AlphaProof Code vs SWE-2: Two Agentic Systems, Two Completely Different Ideas of Proof

Both AlphaProof Code and Cognition SWE-2 are agent systems that write software automatically, and both were sold to the public on the word "agent." Beyond that the two releases barely inhabit the same universe. Cognition SWE-2, announced in September 2026 and available only inside Devin Desktop and the CLI, writes patches: it scored 50.0% on FrontierCode 1.1 Main, within a point of Claude Fable 5.1 at 50.9%, for 64% less money, according to Cognition's own numbers. AlphaProof Code, announced by G​oogle DeepMind on 10 November 2026, writes proofs: roughly 6.5 million lines of Lean establishing that Signal's compiled AArch64 binary satisfies ten security theorems, checked by Lean's kernel rather than by a rubric. One system's output can be wrong in a way that a human reviewer might catch. The other's output cannot be wrong at all — or the kernel rejects it and the work continues.

"Correct" means two different things

Start with the word, because it is where comparisons between these two go wrong. When Cognition reports that SWE-2 resolved 50.0% of FrontierCode 1.1 Main tasks, "resolved" means a patch was produced that the grading process accepted as mergeable. FrontierCode's tasks were written with more than 20 open-source maintainers, at more than 40 hours per task, each maintainer defining what "mergeable" means in their own repository. That is a serious evaluation design and it is still a human judgment about a diff.

When DeepMind reports that a theorem was proved for libsignal v0.105.4, "proved" means Lean's kernel accepted a derivation from the specification. There is no reviewer with an opinion. The gatekeeper agent checks each submission with comparator against a held-out copy of the specification, and Lean checks the rest — which is how a team of five researchers could leave 99% of a codebase unreviewed by humans and still make a claim about all of it.

That difference determines what each system is good for. SWE-2 can write code. AlphaProof Code cannot write production code and does not try; it demonstrates that code someone else wrote does what was specified, and it can also demonstrate that a deliberately broken version does not — the two reintroduced Signal bugs, a weak-state bug and a small-subgroup bug, both produced machine-checked proofs of insecurity.

Where the two are genuinely alike

More than you would expect, and the parallels are the interesting part. Both are expensive to run. AlphaProof Code used teams of 40 to 80 agents running 24 to 72 hours on a single machine per team, across four weeks, with a supervising agent above the team orchestrator and agents auditing their own output to flag roughly 1% for human review. SWE-2's cost is not published per token because Cognition does not sell tokens; the price is a Devin subscription, and the 64%-cheaper claim is a comparison against Claude Fable 5.1 at list rates.

Both rely on the same class of base model. SWE-2 is a post-train of M​oonshot AI's Kimi K3, a 2.8-trillion-parameter mixture-of-experts with roughly 104B parameters active per token, and Cognition says its reinforcement learning added 5 to 6 points across several benchmarks. AlphaProof Code's proof agents were powered by G​emini 4 Carbon, and Claude Opus 5.5 was among the models used to draft and review Signal's Lean specifications. Neither lab built its agents from scratch.

Both depend on a verifier to know whether the work is done. Cognition's verifiers are the test suites and maintainer judgments encoded in FrontierCode. DeepMind's verifier is a proof kernel. The verifier is where each system's ceiling is set, and that is the sentence worth remembering from either paper.

The benchmark honesty problem, side by side

AlphaProof Code has no benchmark table, so the usual caveats about vendor numbers do not apply to it — it has a theorem instead, and a theorem is binary. What it does have is a set of structural exclusions. No post-quantum guarantee for SPQR's healing property, because the ratchet theorems assume the attacker chooses SPQR keys. No payload privacy. No responder authentication for the first message of a handshake that is never promoted. One session per promotion, fixed message lengths, no multi-device state, trust-on-first-use identity keys. An 8-byte MAC forgery term that grows quadratically and goes vacuous at roughly 25,000 messages in a single session. Thirty-seven routines contracted rather than verified, including AES-256-CBC, X25519, Kyber1024 and ML-KEM-768. And the subject is the team's build of libsignal-cli as a fresh process, not the Signal app.

SWE-2's caveats are statistical rather than structural, and the sharpest one is scalability across task length. Going from Terminal-Bench 2.1 to Terminal-Bench 4, Claude Fable 5.1 loses about 35.6 points and GPT-6 Astra about 32.0. SWE-2 loses about 65.5, and its Kimi K3 base about 66.8 — so on long-horizon work it degrades roughly twice as fast as the frontier models it is compared against. Terminal-Bench 4 is the current edition and 2.1 is the superseded one, which means the row where SWE-2 leads is the retired benchmark and the row where it trails is the live one.

Both reports are unusually candid by launch-post standards. The difference is what candour costs. DeepMind's exclusions narrow the claim but leave it true. Cognition's exclusions bound the claim's applicability, and the reader has to decide whether a 27.3% score on the current long-horizon benchmark matters for their workload.

Serving, price, and the thing neither one sells you

• AlphaProof Code — research framework. No endpoint, no weights, no price, no waitlist. The artifacts are a Lean development at github.com/google-deepmind/code-verification with no licence stated on the landing page.

• Cognition SWE-2 — proprietary, Devin-only. No published per-token rate card, no model ID to call, no weights. If you want SWE-2 you buy Devin.

• The base models — Kimi K3 is on OrcaRouter's catalogue at $3 per million input and $15 per million output tokens, provider list price passed through at 0% markup. Claude Opus 5.5, one of the models used to draft Signal's specifications, is at $4 and $20. G​emini 4 Carbon, which powered AlphaProof Code's proof agents, is not something we serve, and nothing here suggests we host AlphaProof Code itself.

That third line is the only part of this comparison with a bill attached that you can pay today. If you are building an agentic coding workflow, the practical choice is not between these two systems — it is between the models underneath them, and both Kimi K3 and Claude Opus 5.5 are one key away with automatic failover if a provider has a bad afternoon. A routing DSL lets you send the same task to two of them and compare the diffs before either touches a repository.

Which failure mode you can live with

SWE-2 fails by producing a patch that does not compile, or compiles and is wrong, or is right and unmergeable. Every one of those failures is discoverable by reading the diff or running the tests, and every one of them costs you a review cycle. The system is designed around this: Cognition reports best-of-N across reasoning-effort settings, which is what you do when the output is probabilistic.

AlphaProof Code fails by not producing a proof, at which point the agents keep working — for hours, sometimes days, with a clean build that can take several hours and proof repairs triggered by a one-line source change. The paper says there is "substantial room for optimization" in both the generated proofs and the techniques, which is a polite way of saying the process is slow and brittle. But when it succeeds there is no review cycle at all, because there is nothing to review. Lean has already agreed.

For a team shipping a feature on Thursday, that trade is obvious and SWE-2 wins it. For a team shipping a cryptographic protocol that will protect conversations for a decade, SWE-2's entire category is the wrong tool, and the fact that it scored 92.8% on Terminal-Bench 2.1 is irrelevant.

What would make these actually comparable

One thing: a benchmark that measures whether an agent's output is right rather than whether it is accepted. FrontierCode approximates this with maintainer-defined mergeability, which is as close as code generation has come. AlphaProof Code sidesteps it entirely by choosing a domain where a kernel can adjudicate. If someone built a suite of tasks where the correct answer is mechanically checkable against a specification — not a test suite, a specification — the two systems could be scored on one axis and the comparison would mean something.

Until then, the useful read is narrower. AlphaProof Code tells you that a five-person team with agent tooling can produce a proof corpus that used to take years, which is a statement about what verification costs. SWE-2 tells you that a post-trained open-weight base can sit within a point of the frontier on a coding benchmark at 64% less, which is a statement about what coding costs. Both costs are falling. They are falling at different rates, for different reasons, and a leaderboard that puts them in the same table would be lying to you.

If you want the practical takeaway: use SWE-2-shaped systems for code that a test suite can gate, and start reading the AlphaProof Code artifact repository if you have code that a test suite cannot gate. The second category is smaller than most teams think and more consequential than any of them admit.