Recommendation for Debug / review
Debugging & Code Review
Our top recommendation for Debugging & Code Review, based on the public evidence we track, is Anthropic: Claude Fable 5.[1][2][3] Assign to security-focused code review and advisory roles, as upstream policy explicitly routes plan-review, code-review, advise, and security work to Fable 5 in the ultra tier (e13). Watch out: Watch for workflow agent spawning bugs where subagents inherit the original session model rather than respecting mid-session /model switches (e15). Z.ai: GLM 5.2 is the next-ranked alternative. Deploy via OpenRouter for automated PR review workflows, with confirmed production runs posting review comments on consumer PRs (e21).
About this recommendation
- Updated
- Sep 25, 2026
- Evidence through
- Sep 25, 2026
- Sources
- 13
- Revision
- v82
Decision audit
Why this result
Inspect the inputs and the computed order behind the recommendation.
Models screened
23
live candidates
Evaluation feeds
6
task-weighted
Winner coverage
87%
intended feed weight
Largest provider share
3 of 6
Anthropic
Sources evaluated
The task sets these weights before any model is scored.
| Evaluation feed | Weight | Winner result | Field measured |
|---|---|---|---|
| SWE-rebench | 30% | #1 | 17/23 |
| Aider Polyglot | 20% | not measured | 4/23 |
| LiveBench Coding | 15% | #3 | 15/23 |
| LMArena WebDev | 15% | #11 | 17/23 |
| OpenRouter usage | 10% | 79/100 | 23/23 |
| SWE-bench Verified | 10% | not measured | 5/23 |
Provider concentration
Each exact model is scored separately; provider identity is not a ranking input.
- Anthropic3 models
- Google1 model
- OpenAI1 model
- Z.ai1 model
Decision table
Every published model is shown in computed order. Practitioner sources are distinct community threads, not the citations repeated in the prose below.
| Rank | Model | Relative score | Coverage | Practitioner evidence | Strongest measured reason |
|---|---|---|---|---|---|
| 01 | Claude Fable 5Anthropic | 74 | 87% | 5 threads · 4 families · 0 cautions | #1 SWE-rebench · #3 LiveBench Coding |
| 02 | GLM 5.2Z.ai | 68 | 87% | 13 threads · 6 families · 7 cautions | #9 SWE-rebench · #16 LiveBench Coding |
| 03 | Claude Opus 5.5Anthropic | 64 | 48% | 2 threads · 2 families · 0 cautions | #1 LiveBench Coding · #1 LMArena WebDev |
| 04 | Claude Opus 4.6Anthropic | 60 | 87% | 1 threads · 1 families · 1 cautions | #16 SWE-rebench · #25 LiveBench Coding |
| 05 | Gemini 2.5 Pro Preview 06-05Google | 25 | 28% | no linked practitioner threads | #2 Aider Polyglot · #6 SWE-bench Verified |
| 06 | o3 ProOpenAI | 23 | 23% | no linked practitioner threads | #1 Aider Polyglot |
Relative score combines normalized benchmark quality and signal coverage; independent practitioner evidence and freshness are bounded tie-breakers. It is an ordering score, not an absolute quality percentage. The writing model receives this order and cannot change it.
Claude Fable 5 is explicitly designated for plan-review, code-review, advise, and security roles in production tier-taxonomy policies, with strong benchmark results including 85.99% on LiveBench Coding.
Best when: Assign to security-focused code review and advisory roles, as upstream policy explicitly routes plan-review, code-review, advise, and security work to Fable 5 in the ultra tier (e13).
Tips
- Assign to security-focused code review and advisory roles, as upstream policy explicitly routes plan-review, code-review, advise, and security work to Fable 5 in the ultra tier (e13).
- Use for cost-sensitive review automation where Dependabot patch/minor PRs can be downgraded to cheaper models while keeping majors on Fable 5 (e14).
- Leverage for agent-heavy coding workflows, resolving 34.05% of Frontier-Bench tasks and 64.5% on SWE-rebench with tool use (e17, e18).
Watch out for
- Watch for workflow agent spawning bugs where subagents inherit the original session model rather than respecting mid-session `/model` switches (e15).
GLM 5.2 runs successfully in production PR review workflows via OpenRouter, though evidence suggests it costs 4x DeepSeek V4 with comparable bug-finding to Qwen 3.7 Max and behind smaller open models on specific bug hunts.
Best when: Deploy via OpenRouter for automated PR review workflows, with confirmed production runs posting review comments on consumer PRs (e21).
Tips
- Deploy via OpenRouter for automated PR review workflows, with confirmed production runs posting review comments on consumer PRs (e21).
Watch out for
- Budget for significantly higher token costs than alternatives, with one user reporting $5 in 30 minutes of relatively slow work, roughly 4x DeepSeek V4 cost (e25).
- Expect bug-finding performance comparable to Qwen 3.7 Max and behind smaller self-hostable models like Gemma 4 and Qwen 3.6 on specific bug hunt benchmarks (e24).
Claude Opus 5.5 leads LiveBench Coding at 89.25% and uniquely caught a critical payment bug in a 3,600-line PR that cheaper API models missed entirely.
Best when: Use for high-stakes code review where cheaper models fail, as it was the only model to identify a must-level payment bug in a large financial PR while GLM-5.3-flash, V4.1-flash, and others missed it (e27).
Tips
- Use for high-stakes code review where cheaper models fail, as it was the only model to identify a must-level payment bug in a large financial PR while GLM-5.3-flash, V4.1-flash, and others missed it (e27).
- Deploy for maximum coding benchmark performance, scoring 89.25% on LiveBench Coding (#1 of 58) (e28).
Claude Opus 4.6 scores 78.18% on LiveBench Coding (#29 of 58), with community reports of effective vulnerability finding in source code and pentesting without jailbreaks.
Best when: Use for vulnerability discovery in source code and pentesting without source, as users report it finds vulns 'just fine' without jailbreaking (e45).
Tips
- Use for vulnerability discovery in source code and pentesting without source, as users report it finds vulns 'just fine' without jailbreaking (e45).
Gemini 2.5 Pro Preview 06-05 scores 83.1% on the Aider polyglot coding benchmark (#3 of 27), placing it competitively for cross-language code editing.
Best when: Use for polyglot code editing where Aider benchmark results guide selection, scoring 83.1% and placing third (e46).
Tips
- Use for polyglot code editing where Aider benchmark results guide selection, scoring 83.1% and placing third (e46).
o3 Pro scores 84.9% on the Aider polyglot coding benchmark (#2 of 27), demonstrating strong cross-language code editing capability.
Best when: Use for polyglot code editing tasks where Aider benchmark performance predicts success, scoring second place at 84.9% (e47).
Tips
- Use for polyglot code editing tasks where Aider benchmark performance predicts success, scoring second place at 84.9% (e47).
Frequently asked
- What is the top-ranked model for Debugging & Code Review?
- Anthropic: Claude Fable 5 ranks first in the current evidence-weighted comparison. Assign to security-focused code review and advisory roles, as upstream policy explicitly routes plan-review, code-review, advise, and security work to Fable 5 in the ultra tier (e13).[1]
- What should I watch out for with Anthropic: Claude Fable 5?
- Watch for workflow agent spawning bugs where subagents inherit the original session model rather than respecting mid-session `/model` switches (e15).[2]
- What is an alternative to Anthropic: Claude Fable 5?
- Z.ai: GLM 5.2 is the next-ranked option. Deploy via OpenRouter for automated PR review workflows, with confirmed production runs posting review comments on consumer PRs (e21).[3]
Sources
- 1
“## Summary A five-model advisor review of the newly merged model-tier taxonomy (`c266410`, agent-harness#309) converged on an author/reviewer split, but rejected making the heavy tier the default implementation tier. Recommended upstream policy: - roadmap/plan/supervise: Claude Opus 5 (`heavy`) - execute/repair: Claude Sonnet 5 (`regular`), with a recorded escalation to Opus 5 after repeated verification failure - plan-review/code-review/advise/security: Claude Fable 5 (`ultra`) - worker/econom…”
ViperJuice · GitHub · Jul 26, 2026 - 2
“## Problem Workflow subagents run on the model the session **started with**, not the model the session is **currently on**. Switching models mid-session with `/model` never reaches workflow agent spawning. Evidence from a live fleet session: 1. Started a pi session with model `aigw/round-robin/qwen3.8-max`. 2. Switched to `dongli/claude-fable-5` via `/model` (~1h later). 3. Ran the built-in `code-review` workflow. All 7 agents spawned with `aigw/round-robin/qwen3.8-max` (recorded in the run fil…”
Explorer1092 · GitHub · Aug 11, 2026 - 3
“Blackbox verification complete: after merging #3, the reusable workflow ran on five consumer PRs (uuta/resorn#145, uuta/SpeakRush#24, uuta/uuter#16, uuta/astro-blog#20, uuta/dotfiles#27). All five AI Review runs succeeded; resorn run 29825001965 logs confirm `openrouter/z-ai/glm-5.2` was the model actually called (not the fallback), and the PR Reviewer Guide comment was posted.”
uuta · GitHub · Jul 21, 2026 - 4
“## What Forge now cost-routes **Dependabot patch/minor** PR reviews to a cheaper model (`claude-sonnet-4-5`), keeping majors and everything else on the fleet default (`claude-fable-5`). The routing decision is made **server-side** in Forge (`/api/v1/reviewer-models`, shipped in TrueTime-Health/forge#176) — but it can only route when the calling workflow hands Forge the PR's author and title. This repo's `claude-code-review.yml` doesn't yet, so its Dependabot PRs still bill the expensive reviewe…”
truetime-helper[bot] · GitHub · Aug 20, 2026 - 5
“using Claude Code at max effort, resolves 34.05% ± 1.71 of Frontier-Bench tasks (#7 of 16 public submissions), measuring a submitted model-and-agent configuration on diverse, difficult agent work.”
Frontier-Bench · Benchmark · Jul 13, 2026 - 6
“Resolves 64.5045045045045% ± 1.4130078505728048 on SWE-rebench (#1 of 13) using tools, a continuously refreshed repository-issue evaluation with configuration recorded separately from the model.”
SWE-rebench · Benchmark · Jul 1, 2026 - 7
“I gave GLM 5.2 a spin on openrouter yesterday and it was mostly fine but it racked up $5 in token use in 30 minutes of (relatively slow) work. It's easily 4x the cost of DeepSeek V4 but I didn't actually feel the results were that much better. I had GPT 5.5 in Codex review it after it was done and there was plenty of slop to go around. Having better luck with MiniMax M3, from a cost benefit ratio.”
cmrdporcupine · Hacker News · Jun 17, 2026 - 8
“I added it to my benchmark based on Mythos-reported bugs, and it's better than GLM 5.1, but still behind several other models, maybe most directly comparable to Qwen 3.7 Max. But, several other open models, including small self-hostable ones (Gemma 4 and Qwen 3.6), found the same number of bugs, 3 of 9. Though it also gets partial credit for reporting one bug in the right spot, but kinda misunderstanding the bug. I also added Kimi K2.7-code in the same run, and it did poorly, consistent with 2.…”
SwellJoe · Hacker News · Jun 17, 2026 - 9
“## 目的 Claude のサブスクリプション（Opus 5.5）を、pi-web-idd の lane の主力にします。#30 で導入する Claude Code の非対話 adapter を土台に、Issue lane と review lane を実際に Claude のサブスクリプションで運用するための残りの作業を整えます。 ## 背景 - 2026-09-24 の比較では、3,600 行の決済 PR（Meltly #5638）の must（取得の失敗中でも決済に進め、画面に出ていない商品が支払いに含まれる）を見つけられたのは、Claude Code 上の Opus だけでした。API の安い model（glm-5.3-flash・v4.1-flash・glm-5.3 high・mimo-v2.6-pro）は見落とし、API の Opus 5.5 は Vercel の 429 により実行できませんでした。 - 仕事量は、1 か月で Issue 67 本（約 126 ポイント）、他人の PR のレビュー 67 本、自分の PR 約 69 本とその指摘への対応です。 - この量…”
penne-0505 · GitHub · Sep 24, 2026 - 10
“Scores 89.25% on LiveBench Coding (#1 of 58), an objective evaluation of code generation and completion.”
LiveBench Coding · Benchmark · Jun 25, 2026 - 11
“Claude Opus 4.6 and 4.8 find vulns in source code just fine and 4.6 will pentest without source for you given a proper harness WITHOUT jailbreaking. WITH jailbreaks, you can probably imagine what they are capable of. Anthropic guardrails seem to be more about protecting their business (distillation), than they are about public safety.”
sciencejerk · Hacker News · Jun 11, 2026 - 12
“Scores 83.1% on the Aider polyglot coding benchmark (#3 of 27), which tests editing real code across many languages.”
Aider polyglot benchmark · Benchmark · Sep 24, 2026 - 13
“Scores 84.9% on the Aider polyglot coding benchmark (#2 of 27), which tests editing real code across many languages.”
Aider polyglot benchmark · Benchmark · Sep 24, 2026
Rankings synthesized from community evidence and open benchmarks. See methodology. Not driven by vendor marketing.