GPT-6 Luna vs GPT-OSS 120B (Workers AI)
Reading a 26.2K-token pull request costs $0.0092 on GPT-OSS 120B (Workers AI), 3.5× the $0.0026 it costs on GPT-6 Luna. Both sit in the economy price tier, so the choice between them is about review quality more than budget. In the benchmark GPT-6 Luna caught 27 percentage points more of the known defects than GPT-OSS 120B (Workers AI) (95% interval +13 pts to +41 pts, a − b).
Side by side
| Figure | GPT-6 Luna | GPT-OSS 120B (Workers AI) |
|---|---|---|
| Price tier | Economy | Economy |
| Bugs caughtshare of known bugs | 67% (better) | 40% |
| False alarmsper clean PR | 0.7 (better) | 2.5 |
| Cost per reviewmeasured | $0.0028 (better) | $0.0075 |
| Cost per bug caught | $0.0046 (better) | $0.0189 |
| Exact linecatches on the expected lines | 97% (better) | 49% |
| Review timemean | 29s (better) | 55s |
| Typical review calllist price | $0.0038 (better) | $0.011 |
| Input / 1M tokens | $0.10 (better) | $0.35 |
| Output / 1M tokens | $0.50 (better) | $0.75 |
| Context window | 1.1M | 128K |
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-6 Luna | 67% | 83% | 25% | 65% | 75% | 56% | 67% | 67% | 67% | 100% | 67% | 100% | 50% | 0% |
| GPT-OSS 120B (Workers AI) | 40% | 42% | 13% | 46% | 71% | 41% | 11% | 44% | 33% | 78% | 33% | 44% | 33% | 0% |
Of the 42 known bugs both were scored on, both caught 24 and neither caught 8. GPT-6 Luna caught 5 bugs that GPT-OSS 120B (Workers AI) missed: 1 in Go, 1 in JavaScript, 1 in Ruby, 1 in Swift and 1 in TypeScript. GPT-OSS 120B (Workers AI) caught 5 bugs that GPT-6 Luna missed: 2 in Go, 1 in Java, 1 in Ruby and 1 in TypeScript.
On pull requests with no known bug, GPT-6 Luna raised 0.7 false alarms per review and GPT-OSS 120B (Workers AI) 2.5. A measured review cost $0.0028 on GPT-6 Luna, 63% less than $0.0075 on GPT-OSS 120B (Workers AI).
| Pull request | Language · kind | GPT-6 Luna | GPT-OSS 120B (Workers AI) |
|---|---|---|---|
| colinhacks/zodfeat(v4): parse input containing reference cycles | TypeScript · Resource handling | 1/1 | 0/6 |
| go-chi/chifeat: middleware.ClientIP, a replacement for middleware.RealIP | Go · Security | 1/1 | 0/3 |
| groue/GRDB.swiftDatabaseQueue and DatabaseSnapshot reads can access task locals | Swift · Concurrency | 1/1 | 0/3 |
| nodejs/undicifeat: add configurable maxPayloadSize for WebSocket | JavaScript · Security | 1/1 | 0/3 |
| rails/railsReimplement Action Cable redis adapter with redis-client | Ruby · Logic | 1/1 | 0/3 |
| go-gitea/giteaperf(gitdiff): optimize inline diff highlighting using cache | Go · Error handling | 0/1 | 1/3 |
| golang/gocrypto/tls: add LocalCertificate to ConnectionState | Go · Logic | 0/1 | 1/6 |
| jhy/jsoupStore cached child els in Attributes, not field | Java · Concurrency | 0/2 | 3/6 |
| socketry/asyncBarrier wait in completion order + `Queue#close`. | Ruby · Concurrency | 0/1 | 1/3 |
| withastro/astroPort Go's csrf protection | TypeScript · Logic | 0/1 | 5/6 |
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-6 Luna and GPT-OSS 120B (Workers AI).