Claude Opus 4.7: Programming Power Peaks but Faces ‘Dumbing‑Down’ Criticism

Release and headline performance

On 16 April 2026 Anthropic announced its flagship large model Claude Opus 4.7. The same day AMD senior director Stella Laurenzo published an audit report that highlighted a noticeable performance decline since February, sparking wide discussion in the developer community.

Programming benchmarks

Claude Opus 4.7 was evaluated on ten benchmarks, surpassing GPT‑5.4 on seven of them. Key results include:

SWE‑bench Verified: 87.6 % (Opus 4.7) vs 80.6 % (Gemini 3.1 Pro) – real GitHub issue fixes.

SWE‑bench Pro: 64.3 % (Opus 4.7) vs 57.7 % (GPT‑5.4) – open‑issue programming.

SWE‑bench Multilingual: 80.5 % (Opus 4.7) – multilingual code ability.

Terminal‑Bench 2.0: 69.4 % (Opus 4.7) – terminal engineering execution.

OSWorld‑Verified: 78.0 % (Opus 4.7) vs 75.0 % – computer GUI operations.

MCP‑Atlas: leading by +9.2 points – agent tool calling.

GPQA Diamond: 94.2 % – graduate‑level scientific QA.

Vision Multimodal: 98.5 % – visual image understanding.

BrowseComp: 79.3 % (Opus 4.7) vs 89.3 % (GPT‑5.4) – web search is the only lagging area.

Conclusion: Programming ability is clearly ahead, while web search remains a weakness.

Notable improvements

The model introduces several new features:

xhigh effort level: a fourth tier (low / medium / high / xhigh) that triggers more self‑verification and multi‑step reasoning. It is suited for large‑scale code refactoring (≈100 k lines), complex algorithm design, and cross‑module system architecture analysis. Activate by sending extra_body={"reasoning_effort": "xhigh"} in the API request.

Vision capability: benchmark score of 98.5 % with higher‑resolution image input, improving recognition of charts, flow diagrams, and UI screenshots.

Context window: a 1 M‑token window at the standard price, roughly equivalent to the full text of "War and Peace" (≈580 k words), a 100 k‑line code file, or an entire medium‑sized project's Git history.

Pricing

Opus 4.7 keeps the same pricing as Opus 4.6:

Input: $5 per million tokens.

Output: $25 per million tokens.

Performance concerns

Stella Laurenzo, an AMD senior director and heavy Claude Code user, audited 6,852 conversation logs, 17,871 reasoning blocks, and over 230 k tool calls. Her systematic re‑testing on BridgeBench (a cross‑architecture benchmark) showed:

Accuracy: 83.3 % → 68.3 % (‑15 percentage points).

Ranking: 2nd → 10th.

The drop means that, on average, more than 1.5 bugs appear for every 10 complex coding tasks.

“Intelligence reduction” three‑fold

Hidden reasoning: Since February, the UI no longer shows step‑by‑step reasoning, only the final answer.

Cache TTL cut: Context cache lifetime reduced from 1 hour to 5 minutes, causing token consumption to rise sharply when users pause.

Pricing model shift: Subscription changed from a fixed $200/month to a $20/month base with pay‑as‑you‑go overage, potentially raising heavy‑user spend to $400‑$600/month.

Why the changes?

Anthropic cites exploding inference costs as the primary driver—high‑frequency users (programmers, data scientists) can consume thousands of tokens per session. Competitive pressure from OpenAI’s Codex ($100/month tier) also influences the new strategy.

Selection advice

Complex code refactoring / agent development: Use Claude Opus 4.7 with xhigh mode.

Simple daily chat / light assistance: Choose Claude Sonnet 4.6 for better cost‑performance.

Strong web‑search requirement: Opt for GPT‑5.4, which leads BrowseComp by 10 points.

Price‑sensitive heavy users: Wait for Opus 4.7 price reduction or switch to another provider.

Final thoughts

From a technical standpoint, Claude Opus 4.7 is a milestone—its programming prowess, multimodal vision, and 1 M‑token context are impressive. However, the disclosed “intelligence reduction” measures cast a shadow over its commercial credibility, and the market will ultimately decide its fate.

Sources: Anthropic official blog, BridgeBench benchmark report, Quantum Bit, tenfy’s blog, APIYI technical review.