Claude Opus 4.7 Boosts Programming Performance by 11% – Why Its ‘No’ Makes It More Reliable

Programming performance

Opus 4.7 achieves 64.3% on SWE‑bench Pro, up from 53.4% in Opus 4.6 (+10.9 pp). On SWE‑bench Verified it scores 87.6% versus 80.8% (+6.8 pp). CursorBench improves from 58% to 70%.

Compared with GPT‑5.4 (57.7% on SWE‑bench Pro) and Gemini 3.1 Pro (54.2%), Opus 4.7 leads.

Key insight : the larger gain on the harder Pro benchmark indicates stronger handling of complex tasks.

Enterprise case studies

Cursor : on 93 programming tasks success rate rose by 13%, solving four tasks that 4.6 could not.

Rakuten : production‑grade bug fixes increased three‑fold.

XBOW : visual accuracy improved from 54.5% to 98.5% (+44 pp).

Notion : agent‑task success up 14% and tool‑error rate fell to one‑third.

Unattended Rust TTS engine

Opus 4.7 built a complete Rust text‑to‑speech engine from scratch without human intervention, demonstrating capability on long‑running, complex pipelines.

Visual capability

Maximum long‑side resolution increased from ~860 px to 2 576 px (≈3×). Total pixel count grew from ~1.2 MP to 3.75 MP (≈3×). XBOW visual test matches the resolution gain (54.5% → 98.5%).

Result: dense screenshots become readable for computer‑use agents; extraction of scientific charts, financial statements, and technical drawings gains precision.

Note : high‑resolution images consume more tokens; pre‑downsampling is advisable when ultra‑high precision is unnecessary.

Reliability – willingness to say “no”

When a user plan contains errors, Opus 4.6 executed blindly; Opus 4.7 points out the problem.

When information is missing, Opus 4.6 attempts to fill (may be wrong); Opus 4.7 reports the error directly.

During output generation, Opus 4.6 may fabricate answers; Opus 4.7 validates internally before returning.

“Claude 在技术讨论中会反驳我，帮我做出更好的决定。” – Replit

“Opus 4.7 遇到缺失数据会直接报错，而 4.6 往往会尝试填充可能错误的值。” – Hex

“它在编写系统级代码前会自行进行数学证明，确保逻辑严谨。” – Vercel

BrowseComp regression

Score dropped from 83.7% to 79.3% because the model now reports missing information instead of fabricating it, allowing GPT‑5.4 (89.3%) to overtake.

New features

xhigh effort level – granularity between high and max for finer trade‑off between reasoning depth and speed; default for Claude Code.

Auto Mode extension – model issues command → classifier evaluates safety; safe commands auto‑approved, risky commands fall back to the traditional permission prompt. Available to Max, Teams, Enterprise.

模型发出命令 → 分类器评估安全性
 ├── 安全 → 自动批准，无需用户介入
 └── 风险 → 保留传统权限提示

/ultrareview command – opens a dedicated session for code review, scanning changes and flagging bugs or design issues.

Task Budgets (API beta) – helps developers plan token spend for long‑running tasks.

Pricing

Input $5 / million tokens, output $25 / million tokens – unchanged from the previous version.

Migration considerations

New tokenizer increases token count by 1.0–1.35× for the same content.

High‑compute mode (“thinking”) raises output token count.

Stricter instruction compliance may require prompt adjustments.

Availability

All Claude product lines (model ID claude-opus-4-7 on Anthropic API).

Amazon Bedrock.

Google Cloud Vertex AI.

Microsoft Foundry.

Conclusion

Opus 4.7 leads in programming (+11 pp on SWE‑bench Pro), tool invocation, and domain‑specific tasks, while its stricter refusal policy causes a modest drop in long‑document search performance. Token consumption may rise up to 35%.