Updated: 2026-07-07

Anthropic released Claude Opus 4.7 positioning it as their most capable model with particular strength in long-horizon agentic work, knowledge work, vision, and memory. After two months running it on real workloads across several teams, the practical changes versus Opus 4.6 are subtler than marketing suggests but consistent.

Key takeaways

  • Opus 4.7 uses fewer tools than 4.6 and reasons more internally; this reduces tokens but can cut too short on tasks needing external information.

  • The effort: xhigh parameter is the primary lever for intensive agentic work.

  • Long-horizon improvement is real but depends on structure: phase-articulated instructions and subagents for bounded work.

  • Tasks Sonnet 4.6 solves at 95% quality don’t justify Opus 4.7’s cost.

  • A router assigning Opus only when complexity warrants saves 40-70% of spend.

Fewer tool calls, more reasoning

The most noticeable behavioural change: Opus 4.7 tends to use tools less than Opus 4.6 and reason more internally. In code exploration tasks, this translates to:

  • Fewer redundant reads of the same file.

  • Fewer similar keyword searches.

  • More inference from context already loaded.

The result is lower token cost (fewer tool calls) and sometimes better answers (more reasoning over what’s already known). The nuance: for tasks genuinely requiring external information, this tendency can cut short too soon. There, the effort parameter acts as counterbalance.

The effort parameter as lever

xhigh is the most useful setting for intensive agentic coding and complex knowledge tasks. Raising effort pushes the model to:

  • Use tools more.

  • Reason deeper.

  • Produce more elaborate responses.

In exchange for more time and tokens. In large code migrations or complex analysis, this lever markedly improves the result; in light tasks, raising effort wastes budget.

Practical rule: tasks with objective criteria (compiles, passes tests, migration runs without error) tolerate high effort well; tasks with subjective criteria (prose quality, design decisions) don’t necessarily improve with more effort and should be measured first.

Long horizon: when it’s a real advantage

The long-task improvement is real but depends on structure. Opus 4.7 sustains coherence better than 4.6 on 50-100 step chains when instructions are well articulated. On worse-structured chains, the difference is smaller.

The pattern that extracts maximum value:

  • Split the task into explicit phases.

  • Use subagents for bounded work.

  • Keep the orchestrator’s context clean.

Opus 4.7 executes this pattern well; without it, the gap with smaller models narrows.

Where it’s not worth it

Not everything justifies Opus 4.7:

  • Tasks already solved by Sonnet 4.6 in production at 95% quality with less cost don’t benefit from switching.

  • Light conversational use is better served by Haiku 4.5.

  • An inference router assigning Opus 4.7 only when complexity justifies it saves 40-70% spend without quality impact.

Conclusion

Claude Opus 4.7 is an incremental improvement with real value in specific cases: long horizon, agentic coding with multi-file tasks, complex analysis requiring deep reasoning. For the rest, it remains overkill. Maturity is using Opus where Opus delivers, Sonnet where it suffices, Haiku where it’s plenty. Knowing how to distinguish is part of the craft.

Frequently asked questions

Does effort: xhigh replace budget_tokens?

Yes. On Opus 4.7, manual thinking with budget_tokens is no longer available; reasoning depth is controlled via thinking: adaptive combined with the effort level, and xhigh is the recommended setting for coding and agentic work.

Is Opus 4.7 worth it for light conversational use?

No. For chat or short tasks, Haiku 4.5 delivers equivalent results at a fraction of the cost; reserving Opus for long horizon and complex analysis is what makes the skills and subagents pattern worth the engineering effort.

How much can a model router actually save?

In the teams I’ve worked with, 40% to 70% of inference spend, with no measurable quality loss, simply by routing Opus only to the tasks that need it.

This article is also available in Spanish.

Sources

  1. Anthropic — Effort, the parameter that controls reasoning depth
  2. Anthropic Engineering — Building effective agents
  3. METR — Measuring AI ability to complete long tasks