dubinc/
The modern link attribution platform. Loved by world-class engineering teams at Twilio, Zillow, Jobber, Vercel, Raycast, and more.
Run by mupt-ai · Released 2026-11-01 · 21 tasks · 9 settings
Which coding agent works best on dubinc/dub.link? Most accurate: Claude Opus 5.5, 81.0% at $2.46 per task. Best for less: DeepSeek V4.1 Flash 1017, 76.2% at $0.31. 9 model settings scored on 21 tasks from its merged pull requests.
All Settings
| Model | Harness | Reasoning | Access | Accuracy | Cost / Task | Frontier |
|---|---|---|---|---|---|---|
| Claude Opus 5.5 | Claude Code | High | API Key | 81.0% | $2.46 | Yes |
| DeepSeek V4.1 Flash 1017 | Codex | High | Vercel AI Gateway | 76.2% | $0.31 | Yes |
| Dari Router | Pi | Default | API Key | 76.2% | $0.66 | No |
| GPT-6.1 Sol | Codex | High | API Key | 76.2% | $0.86 | No |
| Claude Opus 5.5 | Claude Code | Low | API Key | 71.4% | $0.90 | No |
| Claude Haiku 5.5 | Claude Code | High | API Key | 66.7% | $0.55 | No |
| Claude Haiku 5.5 | Claude Code | Low | API Key | 57.1% | $0.15 | Yes |
| GPT-6 Luna | Codex | High | API Key | 42.9% | $0.066 | Yes |
| GPT-6 Luna | Codex | Low | API Key | 4.8% | $0.013 | Yes |
Model Summaries
Claude Haiku 5.5 in Claude Code is a reasonable route for routine feature work, especially with high reasoning, while Claude Opus 5.5 in Claude Code with high reasoning is the strongest choice for difficult, large changes. DeepSeek V4.1 Flash 1017 in Codex with high reasoning, Dari Router in Pi with default reasoning, and GPT-6.1 Sol in Codex with high reasoning are also strong choices for substantial work. Higher reasoning broadens what Claude Haiku 5.5, Claude Opus 5.5 and GPT-6 Luna can handle, but GPT-6 Luna in Codex is not dependable at low reasoning and remains inconsistent at high reasoning.
Claude Opus 5.5 · Claude Code · High
A dependable choice for routine feature work and a strong default for difficult changes spread across many files. Successful work spans UI, APIs, persistent storage and external integrations, with no demonstrated size-related weakness among otherwise tractable changes. The evidence does not establish a distinctive failure pattern or time exhaustion, so there is no supported reason to restrict it to local edits. Keep normal end-to-end verification, and assess architectural refactors and standalone bug fixes separately because those work types are not well established here.
DeepSeek V4.1 Flash 1017 · Codex · High
A strong option for routine changes and difficult, large features spanning UI, APIs, persistent storage and external integrations. Broad multi-file scope is generally handled well, so size alone is not a reason to restrict it to local edits. For highly interdependent work, verify complete requirement coverage and consistency across application layers. The remaining failure evidence concerns incorrect new behavior with existing checks intact, but is too limited to establish a recurring weakness.
Dari Router · Pi · Default
A strong route for routine feature work and many difficult, large changes across UI, APIs, storage and external integrations. It can coordinate broad multi-file work, although that does not guarantee complete behavior across every application layer. Review requirement coverage and consistency between the UI, APIs and stored data rather than treating regression checks as proof of completion. The remaining failure evidence shows missing or incorrect new behavior while existing checks pass, but is too sparse to establish a recurring weakness.
GPT-6.1 Sol · Codex · High
A strong choice for routine feature extensions and difficult, large changes involving UI, APIs, persistent data and external integrations. Broad multi-file scope is usually handled well, although externally visible behavior still needs explicit verification. For data-facing work, check API contracts and backward compatibility alongside the user-facing result, because existing checks can remain green when new behavior is wrong. No time exhaustion was observed, but reliability on standalone bug fixes and architectural refactoring is not established by this evidence.
Claude Opus 5.5 · Claude Code · Low
Reliable for routine feature extensions and much difficult, large multi-file development. It has broad successful coverage across UI, APIs, persistent storage and external integrations, although some cross-layer changes still miss required behavior. Check that requirements are carried through every affected layer, even when regression checks pass. Its observed failures concern missing or incorrect functionality rather than time exhaustion or demonstrated regressions.
Claude Haiku 5.5 · Claude Code · High
A good route for routine feature extensions and many difficult, large changes spread across files. Successful work includes coordinated UI, API and persistence changes as well as external integrations, but demanding end-to-end work remains uneven. Check that all requested behavior works across application layers, since passing existing checks has not guaranteed completion. Its recurring failures concern missing or incorrect functionality rather than demonstrated regressions or time exhaustion.
Claude Haiku 5.5 · Claude Code · Low
Suitable for bounded routine feature work, with demonstrated ability to complete substantial changes across many files, though difficult work is uneven. Successful work spans UI, APIs, persistent data and external integrations. For larger changes, check complete requirement coverage and end-to-end behavior rather than relying on passing regression checks. Its recurring failures leave requested behavior missing or incorrect while existing checks remain green, without time exhaustion.
GPT-6 Luna · Codex · High
Use for bounded routine changes and carefully scoped difficult feature work, not as a dependable owner of broad changes. It can connect UI, APIs and storage and modify external integrations, but large multi-file work remains inconsistent. Check full requirement coverage and agreement between user-facing behavior and stored data. Its recurring failure pattern is to preserve the checked existing behavior while leaving requested functionality missing or wrong.
GPT-6 Luna · Codex · Low
There is no demonstrated dependable assignment class, including routine changes or difficult work spread across many files. Failures extend across UI, APIs, persistent data and external integrations, so isolated success does not establish a reliable specialty. Review requested behavior end to end, validate storage definitions and rerun regression checks before accepting a patch. The dominant problem is missing or incorrect functionality, with additional evidence of invalid storage definitions and broken existing behavior rather than time exhaustion.
Written by gpt-6.1-sol from these results on 2026-11-01.
Results by Task
Whether each setting passed each task, with how its attempt was graded and its transcript.
Tasks
All 21 tasks, each with its instruction, tests, and solution to browse and download.