DeepSeek V4 Flash (Workers AI) vs GPT-5.6 Sol
Reading a 26.2K-token pull request costs $0.1048 on GPT-5.6 Sol, 9.1× the $0.0115 it costs on DeepSeek V4 Flash (Workers AI). DeepSeek V4 Flash (Workers AI) is a economy model and GPT-5.6 Sol a frontier one. In the benchmark GPT-5.6 Sol caught 24 percentage points more of the known defects than DeepSeek V4 Flash (Workers AI) (95% interval −36 pts to −12 pts, a − b).
Side by side
| Figure | DeepSeek V4 Flash (Workers AI) | GPT-5.6 Sol |
|---|---|---|
| Price tier | Economy | Frontier |
| Bugs caughtshare of known bugs | 49% | 72% (better) |
| False alarmsper clean PR | 1.3 | 1.2 (better) |
| Cost per reviewmeasured | $0.0178 (better) | $0.1127 |
| Cost per bug caught | $0.0309 (better) | $0.1732 |
| Exact linecatches on the expected lines | 88% | 92% (better) |
| Review timemean | 137s | 50s (better) |
| Typical review calllist price | $0.0147 (better) | $0.1532 |
| Input / 1M tokens | $0.44 (better) | $4.00 |
| Output / 1M tokens | $1.32 (better) | $20.00 |
| Context window | 1M | 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 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| DeepSeek V4 Flash (Workers AI) | 49% | 46% | 25% | 55% | 46% | 22% | 56% | 61% | 78% | 61% | 44% | 67% | 50% | 0% |
| GPT-5.6 Sol | 72% | 67% | 50% | 69% | 75% | 56% | 100% | 67% | 100% | 100% | 67% | 100% | 50% | 0% |
Of the 42 known bugs both were scored on, both caught 27 and neither caught 7. DeepSeek V4 Flash (Workers AI) caught 3 bugs that GPT-5.6 Sol missed: 1 in Go, 1 in Kotlin and 1 in TypeScript, most of them logic bugs. GPT-5.6 Sol caught 5 bugs that DeepSeek V4 Flash (Workers AI) missed: 2 in TypeScript, 1 in C#, 1 in Go and 1 in JavaScript.
On pull requests with no known bug, DeepSeek V4 Flash (Workers AI) raised 1.3 false alarms per review and GPT-5.6 Sol 1.2. A measured review cost $0.0178 on DeepSeek V4 Flash (Workers AI), 84% less than $0.1127 on GPT-5.6 Sol.
| Pull request | Language · kind | DeepSeek V4 Flash (Workers AI) | GPT-5.6 Sol |
|---|---|---|---|
| withastro/astroPort Go's csrf protection | TypeScript · Logic | 3/4 | 0/1 |
| kotest/kotestFix DataClassEq comparator | Kotlin · Logic | 1/3 | 0/1 |
| traefik/traefikAdd the option to define custom cipher suites for backend serversTransport | Go · API misuse | 1/4 | 0/1 |
| OrchardCMS/OrchardCoreSimplify using `StringValues` in Liquid templates | C# · Security | 0/3 | 1/1 |
| TanStack/queryfix(query-core): fix hydration bugs for already resolved promises | TypeScript · Logic | 0/4 | 1/1 |
| colinhacks/zodfeat(v4): parse input containing reference cycles | TypeScript · Resource handling | 0/4 | 1/1 |
| go-chi/chifeat: middleware.ClientIP, a replacement for middleware.RealIP | Go · Security | 0/3 | 1/1 |
| sindresorhus/p-throttleAdd `weight` option | JavaScript · Logic | 0/3 | 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: DeepSeek V4 Flash (Workers AI) and GPT-5.6 Sol.