Claude Opus 5: Half the Price, but What About Cost per Task?

원문 링크: WordPress 원문

AI NOTES · EN ENGLISH EDITION

Claude Opus 5 Bets on Efficiency, Not Capability: What Half the Price and a Cost-Capability Toggle Actually Mean

KO · 한국어 / EN · English BILINGUAL PAIR

The question in frontier AI is changing. Choosing a model is no longer only about which one produces the highest benchmark score. Teams increasingly care about how many retries and tool calls it takes to finish a job, whether it can hold context through a long workflow, and how much one verified result actually costs.

Anthropic’s Claude Opus 5, released on July 24, 2026, is built around that shift. The company says it approaches Claude Fable 5’s peak performance in several domains while charging half the API price. It also lets customers vary the model’s effort so they can trade speed and token use against capability on each request. This is less a claim of a dramatic new intelligence ceiling than an attempt to make near-frontier performance economical enough for everyday work.

The central takeaway is straightforward: Opus 5 matters less because it sets a new absolute high and more because it may deliver sufficiently high performance more often, at a more predictable cost. Evaluating that claim, however, requires separating vendor benchmarks from the economics of a real production workflow.

In one sentence: Opus 5 is not a universal replacement for Fable 5; it is closer to a daily frontier-class model designed to reduce how often teams need to invoke the most expensive tier.

1. What actually changed

Opus 5 costs $5 per million input tokens and $25 per million output tokens through the standard API. That is the same list price as Opus 4.8 and half the price of Fable 5. Anthropic has made it the default model for Claude Max and the strongest model available on Claude Pro, signaling that it is intended for routine use rather than occasional escalation.

There is also a Fast mode. Anthropic says it runs at roughly 2.5 times the default speed, but it costs twice the standard rate. Fast mode is therefore not free acceleration; it is a deliberate purchase of lower latency.

The more consequential feature is the effort setting. Customers can vary how much computation and token budget the model applies to a request. Lower effort favors speed and conservation. Higher effort gives complex tasks more room for reasoning, verification, and recovery. The important idea is not simply that users have another parameter. It is that a single model can occupy several points on a cost-performance curve.

Lower effort

Best fit Classification, rewriting, short summaries, routine transformations

Economic logic Reduce latency and token use

Main caution May underperform on long or exception-heavy tasks

Higher effort

Best fit Debugging, complex analysis, multi-step agent work

Economic logic Spend more computation to improve completion odds

Main caution Savings disappear if every request is fixed at maximum effort

Fast mode

Best fit Interactive work where a person is waiting

Economic logic Pay twice the standard rate for lower latency

Main caution Measure the value of response time, not speed alone

Fable 5

Best fit Days-long autonomous work and the hardest high-stakes tasks

Economic logic Purchase the highest available ceiling

Main caution Wasteful if routed routine work does not need it

The table reveals the real product story. Opus 5 is not just a score; it is a way to decide how much intelligence to allocate to each request.

2. Read the benchmarks as cost curves, not a victory chart

The most useful part of Anthropic’s launch material is not one headline number. It is the set of curves showing how performance and cost per task change as effort increases.

Anthropic reports that Opus 5 leads Frontier-Bench v0.1 while more than doubling Opus 4.8’s performance at a lower cost per task. On CursorBench 3.2 at maximum effort, it reportedly comes within 0.5% of Fable 5’s peak score at half the cost per task. On OSWorld 2.0, the company says Opus 5 exceeds Fable 5’s best result at just over one-third of the cost. On AutomationBench, its pass rate is reported to be about 1.5 times the next-best model at the same cost per task, and even the lowest effort setting passes more tasks than competing systems.

The right conclusion is not that Opus 5 wins every evaluation. The more consequential claim is that the attainable success rate at a fixed budget has improved. That distinction matters in agentic coding and tool-using workflows, where one failure can trigger another run, more API calls, and human repair.

Benchmark landscape: Frontier-Bench, SWE-bench, ARC-AGI-3, CursorBench, OSWorld, AutomationBench

Three boundaries are essential when interpreting these results.

First, the major figures come from Anthropic’s own evaluations and launch-day customer reports. Results may change with a different codebase, prompt, tool policy, or network environment. Anthropic notes that some Frontier-Bench results used a specific mini-SWE-agent harness and backend and were averaged over five attempts per task.

Second, a benchmark represents a particular task distribution. Strong CursorBench results do not automatically imply the same gain in legal review or scientific research. Anthropic reports improvements over Opus 4.8 across internal life-science evaluations, including gains of 10.2 percentage points on an organic-chemistry task and 7.7 points on a protein-variant task. Those numbers are useful signals, but they remain distinct from independent validation on a team’s own work.

Third, an average score does not show the cost of failure. In production, a model that checks its assumptions and stops before an irreversible change may be cheaper than a nominally higher-scoring model. Conversely, a model that produces long answers or calls tools unnecessarily may cost more despite a good benchmark result.

A serious evaluation should therefore measure four things:

• First-pass completion rate and retry rate

• Input tokens, output tokens, and tool calls required for completion

• Human correction and review time

• Loss and recovery cost when a run fails

3. Why token efficiency can matter more than list price

A price card gives only the simplest part of the bill. A more realistic model is:

true job cost = input + output + retries + tool calls + review time + failure recovery

Consider a service that handles 10,000 jobs per month. Suppose each job averages 20,000 input tokens and 4,000 output tokens. Monthly volume would be 200 million input tokens and 40 million output tokens.

• Opus 5 standard: $1,000 input + $1,000 output = $2,000 per month

• The same usage on Fable 5 or Opus 5 Fast mode: $2,000 input + $2,000 output = $4,000 per month

Now suppose an organization reproduces an early-customer report in which similar quality is maintained while generated tokens fall by 26%. Output volume would fall from 40 million to 29.6 million tokens, and the Opus 5 bill would be $1,740. Output falls by 26%, but the total bill falls by only 13%, because the input side is unchanged. This is why a claim about token reduction should not be copied directly into a claim about invoice reduction.

The saving can be larger if improved completion quality also cuts retries, tool calls, and human repair. One financial-modeling customer quoted by Anthropic reported nine percentage points higher accuracy, one-third fewer turns and tool calls, and 60% less time across effort levels than with Opus 4.8. Another trading evaluation reportedly used roughly one-seventh of the reasoning tokens with less than half the latency. These are company-curated customer accounts, not universal outcomes. They are best treated as hypotheses to reproduce in an internal evaluation, not as guaranteed ROI.

4. The effort setting is a routing policy, not a checkbox

Many teams buy an expensive model and send every request through the same configuration. That destroys most of the economic value of adjustable effort. A practical deployment should separate work into at least three lanes.

Lane 1: Cheap-to-correct tasks

Classification, formatting, short summaries, and routine first-pass writing are easy to inspect and rerun. A smaller model or Opus 5 at lower effort should usually be the starting point. The primary metric is throughput and cost per accepted result, not maximum quality.

Lane 2: Work where quality and cost interact

Code changes, long-document analysis, and workflows involving several tools can become expensive when the first attempt fails. Opus 5 at medium or high effort is a plausible default here. The workflow should still define completion gates such as passing tests, verifying citations, or reconciling a generated table against source data.

Lane 3: High-consequence failure

Large migrations, long-running autonomous agents, and major security, legal, financial, or scientific decisions need more than a high effort value. They may require Fable 5, an independent reviewer model, and explicit human approval.

Effort toggle: low, medium, high — a per-request cost-capability tradeoff as product primitive

A team does not need a sophisticated AI router on day one. A simple escalation policy is enough to start:

• Route low-risk, short work to a small model or low effort.

• Escalate when the input is long, several tools are required, a test fails, or confidence is low.

• After repeated failure—or before an irreversible change—require a stronger model or human review.

• Record the cost of a completed and accepted job , not just tokens by model.

This reframes the endless question of which model is best into a measurable one: which path completes our work most reliably at the lowest total cost?

5. Where the difference may be most visible

The recurring themes in Anthropic’s announcement are not merely answer quality. They are verification, recovery, and long-horizon execution. The company describes cases in which Opus 5 created a missing computer-vision pipeline, found the root cause of a package-manager bug instead of patching the symptom, and built a test harness when no live market feed was available.

Coding and agent workflows

In a large codebase, the first answer matters less than whether the model understands the change surface, runs tests, and tracks down a failure. If Opus 5 genuinely reduces retries and supervision, the economics of coding agents change. Avoiding one confident but incorrect “done” can be worth far more than a modest reduction in tokens.

Knowledge and document work

Legal redlines, financial models, and due-diligence analysis depend on source fidelity and format. Early users quoted by Anthropic reported stronger first-pass work, numerical reasoning, and comparable quality at lower reasoning levels. Because these claims come from private internal evaluations, organizations should reproduce them with their own contract templates, spreadsheets, and acceptance rubrics.

Scientific research

Anthropic says Opus 5 improves on Opus 4.8 in structural biology, organic chemistry, and bioinformatics evaluations. But confidence and evidence remain different things. Citations still need to be verified against original papers; quantitative claims need to be reproduced from source data; and statistical outputs need independent checks. Opus 5 may be valuable for analysis assistance and automating verification steps, but it does not replace experimental evidence or expert judgment.

6. Half the price is not always half the cost

The standard Opus 5 token rates are unambiguously half those of Fable 5. The most economical choice still depends on workload shape.

• Output share: Output is priced at five times the input rate, so verbose generation can dominate the bill.

• Cache behavior: Services with long system prompts and repeated context benefit when tool design preserves the prompt cache. Anthropic’s beta for changing tools mid-conversation without invalidating that cache is operationally significant.

• Value of latency: Fast mode can make sense when a person is waiting for a result. It is harder to justify for overnight batch work where latency has little economic value.

• Retry rate: A cheap model that needs three attempts may cost more than an expensive one that finishes in one.

• Safety routing: Automatic Fallbacks can turn classifier refusals into usable responses by routing to another model. Teams still need observability so they know which model answered and what it cost.

Cost optimization is therefore a system design problem involving caching, effort, tools, fallbacks, and verification—not merely a model-selection decision.

7. Competition is moving from a single best model to a model portfolio

Ars Technica interprets Opus 5 primarily as a cost story rather than a radical capability leap. Its coverage also points to price pressure from open-weight models and to routers from companies such as Cursor and Meta that select models based on the request. CNBC similarly places the launch in a market where enterprise buyers are less willing to experiment with expensive AI without a clear return on investment.

As teams mix small models, open models, and frontier systems, the unit of competition changes. Three factors become more important than a single leaderboard rank:

• The amount of accepted work completed for a fixed budget

• The quality of routing, caching, and developer controls

• The predictability of failures and safety behavior

Opus 5 is well positioned on that axis. It combines better performance at the Opus 4.8 price with effort controls, Fast mode, automatic fallback, and mid-conversation tool changes. The launch is therefore as much about economic primitives for building services as it is about model intelligence.

Competitive landscape: model routers and open weights — efficiency per dollar is the new battleground

8. Clear limitations remain

Opus 5 is not Anthropic’s strongest system in every area. The company explicitly says it remains behind Mythos 5 in offensive cybersecurity and long-running autonomous biology research. Anthropic intentionally limited cyber training; the model has improved at finding vulnerabilities but remains substantially weaker at converting them into working exploits.

“Near Fable performance” also does not mean equivalence on every task. Days-long autonomous projects and work with extreme failure costs may still warrant a higher-tier model and direct supervision. At the other end of the spectrum, Opus 5 may be unnecessary for routine classification or simple summarization.

The largest evidentiary limitation is that much of the launch case is vendor-led. Until independent benchmarks, production invoices, and organization-specific quality evaluations accumulate, “similar performance at half the price” should be treated as a compelling operational hypothesis, not a universal fact.

9. What a team can do this week

Instead of switching everything at once, assemble a small evaluation of 30 to 100 representative jobs.

• Compare the current model and Opus 5 under the same prompts, tools, and stopping conditions.

• Test lower, medium, and higher effort separately; record both quality and cost.

• Track retries, tool calls, human correction time, and test pass rates in addition to token use.

• Segment results by task difficulty rather than averaging easy and hard jobs together.

• Keep separate review gates for privacy, security, legal, financial, and scientific outputs.

• After two weeks, select the route with the lowest cost per accepted job—not the model with the most impressive headline.

The most useful question raised by Opus 5 is not where it ranks. It is this: Are we reserving expensive intelligence for the moments that actually require it? Teams that can answer that question with their own data may turn the lower price and effort control into genuine savings. Teams that cannot may simply convert another efficiency feature into a more expensive default.

Sources

• Anthropic — Introducing Claude Opus 5

• Ars Technica — Anthropic’s Opus 5 is about token efficiency, not a capability leap

• ZDNET — Claude Opus 5 arrives with near Fable performance at half the price

• CNBC — Anthropic’s new AI model rivals Fable 5 and is cheaper

• TechCrunch — Anthropic launches Opus 5

Updated July 28, 2026. Prices and product availability may change. Benchmark claims and customer examples are attributed to Anthropic unless otherwise noted.