Claude Opus 5 Stumbles: Higher Reasoning Leads to Lower Scores

Episode 1 · Opus 5 Coding Paradox: Higher Reasoning Leads to Lower Scores (2026-07-24, 13 posts)

Claude Opus 5 reveals a counterintuitive phenomenon in the FrontierCode 1.1 benchmark: increasing reasoning strength to xhigh or max degrades coding scores, making medium effort the optimal setting. This provides practical reference for developers in model selection and configuration.

Confirmed

In FrontierCode 1.1 tests, Opus 5 performs better at medium reasoning effort than at xhigh and max. According to system card data cited by @inductionheads and @keunwoochoi, approximately 20-30% of reported benchmarks show max thinking scoring lower than xhigh. @silasalberti and @hero88645 point out that the core reason for the score drop is that FrontierCode evaluates not just correctness, but also code merge-ability. Its scope metric penalizes the excessive modifications or non-compliant code scope expansion generated by Opus 5 under high reasoning states. Additionally, @thesaraharminta supplemented with Vibe Code Bench data, noting that Opus 5's accuracy rises from low (76.7%), medium (82.0%) to high (88.0%), proving that high reasoning gears still hold an advantage in standard code generation and achieve the best cost-effectiveness.

Unconfirmed

@draginol proposes a speculation that Opus 5's poor performance under high effort might not be a decline in the main model's capability, but rather because it relies more on non-Opus 5 sub-agents to share tasks under this mode, dragging down the overall evaluation score. This mechanism still requires more internal details for confirmation.

Why it matters

This finding breaks the conventional assumption that "more thinking equals better performance." @seh0872 and @brandongalang believe this proves that Opus 5 at medium reasoning effort is the most suitable default state for daily use. It not only avoids code boundary issues caused by "overthinking" but also effectively controls the high cost of the Opus tier, providing a more cost-effective configuration solution for AI programming workflows.

• Anthropic effort settings and benchmark data offer a practical guide for choosing Sonnet vs. Opus — seh0872 · 2026-07-24
• FrontierCode results suggest Claude Opus 5 is the best default at medium reasoning — brandon_galang · 2026-07-25
• Charts Show Opus 5 Peaks in Coding Performance with Medium Thinking — dejavucoder · 2026-07-25
• FrontierCode charts show Claude Opus 5 peaking at medium reasoning effort — zainhas · 2026-07-25
• Claude Opus 5 scores better on FrontierCode at medium reasoning than at max — zainhas · 2026-07-25
• FrontierCode 1.1 says Opus 5 drops in score at xhigh and max reasoning settings — inductionheads · 2026-07-25
• FrontierCode 1.1 shows Opus 5 can score lower under stricter reasoning settings — andrew_n_carr · 2026-07-25
• Opus 5 max thinking reportedly underperforms xhigh on 20–30% of benchmarks — keunwoochoi · 2026-07-25
• FrontierCode’s design explains why Opus 5 scores drop as reasoning increases — silasalberti · 2026-07-25
• Opus 5 beats higher effort on FrontierCode at medium effort — kieranklaassen · 2026-07-25
• Claude Opus 5 may be scoring lower on hard benchmarks because it offloads work to subagents — draginol · 2026-07-25
• Claude Opus 5 peaks at high reasoning level on Vibe Code Bench, then gets pricier — thesaraharminta · 2026-07-25
• Opus 5’s coding scores reportedly drop above “high” effort, not at max — hero88645 · 2026-07-25

Episode 2 · Claude Opus 5 Shows Erratic Behavior in Probability Tests (2026-07-25, 3 posts)

Benchmark screenshots reveal that Claude Opus 5 exhibits unstable reasoning, repeatedly changing its answer up to 30 times on a single probability question before converging, raising concerns about its reliability.

• Claude Opus 5 flips between 1/3 and 2/5 before finally settling on an answer — Sauers_ · 2026-07-25
• Claude Opus 5 flips 30 times on a probability problem before settling on one answer — scaling01 · 2026-07-25
• A second Claude Opus 5 screenshot shows the same kind of reasoning instability — scaling01 · 2026-07-25

Episode 3 · GPT-5.6 Beats Claude Opus 5 in Bug Hunt Benchmark (2026-07-25, 3 posts)

In a Bug Hunt Bench test involving 45 hidden bugs in a VS Code extension, GPT-5.6 Sol successfully fixed 22 bugs, outperforming Claude Opus 5 which fixed 11 to 12 bugs and older versions that missed these specific issues.

• Bug Hunt Bench: GPT-5.6 Sol fixes 22 bugs, Opus 5 12, on a 45-bug repo — PawelHuryn · 2026-07-25
• Claude Opus 5 fixes 11 of 45 hidden bugs, versus 2 for Opus 4.8 — PawelHuryn · 2026-07-25
• Claude Opus 5 spots code issues earlier versions and Codex kept missing — PawelHuryn · 2026-07-25