Claude Opus 5 Medium vs Xhigh: Which Effort Setting Should Developers Choose?

Claude Opus 5 (Adaptive Reasoning, Xhigh Effort) wins quality-sensitive coding, while Claude Opus 5 (Adaptive Reasoning, Medium Effort) is faster at the same listed cost.

The Xhigh configuration reaches 77 on the Artificial Analysis coding index versus 74.3 for Medium. Medium reaches 54.838 median output tokens per second versus 53.917 for Xhigh. Both report 0.3 seconds median latency. Data provided by https://artificialanalysis.ai/. Artificial Analysis supplies the comparison snapshot.

The practical choice is effort policy, not a different official model version. Anthropic documents claude-opus-5 as the model ID and stable alias, while effort controls the reasoning configuration (Models overview; What’s new in Claude Opus 5).

Claude Opus 5 (Adaptive Reasoning, Xhigh Effort) is a higher-scoring effort configuration, not a separate production model, while Medium is the lower-effort control.

Both entries carry the release date 2026-07-24 in the data brief. The official vocabulary resolves the naming ambiguity: use claude-opus-5 in production and set the requested effort level. The research material found no official claude-opus-5-medium or claude-opus-5-xhigh API ID (Models overview; What’s new in Claude Opus 5).

Decision area	Medium	Xhigh	What it means
Coding index	74.3	77	Xhigh has the stronger measured coding result
Intelligence index	56.3	60.1	Xhigh has the stronger measured general result
Median output speed	54.838	53.917	Medium produces tokens faster
Median latency	0.3	0.3	The supplied latency metric is tied
Blended price per 1M tokens	$10	$10	Unit pricing does not decide the choice

Anthropic’s current model overview keeps Opus 5 available and does not present Medium or Xhigh as separate lifecycle products. The official release positions the model for deep reasoning, agentic coding, code review, long-context work, and multi-agent tasks (Introducing Claude Opus 5). The evidence does not show that Xhigh adds a separate capability surface. It shows a different effort setting with stronger evaluation results in the supplied snapshot.

Claude Opus 5 (Adaptive Reasoning, Xhigh Effort) is the performance pick for tasks where deeper reasoning matters more than streaming speed.

The supplied coding result favors Xhigh at 77 versus 74.3 for Medium. That gap supports choosing Xhigh for complex repository changes, difficult debugging, and agent workflows where an incorrect plan creates expensive rework. Medium’s output speed of 54.838 versus 53.917 for Xhigh supports a different use case: interactive development where the developer frequently reviews and redirects the model.

Anthropic explicitly positions Opus 5 for long-horizon agentic coding, multi-file development, code review, visual understanding, and complex document work (What’s new in Claude Opus 5; Introducing Claude Opus 5). The performance data does not prove that Xhigh completes every real task faster. It only shows a stronger evaluation result and slightly slower token emission in this snapshot.

Community evidence is divided. A Reddit discussion describes long autonomous work as useful, while other users report slow responses, over-planning, verbosity, and instruction drift (Reddit discussion). Hacker News users praise the model’s ability to build supporting workflows, but warn that it may continue consuming tokens instead of requesting missing input (Hacker News discussion). Lenny’s public review also describes a model that can require repeated human direction in coding-agent work (Lenny’s review).

The evidence is insufficient to quantify rework, pass rates, tool-call counts, or wall-clock completion. Treat Xhigh as the stronger quality hypothesis, not as a guaranteed task-level winner.

Claude Opus 5 (Adaptive Reasoning, Medium Effort) and Claude Opus 5 (Adaptive Reasoning, Xhigh Effort) tie on listed token prices, so effort does not change the price card.

Both settings are listed at $10 per 1M blended tokens, with $5 input tokens and $25 output tokens. That tie shifts the cost decision from unit price to work completed per request. A setting that produces longer reasoning, more progress updates, extra validation, or additional tool activity can consume more budget even when the rate is unchanged.

Anthropic states that thinking tokens and the final response share the same max_tokens ceiling. The documentation also warns that default responses and delivery documents can become longer, with more progress reporting, verification, and sub-agent delegation (What’s new in Claude Opus 5). Those behaviors make Xhigh potentially more expensive in agent loops that pay for repeated turns or large tool traces.

Prompt caching adds another cost variable. Anthropic’s pricing page lists cache hits at $0.50 per 1M tokens, so repeated instructions and stable repository context should be measured separately from fresh prompts (Anthropic pricing). Hacker News discussion also raises the risk of extra token consumption when the model lacks necessary input but continues building an alternative workflow (Hacker News discussion).

The supplied evidence does not report token consumption, output length, tool-call volume, or cost per completed feature for either effort setting. Therefore, neither configuration can be called cheaper per successful task from the price card alone.

Claude Opus 5 (Adaptive Reasoning, Xhigh Effort) is the safer default for autonomous, high-stakes coding, while Claude Opus 5 (Adaptive Reasoning, Medium Effort) suits guided, interactive development.

Workflow	Recommended setting	Reason
Autonomous multi-file change	Xhigh	The supplied coding index is 77 versus 74.3, and Anthropic targets long-horizon agentic coding
Interactive coding and review	Medium	The supplied median output speed is 54.838 versus 53.917, with no listed price penalty
Repeated long prompts	Either	The base price is tied, while cache behavior and prompt shape determine spend
Strict tool workflows	Medium first	Lower effort is easier to test before increasing autonomy
High-cost failure scenarios	Xhigh with approvals	Stronger evaluation results do not remove the need for human review

For API integration, send the official claude-opus-5 model ID and set the effort parameter. Do not send the evaluation slugs as if they were separate model IDs (Models overview; What’s new in Claude Opus 5).

Xhigh should still run inside explicit tool permissions, tests, approval points, and output limits. Anthropic documents cases where disabled thinking can turn tool calls into ordinary text, expose internal XML, or interact poorly with high effort settings (What’s new in Claude Opus 5). The launch material also states that important limitations remain for long autonomous scientific research, so higher effort should not be treated as a general solution to model boundaries (Introducing Claude Opus 5).

The community signal does not justify a universal default. Reddit reports both successful long-running work and frustration with over-thinking or instruction drift (Reddit discussion). Lenny’s review describes the need for repeated human direction, while Hacker News reports strong autonomy alongside possible unnecessary token use (Lenny’s review; Hacker News discussion). Run the same repository tasks through both settings and measure successful completion, corrections, tool activity, and total spend before making a permanent choice.

Claude Opus 5 (Adaptive Reasoning, Medium Effort) and Claude Opus 5 (Adaptive Reasoning, Xhigh Effort) need production checks for API naming, effort control, and token budgeting before rollout.

Pre-commit checks

• Send the official claude-opus-5 model ID and treat Medium and Xhigh as effort settings (Models overview).
• Keep thinking behavior compatible with the selected effort. Anthropic warns that disabled thinking can expose tool-call text or internal XML in visible output (What’s new in Claude Opus 5).
• Budget max_tokens for both internal thinking and final response text, not only the visible answer (What’s new in Claude Opus 5).
• Add approval points, tests, and scoped tool permissions because community reports describe instruction drift, over-planning, and autonomous work that continues without enough input (Reddit discussion; Hacker News discussion).
• Measure completed-task cost and rework. The supplied snapshot gives unit prices and evaluation indicators, not operational cost per feature.