Claude Opus 4.7 Review: The Coding Partner Hits Its Peak

I put Anthropic's latest flagship, Claude Opus 4.7, through two weeks of real production work. Verdict: the most polished coding LLM available right now.

Scorecard

What's new

The biggest jump from previous versions is sub-agent coordination. When a parent agent delegates, the child interprets correctly and returns a usable result without redundant back-and-forth. Agent SDK users will feel this immediately.

Versus the competition

Against GPT-5 family, Claude wins on long context, parallel tool calls, and instruction faithfulness. GPT still has the edge on visual reasoning and is slightly faster end-to-end.

Bottom line

95/100. Make this your default coding partner. If cost matters, pair it with Haiku 4.5 for the simple stuff.

What "tool-call reliability" actually looks like

Tool-call reliability sounds like a benchmark term. In daily use it shows up as a specific pattern: the model picks the right tool the first time, supplies all required arguments, handles the result correctly, and doesn't loop.

Across the two-week trial running an agent that orchestrated bash, file editing, web search, and a custom code-review tool, the failure modes broke down like this:

The most important number is the bottom one. Once a model stops inventing tool names, you can run it unattended on long-horizon tasks. That's the threshold that turned 4.7 from "great pair programmer" into "background worker."

The 1M context window: still under-used

Anthropic shipped the 1M context window quietly with 4.6 and refined the cache behavior in 4.7. The result is that putting a whole monorepo in the prompt is now genuinely viable, with the obvious caveat that you pay for every token you send.

Real-world patterns that emerged across two weeks:

• Codebase Q&A is best with a 200–400K window and tight system instructions. Going further dilutes retrieval quality
• Cross-file refactors are where the long context actually shines — point the model at every file it might need to touch and let it reason about consistency
• Multi-document summarization of legal or technical materials works well up to roughly 600K tokens before quality starts to degrade noticeably

Anthropic's prompt caching is essential to make any of this affordable. Without it, the per-call cost of a 500K prompt is hard to justify outside of single-shot tasks.

Where it's still weaker than GPT-5

This isn't a Claude fan letter. Two areas where GPT-5 family still feels stronger in honest comparison:

1. Visual reasoning — analyzing screenshots, chart interpretation, OCR on complex layouts. GPT's multimodal pipeline still has the edge
2. Latency on short tasks — for sub-second responses on simple queries, GPT and Gemini Flash are still ahead. Opus 4.7's first-token latency is roughly 30% higher than GPT-5's mini variant

For pure speed-to-first-token, route the easy work elsewhere and reserve Opus for what it does best.

Pricing the cost trade-off honestly

The "★★★☆☆" cost rating reflects the fact that Opus 4.7 is expensive on output tokens compared to GPT-5 mini or Haiku 4.5. The math that justifies it:

• Output token cost is roughly 3–4× a mid-tier model
• First-pass correctness on coding tasks is roughly 2× a mid-tier model, meaning fewer retries
• Time saved verifying and debugging the output is the biggest invisible win

For experienced engineers using it on production code, the net is still positive — but only because the human time saved dwarfs the API cost. For learning use cases or hobby projects, the cost gap matters more.

The recommended setup

After two weeks I converged on this stack:

• Opus 4.7 for: cross-file refactors, debugging complex bugs, anything involving more than three tools
• Haiku 4.5 for: file searches, simple edits, formatting tasks, one-shot queries
• A routing rule that escalates to Opus when Haiku returns low-confidence answers

Total monthly cost on a moderate workload: roughly 40% of what Opus-everywhere would have been, with no perceptible quality regression on the tasks Haiku handles.

Read also

• Three Rules Behind a ¥500K/Month Indie SaaS
• Smart Glasses I Can't Take Off

Related: privacy around AI usage

When you run Claude against repositories or production data, a VPN keeps your traffic shape away from your default ISP routing. Useful for client work.