DeepSeek V4 Pro (Workers AI) vs Gemma 4 26B A4B (Workers AI)
Reading a 26.2K-token pull request costs $0.0346 on DeepSeek V4 Pro (Workers AI), 13× the $0.0026 it costs on Gemma 4 26B A4B (Workers AI). DeepSeek V4 Pro (Workers AI) is a mid-range model and Gemma 4 26B A4B (Workers AI) a economy one. In the benchmark DeepSeek V4 Pro (Workers AI) caught 28 percentage points more of the known defects than Gemma 4 26B A4B (Workers AI) (95% interval +14 pts to +42 pts, a − b).
Side by side
| Figure | DeepSeek V4 Pro (Workers AI) | Gemma 4 26B A4B (Workers AI) |
|---|---|---|
| Price tier | Mid-range | Economy |
| Bugs caughtshare of known bugs | 54% (better) | 26% |
| False alarmsper clean PR | 2.9 | 0.4 (better) |
| Cost per reviewmeasured | $0.0643 | $0.0048 (better) |
| Cost per bug caught | $0.0872 | $0.0209 (better) |
| Exact linecatches on the expected lines | 89% | 91% (better) |
| Review timemean | 390s | 308s (better) |
| Typical review calllist price | $0.0442 | $0.0033 (better) |
| Input / 1M tokens | $1.32 | $0.10 (better) |
| Output / 1M tokens | $3.96 | $0.30 (better) |
| Context window | 1M | 256K |
Head to head
The known bugs both were scored on, by who caught each in at least one run.
By language
| Model | All | TypeScript6 bugs | Go4 bugs | Java4 bugs | Rust4 bugs | C#3 bugs | JavaScript3 bugs | Kotlin3 bugs | PHP3 bugs | Python3 bugs | Ruby3 bugs | Swift3 bugs | C++2 bugs | C1 bug |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| DeepSeek V4 Pro (Workers AI) | 54% | 79% | 25% | 49% | 54% | 59% | 17% | 30% | 69% | 72% | 67% | 75% | 38% | 33% |
| Gemma 4 26B A4B (Workers AI) | 26% | 33% | 0% | 30% | 38% | 7% | 0% | 56% | 33% | 44% | 67% | 0% | 0% | 0% |
Of the 42 known bugs both were scored on, both caught 15 and neither caught 7. DeepSeek V4 Pro (Workers AI) caught 19 bugs that Gemma 4 26B A4B (Workers AI) missed: 3 in Swift, 2 in C++, 2 in Java, 2 in PHP, 2 in Rust, 2 in TypeScript, 1 in C, 1 in C#, 1 in Go, 1 in JavaScript, 1 in Python and 1 in Ruby. Gemma 4 26B A4B (Workers AI) caught one bug that DeepSeek V4 Pro (Workers AI) missed: 1 in Rust.
On pull requests with no known bug, DeepSeek V4 Pro (Workers AI) raised 2.9 false alarms per review and Gemma 4 26B A4B (Workers AI) 0.4. A measured review cost $0.0643 on DeepSeek V4 Pro (Workers AI), 14× the cost of $0.0048 on Gemma 4 26B A4B (Workers AI).
| Pull request | Language · kind | DeepSeek V4 Pro (Workers AI) | Gemma 4 26B A4B (Workers AI) |
|---|---|---|---|
| OrchardCMS/OrchardCoreSimplify using `StringValues` in Liquid templates | C# · Security | 3/3 | 0/3 |
| aio-libs/aiohttpkeep websocket compression state across interleaved control frames | Python · Logic | 4/4 | 0/3 |
| better-auth/better-authfix(saml): IdP-Initiated Callback Routing | TypeScript · Security | 4/4 | 0/2 |
| pointfreeco/swift-navigationUIKit: Fix switch presented child double dismiss | Swift · Logic | 3/3 | 0/4 |
| traefik/traefikAdd the option to define custom cipher suites for backend serversTransport | Go · API misuse | 4/4 | 0/3 |
| TanStack/queryfix(query-core): fix hydration bugs for already resolved promises | TypeScript · Logic | 3/4 | 0/2 |
| doctrine/ormFix collection filtering API for IN/NOT IN comparisons that require type conversions | PHP · Logic | 3/4 | 0/3 |
| groue/GRDB.swiftDatabaseQueue and DatabaseSnapshot reads can access task locals | Swift · Concurrency | 3/4 | 0/2 |
| nushell/nushellreorganize nushell configuration | Rust · Error handling | 3/4 | 0/2 |
| duckdb/duckdbFix vsize = 2 issue in JSON create | C++ · Concurrency | 2/4 | 0/3 |
| jhy/jsoupStore cached child els in Attributes, not field | Java · Concurrency | 3/6 | 0/8 |
| nodejs/undicifeat: add configurable maxPayloadSize for WebSocket | JavaScript · Security | 2/4 | 0/3 |
| pointfreeco/swift-composable-architectureRemove child store from parent when invalidated | Swift · Resource handling | 2/4 | 0/4 |
| socketry/asyncBarrier wait in completion order + `Queue#close`. | Ruby · Concurrency | 2/4 | 0/3 |
| tokio-rs/tokiotime: move lazy-registration state into `Sleep` | Rust · Logic | 2/5 | 0/2 |
| c-ares/c-aresEvent Thread could stall when not notified of new queries on existing connections that are in a bad state | C · Resource handling | 1/3 | 0/4 |
| symfony/symfony[Security] Allow multiple OIDC discovery endpoints | PHP · Security | 1/3 | 0/3 |
| ggml-org/llama.cpprpc : cache and reuse compute graphs | C++ · Security | 1/4 | 0/3 |
| mockito/mockitoAdd core API to enable Kotlin singleton mocking | Java · API misuse | 1/5 | 0/3 |
| denoland/denoperf(libs/core): poll only ready handles and yield between I/O batches | Rust · Concurrency | 0/4 | 2/2 |
Defects caught over all runs: a PR reviewed three times with one known bug shows 3/3 when it was caught every time.
Every bug each was scored on: DeepSeek V4 Pro (Workers AI) and Gemma 4 26B A4B (Workers AI).