The short version: a rigorous third-party benchmark that runs the same test across all five Opus 4.8 effort levels and publishes a score for each one does not exist yet. Almost every benchmark you see quoted comes from Anthropic's own system card, run at max effort (Terminal-Bench is the one exception, run at high). So the published numbers show what the model can do at the top of the dial, not how it degrades as you turn it down. Below is everything that genuinely exists.

The five levels (there are five, not three)

In Claude Code and the API: low, medium, high, xhigh, max. In the claude.ai / Cowork app the dial reads High / Extra / Max, and "Extra" is exactly xhigh. Same control, two vocabularies. On the Max plan the app only exposes High, Extra, and Max (not low/medium).

EffortBest forTrade-off
LowClassification, quick lookups, high-volume simple jobs, subagentsBiggest token savings, some capability drop
MediumRoutine drafting, summaries, everyday questionsBalanced, moderate savings
High (default)Complex reasoning, nuanced analysis, hard codingAnthropic's recommended floor for anything intelligence-sensitive
Xhigh ("Extra")Long-horizon agentic + coding, lots of tool callsClaude Code's default; best coding results
MaxHardest problems where correctness beats costMost thinking, slowest, burns the most quota

What third parties actually published per-effort

1. DataCamp built the only real measurement harness

Their Opus 4.8 effort tutorial runs the same systems-design prompt at low/medium/high and scores each with Claude Haiku 4.5 as a rubric judge (completeness, correctness, edge-case coverage, implementation). The quality baseline:

EffortQuality score (0-100)
Low68
Medium82
High91

Honesty flag: those 68 / 82 / 91 are the tutorial's illustrative default values on one prompt, not a peer-reviewed benchmark. Directionally real (biggest jump is low to medium), but treat them as "one person's rubric on one question." They did not test xhigh or max.

2. Cursor's per-effort statement (from Anthropic's launch)

"On CursorBench, Claude Opus 4.8 exceeds prior Opus models across every effort level... using fewer steps for the same intelligence." Real, but no numbers per level.

3. Anthropic's own token / effort note (the closest thing to a cost benchmark)

At the default high, coding tasks spend roughly the same tokens as Opus 4.7's default but perform better. Xhigh and max deliberately spend more tokens for better results. High is engineered to be the free-lunch tier.

4. Cost facts confirmed across sources

API price is flat $5 / $25 per million tokens regardless of effort (you just spend more tokens at higher effort). Fast Mode dropped to $10 / $50, three times cheaper than on Opus 4.7. On the Max $200 plan none of that is dollars, it's quota burn, and higher effort drains your 5-hour and weekly allowance faster.

The practical rule

Leave it on High for real thinking work. Bump to Xhigh / Extra only for heavy coding sessions. Consciously drop to Medium / Low for routine drafting, triage, and summaries. That's the single biggest lever for not exhausting the $200 plan by mid-week.

What doesn't exist (and what I offered to build)

Nobody has published "Opus 4.8: low vs medium vs high vs xhigh vs max on SWE-bench." I offered to build it: pick 8-10 representative prompts (coding, reasoning, research-synthesis, classification), run each at all five effort levels, capture actual token count + latency + a scored quality rubric per level, and produce a real table showing where quality gains flatten and where you're just burning quota. A couple dollars of API spend, measured on the kind of work you actually do.

Sources