Claude Opus 5's Effort Parameter: How to Cut API Costs Without Losing Quality

Anthropic shipped Claude Opus 5 on July 24, 2026, and the headline feature isn't a benchmark score — it's a dial. For the first time, Opus ships with five explicit effort settings (low, medium, high, xhigh, max) that let you trade reasoning depth for latency and token spend on every single request. If you're building on the Claude API, this changes how you should be pricing, routing, and testing your workloads — and most teams haven't touched their defaults yet.

This matters more than a typical model bump. Anthropic is explicitly telling developers to stop treating "which model" as the only cost lever and start treating "how hard should it think" as a separate one. Get that wrong and you're either overpaying for simple classification calls or underpowering your hardest agentic tasks.

What Actually Shipped on July 24

Claude Opus 5 comes close to Claude Fable 5's capability on most benchmarks while costing half as much per token — $5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.8's pricing. Fable 5 remains Anthropic's flagship at $10/$50 per million tokens.

The specs worth knowing before you migrate:

• Context window: 1M tokens — this is both the default and the maximum. There's no smaller-context variant to fall back to.
• Max output: 128K tokens
• Extended thinking: on by default
• Batch API: half price ($2.50/$12.50 per million tokens)
• Prompt cache hits: $0.50 per million tokens
• Fast mode: roughly 2.5x the default speed at $10/$50 per million tokens

According to The Decoder's reporting, Opus 5 landed within 0.5% of Fable 5's peak CursorBench 3.2 score at max effort, and actually beat Fable 5's best OSWorld 2 result — a strong signal for teams running computer-use or coding agents who've been paying Fable 5 prices out of necessity.

The Five-Tier Effort System, Explained

Here's the part that's genuinely new. Instead of one Opus behavior, you now pick a point on a ladder:

Per Anthropic's own guidance, the low and medium settings aren't a compromise tier — the company says they deliver good results at a fraction of the token use and latency of earlier Opus versions, and they outperform equivalent settings on Opus 4.8. That's a direct invitation to move workloads that used to require a full Opus call down to low or medium instead of downgrading to Sonnet.

For coding and agentic work specifically, Anthropic still recommends starting at xhigh — not high, and not max. That's a meaningful signal if you're building autonomous coding agents: max exists for edge cases, not as the default ceiling.

Why This Beats Just Switching Models

Before this release, the standard cost-optimization move was model selection: use Haiku for cheap tasks, Sonnet for the middle tier, Opus for the hard stuff. The effort parameter adds a second axis within Opus itself, which means you can now:

Keep one model family in your codebase instead of maintaining routing logic across Haiku, Sonnet, and Opus for tasks of varying difficulty.

A/B test effort tiers on the same prompt to find the cheapest setting that still hits your quality bar, rather than guessing which model tier is "enough."

Route dynamically at runtime — e.g., a support bot that starts every ticket at low, and only escalates to high or xhigh if the initial response scores low on a confidence check.

If you're already running a Claude-based agent or chatbot, this is worth an afternoon of testing. Take your current production prompts, run them at low and medium, and measure quality against your existing high-effort (or model-default) baseline. Anthropic's public claim is that a meaningful share of "Opus-required" workloads will pass at medium — but that's task-dependent, and the only way to know is to test your actual prompts.

A Practical Cost Example

Say you're running a document-summarization pipeline processing 10 million input tokens and generating 2 million output tokens per day at Opus 5 pricing.

At default (high) effort, assuming no batching or caching, that's roughly $50 (input) + $50 (output) = $100/day, or about $3,000/month. If testing shows medium effort holds quality for your summarization task, and Anthropic's guidance about medium using "a fraction" of the tokens holds — even a conservative 30-40% token reduction on the reasoning overhead — you're looking at meaningfully lower spend without a model downgrade. Stack that with Batch API (50% off) for non-realtime jobs, and prompt caching for repeated system prompts, and the same workload can drop well below $1,500/month.

The math changes per use case, but the principle holds: effort tuning is now a bigger lever than most teams' existing cost-optimization work.

How to Get Started

If you're on the Claude API, the effort parameter is available now for Opus 5 requests. A few concrete steps:

• Audit your current Opus traffic. Pull a sample of real production prompts and re-run them at low and medium. Compare outputs against your existing quality bar — not against a benchmark, against what your users actually need.
• Segment by task type. Classification, tagging, and short-form extraction are strong low/medium candidates. Multi-step agentic work and code generation should stay at xhigh per Anthropic's own recommendation.
• Don't default to max. It's built for genuinely hard reasoning tasks — deep research, complex proofs, architecture decisions with many interacting constraints — not as a "just to be safe" setting. Anthropic's benchmark data suggests xhigh already captures most of the coding and agentic performance ceiling.
• Re-test after any prompt change. Effort/quality tradeoffs are prompt-specific. A prompt rewrite can shift which tier is "good enough."

Key Takeaways

• Claude Opus 5 launched July 24, 2026 with five effort tiers (low, medium, high, xhigh, max) controlling reasoning depth per request.
• Pricing is unchanged from Opus 4.8: $5/$25 per million input/output tokens — half of Fable 5's $10/$50.
• Opus 5 comes within 0.5% of Fable 5 on CursorBench 3.2 at max effort, and beats it on OSWorld 2, while costing half as much.
• Anthropic recommends low/medium for most traffic and xhigh (not max) as the starting point for coding and agentic tasks.
• The effort parameter is a cost lever independent of model choice — test it on your actual prompts before assuming you need a smaller model.

Next Steps

If you're preparing for the Claude Certified Architect (CCA) exam, effort-parameter tuning and cost-optimization strategy are exactly the kind of applied API knowledge the certification tests. AI for Anything's CCA practice test bank covers Claude API pricing models, agent architecture, and real exam-style scenarios so you walk in ready — not guessing.

Sources:

• Introducing Claude Opus 5 — Anthropic
• Anthropic claims its new Claude Opus 5 delivers near-Fable 5 performance at half the token price — The Decoder
• Anthropic debuts Claude Opus 5 with feature that lets users toggle between cost and capability — Fortune
• Anthropic launches Claude Opus 5, a cheaper AI model for coding, agents and enterprise workflows — VentureBeat