Claude Opus 4.7 — when extended thinking earns its cost

What Opus 4.7 actually changed

Anthropic released Claude Opus 4.7 on April 16, 2026. The headline was not a new benchmark — it was that Opus, which had cost $15/$75 per million tokens on earlier generations, now ships at $5 in / $25 out with a 1M token context window. That is standard Sonnet-tier pricing for a flagship model, and it changes which tasks are cost-justified on Opus.

Extended thinking is the key feature

When you enable extended thinking, Claude spends hidden reasoning tokens before emitting the answer. You set a budget; Claude spends up to that budget. The bigger the budget, the deeper the search over approaches. Opus 4.7 handles budgets of tens of thousands of tokens gracefully; Sonnet starts to lose coherence at that scale.

Where extended thinking is decisive

• Multi-step mathematical proofs where every step has to be verified
• Debugging code that involves invariants across 5+ files — the thinking token phase lets Claude trace the dependency graph
• Legal or policy analysis where multiple clauses interact and the reasoning has to be auditable
• Planning agentic work where the wrong initial plan costs hours of execution

Where it is wasted

• Summarization (the model already knows how)
• Simple Q&A with known facts
• Creative writing — thinking tokens tend to make the final text worse by over-constraining
• Anything where you would accept the first plausible answer

A production-grade call

from anthropic import Anthropic client = Anthropic() resp = client.messages.create( model="claude-opus-4-7", max_tokens=8192, thinking={ "type": "enabled", "budget_tokens": 16000 }, messages=[ {"role": "user", "content": "Given the attached 400-page regulatory filing, identify every section that conflicts with the EU AI Act provisions on high-risk systems."} ] ) # Inspect what Claude was thinking (optional, advanced) for block in resp.content: if block.type == "thinking": pass # hidden reasoning, not shown to end user elif block.type == "text": print(block.text) print(resp.usage) # input_tokens, output_tokens, thinking_tokens separateThe thinking block is part of the response but typically not shown to the end user. You can log it for audit, but never paste it back as context — it is one-shot reasoning.

The cost model you need to internalize

• Input: $5 per million tokens (cached: ~$0.50)
• Output: $25 per million tokens (includes thinking)
• Extended thinking with 16k budget ≈ $0.40 of output cost per call baseline
• Prompt caching makes repeat calls over long documents economically viable
• At small call volume, Opus with thinking is usually cheaper than failing and retrying on a smaller model

The right question is not 'can I afford Opus with extended thinking?' — it is 'can I afford to be wrong on this task?'

— Common framing from Anthropic's solutions engineers