Claude Opus 5 Medium Just Beat Sonnet 5 Max—at Half the Cost

The usual model-routing rule is simple: start with the cheaper model, then escalate difficult work to Opus. Claude Opus 5 makes that rule look dated.

For serious agent work, the better default may now be the stronger model at a lower effort setting. You pay more per token, but the agent can need far fewer tokens, tool calls and retries to finish the job.

That is not a theoretical argument. On CursorBench 3.2, Opus 5 at medium effort scored 64.3% at $3.29 per task. Sonnet 5 at max effort scored 61.5% at $6.45. Opus used 23,612 tokens and 44 steps; Sonnet used 92,882 tokens and 86 steps.

In that test, Opus 5 medium was slightly better, used about 75% fewer tokens, and cost 49% less.

Source: CursorBench 3.2

What about Sonnet 5 at high and xhigh?

The same benchmark makes the answer unusually clear. Opus 5 medium did not merely edge out Sonnet max:

• Against Sonnet 5 high, Opus medium scored 7.4 percentage points higher for just $0.10 more per task. It used 40% fewer tokens and 13 fewer steps.
• Against Sonnet 5 xhigh, Opus medium scored 5.6 points higher, cost 21% less, used 55% fewer tokens and took 23 fewer steps.
• Against Sonnet 5 max, Opus medium scored 2.8 points higher, cost 49% less, used 75% fewer tokens and took 42 fewer steps.

So high is the only one of those Sonnet settings that was marginally cheaper per task—and it produced the lowest score. Xhigh and max both cost more while scoring lower. This is one benchmark, not a universal law, but it strongly supports Opus medium as the safer default for substantial agent work.

Chart recreated from the public CursorBench 3.2 results. Scores and average task costs are shown exactly as reported.

The important distinction: token price is not task price

Sonnet 5 is still cheaper on the rate card. Through August 31, 2026, it costs $2 per million input tokens and $10 per million output tokens. Its standard price from September 1 is $3/$15. Opus 5 costs $5/$25.

If both models use exactly the same tokens, Sonnet wins on price. Opus costs 2.5 times as much as promotional Sonnet and 1.67 times as much as standard-price Sonnet.

Agents do not buy equal token bundles, however. They buy completed tasks. A weaker run can explore more branches, make more tool calls, repeat checks, miss the root cause and need another attempt. CursorBench shows how quickly the cheaper-per-token model can become the more expensive worker.

The comparison becomes even less favorable to max-effort Sonnet when its introductory pricing ends. Holding the same benchmark usage constant, its $6.45 task would rise to about $9.68. That would make Opus 5 medium roughly 66% cheaper in this specific test.

Source: Anthropic API pricing

The result is not limited to one coding score

Artificial Analysis tests models on GDPval-AA v2: 220 realistic professional tasks involving documents, spreadsheets, diagrams, and research across 44 occupations.

There, Opus 5 medium scored 1,632 Elo, while Sonnet 5 max scored 1,603. The gap is modest and the reported uncertainty intervals overlap, so it should not be sold as a decisive win. It does show that the Cursor result is not an isolated story about raw coding: medium-effort Opus can reach the same quality band as max-effort Sonnet on real knowledge work.

Anthropic's own launch evaluations point in the same direction. The company reports state-of-the-art results for Opus 5 on Frontier-Bench, a leading result on Zapier AutomationBench even at low effort, and stronger computer-use performance than Fable 5 at just over one-third of the cost. These are vendor-reported results, but they cover coding, browser work, business automation and long-horizon agents rather than one narrow benchmark.

Sources: GDPval-AA v2, Anthropic's Opus 5 launch

Anthropic's GDPval-AA v2 launch chart plots the Opus and Fable effort ladders against cost. It does not include Sonnet 5; the Opus-versus-Sonnet figures above come from Artificial Analysis's current leaderboard. Source: Anthropic.

Why Opus 5 changes the view of agents

Earlier agents often compensated for model limits with scaffolding: detailed plans, verifier agents, repeated retries and tightly scripted tool sequences. Opus 5 is better at finding root causes, verifying its own work and continuing until the task is actually complete. Anthropic also says it delegates to subagents more readily and warns that old prompts demanding extra verification can now cause over-verification.

This matters because orchestration is not free. Every retry adds tokens, latency and another chance for the workflow to drift. A model that needs fewer corrective loops can be cheaper even when each token costs more.

Opus 5 also gives teams one practical control surface: low, medium, high, xhigh and max effort. Effort changes the whole response, including thinking, tool calls and function arguments. It is not a strict token budget, but it lets one model cover a wide operating range.

Sources: What's new in Opus 5, Anthropic effort documentation

Opus 5 versus Fable 5: the premium tier has become harder to justify

Fable 5 remains Anthropic's most capable widely released model. It also costs exactly twice as much as Opus 5: $10/$50 versus $5/$25 per million input/output tokens.

CursorBench makes the trade-off unusually clear:

• Medium: Opus 5 scored 64.3% at $3.29; Fable 5 scored 65.2% at $6.80.
• High: Opus 5 scored 66.7% at $3.91; Fable 5 scored 66.5% at $8.77.
• Max: Opus 5 scored 70.0% at $8.23; Fable 5 scored 70.5% at $17.32.

Anthropic's CursorBench launch chart visualizes score versus task cost across effort levels. Source: Anthropic; underlying benchmark: CursorBench 3.2.

Across those settings, Opus delivered virtually the same result for roughly half the task cost. Fable may still earn its place when your own evaluation shows that the final fraction of capability changes the outcome. It also has different safeguards and a mandatory 30-day retention policy, while general Opus 5 access has no model-specific retention requirement.

The evidence does not support calling Opus the default for literally every workload. Sonnet and Haiku still make sense for simple, high-volume or latency-sensitive requests. Mythos remains stronger for some high-risk cyber and biological research. But for general-purpose agents doing valuable work, Fable now looks like an exception path rather than the automatic top-tier choice.

Sources: CursorBench 3.2, Claude model overview, Fable 5 documentation

A practical routing policy

A sensible starting policy for agentic work is now:

• Opus 5 medium: the everyday default for bounded coding, research and business workflows where the result matters.
• Opus 5 high: ambiguous tasks, deeper reasoning and work that must survive several tool calls.
• Opus 5 xhigh: long-running coding and high-autonomy agents. This is Anthropic's recommended starting point for demanding agentic work.
• Opus 5 max: reserve for frontier problems where evaluation data proves that unconstrained effort pays back.
• Sonnet 5: high-volume, faster and simpler workflows where your tests show that lower per-token pricing survives at the task level.
• Fable 5: only when its measurable quality edge is worth roughly double the price, or when a workload specifically requires it.

Do not change effort blindly in the middle of a cache-heavy conversation: Anthropic notes that changing effort invalidates cached prefixes. Route by workload, hold effort steady inside the session, and measure cost per accepted result rather than cost per million tokens.

The bottom line

Opus 5 does not make Sonnet 5 obsolete. It does make Sonnet 5 at high, xhigh and max difficult defaults to defend for demanding agents. On CursorBench, Opus medium beat all three: high was only $0.10 cheaper but scored 7.4 points lower, while xhigh and max both cost more and still scored lower.

The sharper operating model is: start valuable agent work on Opus 5 medium. Raise Opus effort when your evaluations show that complexity justifies it. Route down to Sonnet or Haiku when speed, volume, or latency matters more—and only when your own measurements show that the completed task is genuinely cheaper.

This is one benchmark, not a law for every workload. But it changes the default question. Do not ask which model sells the cheapest tokens. Ask: What is the lowest effort on the strongest model that reliably finishes the job?