Claude Haiku 5.5 vs Gemma 4 26B A4B (Workers AI)
Claude Haiku 5.5 and Gemma 4 26B A4B (Workers AI) charge practically the same to read a 26.2K-token pull request ($0.0026). Above 100K tokens long-context rates apply, and Gemma 4 26B A4B (Workers AI) becomes the cheaper one. Both sit in the economy price tier, so the choice between them is about review quality more than budget. In the benchmark Claude Haiku 5.5 caught 38 percentage points more of the known defects than Gemma 4 26B A4B (Workers AI) (95% interval +22 pts to +53 pts, a − b).
Side by side
| Figure | Claude Haiku 5.5 | Gemma 4 26B A4B (Workers AI) |
|---|---|---|
| Price tier | Economy | Economy |
| Bugs caughtshare of known bugs | 64% (better) | 26% |
| False alarmsper clean PR | 1.4 | 0.4 (better) |
| Cost per reviewmeasured | $0.0038 (better) | $0.0048 |
| Cost per bug caught | $0.0064 (better) | $0.0209 |
| Exact linecatches on the expected lines | 82% | 91% (better) |
| Review timemean | 15s (better) | 308s |
| Typical review calllist price | $0.0038 | $0.0033 (better) |
| Input / 1M tokens | $0.10 | $0.10 |
| Output / 1M tokens | $0.50 | $0.30 (better) |
| Context window | 1M | 256K |
- Above 100K tokens, long-context rates mean Gemma 4 26B A4B (Workers AI) is the cheaper of the two.
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 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Claude Haiku 5.5 | 64% | 83% | 75% | 40% | 75% | 67% | 33% | 67% | 100% | 39% | 67% | 67% | 50% | 0% |
| 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 14 and neither caught 12. Claude Haiku 5.5 caught 14 bugs that Gemma 4 26B A4B (Workers AI) missed: 3 in Go, 2 in PHP, 2 in Swift, 2 in TypeScript, 1 in C++, 1 in C#, 1 in JavaScript, 1 in Ruby and 1 in Rust. Gemma 4 26B A4B (Workers AI) caught 2 bugs that Claude Haiku 5.5 missed: 1 in Kotlin and 1 in Ruby, all of them logic bugs.
On pull requests with no known bug, Claude Haiku 5.5 raised 1.4 false alarms per review and Gemma 4 26B A4B (Workers AI) 0.4. A measured review cost $0.0038 on Claude Haiku 5.5, 20% less than $0.0048 on Gemma 4 26B A4B (Workers AI).
| Pull request | Language · kind | Claude Haiku 5.5 | Gemma 4 26B A4B (Workers AI) |
|---|---|---|---|
| OrchardCMS/OrchardCoreSimplify using `StringValues` in Liquid templates | C# · Security | 1/1 | 0/3 |
| better-auth/better-authfix(saml): IdP-Initiated Callback Routing | TypeScript · Security | 1/1 | 0/2 |
| colinhacks/zodfeat(v4): parse input containing reference cycles | TypeScript · Resource handling | 1/1 | 0/2 |
| doctrine/ormFix collection filtering API for IN/NOT IN comparisons that require type conversions | PHP · Logic | 1/1 | 0/3 |
| duckdb/duckdbFix vsize = 2 issue in JSON create | C++ · Concurrency | 1/1 | 0/3 |
| go-gitea/giteaperf(gitdiff): optimize inline diff highlighting using cache | Go · Error handling | 1/1 | 0/3 |
| golang/gocrypto/tls: add LocalCertificate to ConnectionState | Go · Logic | 1/1 | 0/3 |
| groue/GRDB.swiftDatabaseQueue and DatabaseSnapshot reads can access task locals | Swift · Concurrency | 1/1 | 0/2 |
| nodejs/undicifeat: add configurable maxPayloadSize for WebSocket | JavaScript · Security | 1/1 | 0/3 |
| pointfreeco/swift-composable-architectureRemove child store from parent when invalidated | Swift · Resource handling | 1/1 | 0/4 |
| socketry/asyncBarrier wait in completion order + `Queue#close`. | Ruby · Concurrency | 1/1 | 0/3 |
| symfony/symfony[Security] Allow multiple OIDC discovery endpoints | PHP · Security | 1/1 | 0/3 |
| tokio-rs/tokiotime: move lazy-registration state into `Sleep` | Rust · Logic | 1/1 | 0/2 |
| traefik/traefikAdd the option to define custom cipher suites for backend serversTransport | Go · API misuse | 1/1 | 0/3 |
| kotest/kotestFix DataClassEq comparator | Kotlin · Logic | 0/1 | 1/3 |
| rails/railsReimplement Action Cable redis adapter with redis-client | Ruby · Logic | 0/1 | 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: Claude Haiku 5.5 and Gemma 4 26B A4B (Workers AI).