GPT-OSS 120B (Workers AI) vs GPT-6.1 Sol
Reading a 26.2K-token pull request costs $0.0524 on GPT-6.1 Sol, 5.7× the $0.0092 it costs on GPT-OSS 120B (Workers AI). GPT-OSS 120B (Workers AI) is a economy model and GPT-6.1 Sol a mid-range one. In the benchmark GPT-6.1 Sol caught 49 percentage points more of the known defects than GPT-OSS 120B (Workers AI) (95% interval −62 pts to −35 pts, a − b).
Side by side
| Figure | GPT-OSS 120B (Workers AI) | GPT-6.1 Sol |
|---|---|---|
| Price tier | Economy | Mid-range |
| Bugs caughtshare of known bugs | 40% | 89% (better) |
| False alarmsper clean PR | 2.5 | 0.9 (better) |
| Cost per reviewmeasured | $0.0075 (better) | $0.0401 |
| Cost per bug caught | $0.0189 (better) | $0.0514 |
| Exact linecatches on the expected lines | 49% | 95% (better) |
| Review timemean | 55s | 28s (better) |
| Typical review calllist price | $0.011 (better) | $0.0766 |
| Input / 1M tokens | $0.35 (better) | $2.00 |
| Output / 1M tokens | $0.75 (better) | $10.00 |
| Context window | 128K | 1.1M |
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 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| GPT-OSS 120B (Workers AI) | 40% | 42% | 13% | 46% | 71% | 41% | 11% | 44% | 33% | 78% | 33% | 44% | 33% | 0% |
| GPT-6.1 Sol | 89% | 83% | 75% | 70% | 100% | 89% | 100% | 100% | 100% | 100% | 100% | 100% | 50% | 100% |
Of the 42 known bugs both were scored on, both caught 26 and neither caught 1. GPT-OSS 120B (Workers AI) caught 3 bugs that GPT-6.1 Sol missed: 1 in Go, 1 in Java and 1 in TypeScript, most of them concurrency bugs. GPT-6.1 Sol caught 12 bugs that GPT-OSS 120B (Workers AI) missed: 2 in Go, 2 in JavaScript, 1 in C, 1 in C#, 1 in Kotlin, 1 in PHP, 1 in Ruby, 1 in Rust, 1 in Swift and 1 in TypeScript.
On pull requests with no known bug, GPT-OSS 120B (Workers AI) raised 2.5 false alarms per review and GPT-6.1 Sol 0.9. A measured review cost $0.0075 on GPT-OSS 120B (Workers AI), 81% less than $0.0401 on GPT-6.1 Sol.
| Pull request | Language · kind | GPT-OSS 120B (Workers AI) | GPT-6.1 Sol |
|---|---|---|---|
| jhy/jsoupStore cached child els in Attributes, not field | Java · Concurrency | 3/6 | 0/2 |
| go-gitea/giteaperf(gitdiff): optimize inline diff highlighting using cache | Go · Error handling | 1/3 | 0/1 |
| trpc/trpcfix(server): pass maxDurationMs timeout signal to subscription handlers | TypeScript · Concurrency | 1/3 | 0/1 |
| 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 | 0/3 | 1/1 |
| colinhacks/zodfeat(v4): parse input containing reference cycles | TypeScript · Resource handling | 0/6 | 1/1 |
| dotnet/orleansfix(serialization): cap codec allocations | C# · Logic | 0/3 | 1/1 |
| go-chi/chifeat: middleware.ClientIP, a replacement for middleware.RealIP | Go · Security | 0/3 | 1/1 |
| groue/GRDB.swiftDatabaseQueue and DatabaseSnapshot reads can access task locals | Swift · Concurrency | 0/3 | 1/1 |
| kotest/kotestFix DataClassEq comparator | Kotlin · Logic | 0/3 | 1/1 |
| nodejs/undicifeat: add configurable maxPayloadSize for WebSocket | JavaScript · Security | 0/3 | 1/1 |
| nushell/nushellreorganize nushell configuration | Rust · Error handling | 0/6 | 1/1 |
| rails/railsReimplement Action Cable redis adapter with redis-client | Ruby · Logic | 0/3 | 1/1 |
| sindresorhus/p-throttleAdd `weight` option | JavaScript · Logic | 0/3 | 1/1 |
| symfony/symfony[Security] Allow multiple OIDC discovery endpoints | PHP · Security | 0/3 | 1/1 |
| traefik/traefikAdd the option to define custom cipher suites for backend serversTransport | Go · API misuse | 0/6 | 1/1 |
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: GPT-OSS 120B (Workers AI) and GPT-6.1 Sol.