Key points
This report (dated June 2026, AI-assisted literature review) examines whether Coq (renamed Rocq Prover in 2025, current version Rocq 9.2.0) and Isabelle/HOL remain the strongest formal reasoning tools, evaluating Lean 4 and other rivals.
Core conclusion
Coq and Isabelle remain unrivaled kings, but in complementary domains. Lean 4 is a rising challenger but poses no replacement threat yet. The future is a complementary tri-polar landscape, not a zero-sum contest.
Coq / Rocq — the industrial giant of dependent type theory
- Based on the Calculus of Inductive Constructions (CIC); supports propositions-as-types, dependent types, inductive definitions, and code extraction to OCaml/Haskell.
- Renamed Rocq Prover in 2025 with a jump to version 9.x; ACM Software System Award recipient; 40+ years of development led by INRIA.
- Landmark industrial milestones: CompCert (the only end-to-end formally verified optimizing C compiler), VST 3.0 (separation logic for C verification, Iris-based), Iris concurrency logic, Four Color Theorem (2005), Feit-Thompson theorem (2012), MetaRocq self-bootstrap verification.
- Industrial users include Google, AbsInt, BlueRock, Formal Vindication. LLM4Rocq explores LLM integration.
- Built on classical higher-order logic with the human-readable Isar proof language and Sledgehammer automation (delegates to E, Vampire, Z3, CVC4).
- Landmarks: seL4 microkernel (first full OS kernel verification, 2009), AWS Nitro Isolation Engine (2026 — the first formally verified hypervisor deployed in a commercial cloud, ~330,000 lines of proofs, written in μRust, running in production on Graviton5), Isabelle/Solidity (2025), AutoCorrode Rust verification (2025).
- Archive of Formal Proofs (AFP): 982 entries, ~5.2 million lines, ~316,900 lemmas, 593 authors; 640 PutnamBench formalizations.
- Dependent type theory; self-implemented in Lean 4; created by Leonardo de Moura; best-in-class VS Code experience; metaprogramming and proofs share one language.
- mathlib4: >5,000,000 lines (March 2026), 132,448 definitions, 278,346 theorems, 772 contributors; leads PutnamBench with 672 formalizations (vs Isabelle 640, Coq 412). Endorsed by Terence Tao.
- Weaknesses: zero industrial verification milestones (no CompCert/seL4/AWS Nitro equivalent), C++ runtime kernel with less long-term scrutiny, automation below Sledgehammer level, younger ecosystem.
- Agda: deep dependent types, beloved by PL researchers, weak automation.
- F*: refinement types + SMT, Project Everest (TLS), specialized for program verification.
- HOL Light / HOL4: minimal trusted computing base, Flyspeck floating-point work, small ecosystem.
Isabelle/HOL — the gold standard for system verification
Lean 4 — the strongest challenger
Other competitors
Domain-to-tool mapping
| Domain | Best tool | |---|---| | C compiler verification | Coq (CompCert) | | OS kernel verification | Isabelle (seL4) | | Cloud infrastructure isolation | Isabelle (AWS Nitro) | | Concurrent program verification | Coq (Iris) | | Large-scale math formalization | Lean 4 / Isabelle | | Teaching / onboarding | Lean 4 | | Minimal TCB | HOL Light | | AI + formal proofs | Lean 4 (Pantograph) / Coq (LLM4Rocq) |
Discussion: are Coq and Isabelle still the kings?
Yes, for now. Industrial verification cannot be rushed: CompCert took ~20 years, seL4 verification spanned 2004–2009 and beyond to AWS Nitro (2026). The two tools embody fundamentally different philosophies—Coq's constructive CIC enables proof-to-program extraction but restricts classical reasoning; Isabelle's classical HOL enables Sledgehammer's powerful automation but sacrifices extraction. Lean 4 takes a middle path.
Warnings: mathlib4's explosive growth and 772 contributors could gradually marginalize Coq/Isabelle in mathematical formalization; Lean 4's developer experience represents a generational gap over CoqIDE and Isabelle/jEdit; and LLM integration (Pantograph vs LLM4Rocq) may reshape the landscape within five years.