Claude Opus 4.6
Claude Opus 4.6
Legacy model. Claude Opus 4.6 is a real but legacy model in the Claude 4 family. The current flagship Opus is claude-opus-4-8, which supersedes it. Pricing and context window specs below apply to both versions; see the Extended Thinking section for differences.
Anthropic's top-tier model in the Claude 4 family. Not primarily a coding benchmark winner — Opus is positioned for complex multi-step reasoning, long-horizon agent tasks, and situations where claude-sonnet-4-6 makes too many mistakes to recover from.
Key Specs
- Context window: 1M tokens
- Output tokens: 128K — unusually high, making it practical for tasks that generate large artifacts (full codebases, long reports, extensive plans)
- Pricing: $5 input / $25 output per million tokens (same for Opus 4.6 and Opus 4.8)
- Designed for orchestration roles in multi-agent-setup
Extended Thinking
Claude Opus 4.6 supports extended-thinking with manual budget_tokens control as well as adaptive thinking. The successor, claude-opus-4-8, uses adaptive thinking only — it does not support manual budget_tokens configuration.
When Opus Outperforms Sonnet
- Agent team orchestration: When coordinating multiple subagents, the planner needs fewer mistakes. Opus's lower error rate on ambiguous tasks makes it worthwhile as the top-level orchestrator.
- Complex synthesis: Integrating information across very long contexts where reasoning chains need to stay coherent.
- High-stakes single-shot tasks: When you can't iterate interactively and need the first attempt to be solid.
Pricing Reality
At $25/M output tokens with 128K output capacity, a single maxed-out response costs ~$3.20 in output alone. This is a model you use when task quality justifies cost, not for routine work. In practice, most agentic tasks don't hit 128K output, but the ceiling matters for document-generation workloads.
Strengths
- Highest reasoning ceiling in the Claude family
- 128K output token limit opens up large artifact generation
- Solid instruction following on complex, multi-constraint prompts
- Well-suited for orchestrator role in agentic-workflows
Weaknesses
- Expensive — 5x Sonnet's input cost, 1.67x output cost
- Slower than Sonnet at generation — noticeable in interactive use
- On pure coding benchmarks, often within noise of Sonnet. The extra cost is rarely justified for standard coding tasks.
Use Cases
Best for: orchestrating multi-agent-setup, generating large codebases or documentation in a single pass, tasks requiring sustained multi-step reasoning across long contexts.
Skip for: routine coding tasks, quick lookups, anything Sonnet handles reliably.
Related
claude-sonnet-4-6 · multi-agent-setup · agentic-workflows · extended-thinking