45B valuation
- Certora claims $75B in DeFi assets secured
Named risks
- No confirmed customers or revenue as of July 2026.
- Product pivot from AI mathematician to binary verification within 4 months signals unclear PMF.
- Core toolchain (Lean 4, Mathlib) is open-source and not proprietary per independent commentary.
- Target customer base (frontier AI labs) can build equivalent capability in-house.
- Conflicting funding reports ($500K note vs $5M claimed) unresolved by primary press.
Rivals, as read in July 2026
| Rival | Direction | Evidence |
|---|---|---|
| Harmonic AI | losing to | Aristotle hit IMO gold-equivalent Sept 2025, $1.45B valuation Nov 2025 vs Cajal's no published benchmark |
| Certora | losing to | Shipped AutoProver July 2026, claims $75B DeFi secured, directly competes with Cajal's binary-verification pivot |
| Axiom Math | gaining on | Reportedly raising ~$200M and hiring named mathematicians as of Apr 2026 (single-sourced) |
| Logical Intelligence | losing to | Aleph scored 99.4% on PutnamBench May 2026, no equivalent Cajal figure exists |
| OpenAI / DeepMind | contesting | OpenAI refuted Erdos unit-distance conjecture May 2026; DeepMind runs AlphaProof/AlphaEvolve in-house, target customers double as competitors |
Questions people ask
- Is Cajal profitable in 2026?
- No revenue or paying customers have been confirmed by any third-party source as of July 2026. Cajal is a two-person YC W26 startup (per YC's own listing, May 2026) with no disclosed pricing, revenue, or named clients, so profitability cannot be assessed.
- Could AI replace Cajal's product?
- Cajal's product Tau is itself an AI system built on open-source Lean 4 and Mathlib, tools independent commentary (May 2026) says have architecture that 'is not secret.' Frontier labs like OpenAI and DeepMind, its target customers, can build equivalent in-house proof-search systems.
- Who is Cajal's biggest competitor?
- Harmonic AI is the closest rival: its Aristotle system achieved gold-medal-equivalent IMO 2025 performance (published Sept 2025) and Harmonic closed a $120M Series C at a $1.45B valuation in November 2025, far outpacing Cajal's unpublished results.
- What does Cajal actually build?
- Cajal builds Tau, a multi-agent AI system generating Lean 4 formal proofs, originally pitched (Feb 2026 YC launch) as an AI mathematician for quantum computing and finance, later repositioned by mid-2026 toward verifying compiled binaries for security and code-correctness certificates.
- Does Cajal have any named customers?
- No source confirms a named paying customer. An apis.io listing dated July 21, 2026 states there is no third-party verification of paying customers, and Cajal's stated targets, frontier AI labs, are also its most capable potential competitors.
- Who owns Cajal?
- Cajal Technologies is privately held, founded in 2025, and backed by Y Combinator (Winter 2026 batch, partner Diana Hu). Funding reports conflict between a $500K convertible note led by Toloka.vc and a later-claimed $5M raised as of June 2026, unconfirmed by primary press.
Timeline
- 2026-07-21apis.io listing confirms no verified paying customers for Cajal
- 2026-07-14Certora ships AutoProver, agentic verification for Solidity/Rust, claims $75B DeFi secured
- 2026-06-15Cajal releases Talos, open-source Wasm interpreter in Lean 4 under AGPLv3
- 2026-06-10StartupHub.ai reports $5M total raised and notes pivot to binary verification framing
- 2026-05-09Logical Intelligence's Aleph reports 99.4% on PutnamBench
- 2026-05-07Independent commentary notes Cajal's Lean/Mathlib architecture 'is not secret'
- 2026-05-06CB Insights snapshot lists Cajal team size and notes $500K convertible note
- 2026-02-24Cajal launches on Y Combinator as 'AI mathematician' for scientific discovery
- 2025Cajal Technologies founded
Revision history
27 July 2026
First assessed. Losing its protection. The text above is fixed as of this date and is not edited afterwards.
In the same band
Sources
- IMO-level Automated Theorem Proving (2025-09-10)
- Mathematicians say AI cracked research proofs. The best tools aren't public (2026-04-14)
- Cajal: Scaling formal verification to accelerate scientific discovery
- Products, Competitors, Financials, Employees, Headquarters Locations (2026-05-06)
- Launch YC: Cajal: Scaling Formal Verification for Scientific Discovery (2026-02-24)
- cajal-technologies/talos: Wasm interpreter in lean, designed for reasoning
- Agentic Formal Verification for Every Developer (2026-07-14)
- A Rust-to-Lean Verification Pipeline with AI Provers (2026-05-09)
- Leiden Declaration on Artificial Intelligence and Mathematics (2026-07-27)
- Cajal Technologies, API Provider, Schemas (2026-07-21)
- Claude's Corner: Cajal, The Machine That Checks Its Own Math (2026-05-07)
- Cajal - $5M Raised - Reviews & Alternatives (2026-06-10)