Claude Opus 5 Is an Everyday Model—With Limits

The short version: Anthropic released Claude Opus 5 on July 24 as its new default model for Claude Max and its strongest model on Claude Pro. The launch benchmarks make a strong case for Opus 5 as the new default for difficult work. They do not make it the right model for every prompt—and independent code-review results argue against using it as the only reviewer.

The useful conclusion is more specific: Opus 5 can be a daily driver for high-value, ambiguous work. Sonnet 5 should still handle routine volume. Fable 5 should be an escalation, not a reflex.

What did Anthropic release with Claude Opus 5?

Opus 5 keeps the previous Opus price while moving much closer to Anthropic’s more expensive Fable tier. It is available across Claude products and through the API as claude-opus-5.

Model	Standard API price per 1M tokens	Best fit	Practical role
Claude Sonnet 5	$2 input / $10 output through Aug. 31; then $3 / $15	Routine coding, chat, extraction, and scaled agents	Volume default
Claude Opus 5	$5 input / $25 output	Difficult coding, research, documents, and multi-step agents	High-value daily driver
Claude Fable 5	$10 input / $50 output	The longest, most open-ended work where internal evals prove an advantage	Escalation tier

Opus 5 also has a one-million-token context window, five effort settings from low to max, and a Fast mode that runs about 2.5 times faster for twice the normal token price. Thinking is enabled by default.

That combination matters more than the family name. Teams are no longer choosing one fixed intelligence level; they are selecting a point on a quality, latency, and cost curve for each task.

Which Claude Opus 5 benchmarks actually matter?

Anthropic’s headline results are unusually strong:

Evaluation	Opus 5 result	What it suggests	Important caveat
Frontier-Bench v0.1	43.3%	Strong end-to-end agentic coding	Anthropic’s internal run used its selected harness and five attempts per task
ARC-AGI-3	30.2%	Large gain on novel problem-solving	High-effort result; Fable 5 was not included in the reported comparison
GDPval-AA v2	1,861 Elo	Strong professional knowledge work	A benchmark is still a proxy for a team’s documents, tools, and acceptance criteria
SWE-bench Pro	79.2%	Frontier-level repository issue resolution	Slightly behind Mythos 5 and Fable 5 in the reported table
AA-Briefcase, max	1,720 Elo	Leading agentic document and analysis work	Independent but proprietary benchmark; max effort averaged more than 36 minutes per task

The most useful independent signal comes from Artificial Analysis. It tested all five effort settings on AA-Briefcase, which requires agents to work across thousands of private input files and produce reports, presentations, and spreadsheets. Opus 5 at max effort scored 1,720 Elo versus Fable 5’s 1,574, while costing $17.79 per task instead of $22.30. At high effort, it still edged Fable 5 while costing $10.41 per task—less than half as much.

The trade-off is time. High, x-high, and max effort averaged roughly 26, 34, and 36 minutes per task. This is not evidence that max effort should become a universal default. It shows that Opus 5 can buy more quality when the task is worth waiting for.

The broader comparison reported by Decrypt reaches the same headline conclusion: Opus 5 beats or approaches Fable 5 on many published evaluations for half the per-token price. The better buying question is not “Which model won the chart?” It is “At which effort level does the quality improvement exceed the extra cost and waiting time on our work?”

Is Claude Opus 5 good for code review?

Yes—as a specialist, not as the only safety net.

CodeRabbit tested Opus 5 across about 100 verified error patterns from real open-source pull requests, with three runs per configuration. At x-high effort, Opus 5 produced more precise actionable comments than CodeRabbit’s production baseline: 39.3% versus 35.2%. But it caught fewer known issues: 55.2% versus 61.1%.

It also produced 92 nitpicks versus 23 for the baseline, and its precision across the complete post-pipeline comment stream fell to 28.6% versus 32.8%. More effort did not improve every metric.

The category breakdown is equally important. CodeRabbit found promising results on configuration errors and code quality, but weaker performance on logic errors, race conditions, and API misuse. Its conclusion is narrower than Anthropic’s general code-review claim—and more actionable:

• Use Opus 5 for a second architectural or maintainability pass.
• Keep tests, static analysis, and a recall-oriented reviewer for correctness-heavy changes.
• Do not assume x-high effort is automatically safer.
• Ask for broad findings, then filter separately; a conservative prompt can suppress recall.

There is another cost signal. In CodeRabbit’s runs, Opus 5 averaged about 60,500 input and 9,500 output tokens per review call, compared with about 40,500 input and 5,800 output tokens for its GPT-5.6-based baseline. A higher-quality comment subset can still be a poor bargain if developers must sift through a much longer review.

Should Claude Opus 5 be your daily driver?

For a Claude Max subscriber doing complex knowledge work or software development: probably yes. For an API routing every task through one model: no.

Use Opus 5 as the default when a wrong conclusion is expensive and the work benefits from sustained judgment:

• ambiguous feature design and multi-file implementation;
• root-cause analysis across a large codebase;
• due diligence, financial analysis, and complex document synthesis;
• spreadsheet, presentation, and research deliverables that require iterative verification;
• long-running agents that must coordinate tools or independent workstreams.

Start with Sonnet 5 when the task is frequent, bounded, and easy to verify:

• routine chat and drafting;
• small code edits with clear tests;
• extraction, classification, routing, and summarization;
• high-volume automations where latency and unit cost dominate;
• first-pass work that can be escalated only when confidence is low.

Reserve Fable 5 for tasks where your own evaluation shows a real advantage: very long orchestration, difficult investigative reasoning, or work that repeatedly fails under Opus. One early developer report on Reddit described sound code but hasty conclusions during debugging, and preferred Fable for investigation. That is an anecdote, not a benchmark, but it is a useful reminder that clean code and correct diagnosis are different outcomes.

Which effort setting should you use?

Treat effort as routing, not a quality slider that only moves upward.

Workload	Starting setting	Why
Routine questions, triage, and narrow edits	Low or medium	Lower latency and token use; Anthropic says quality remains strong on many tasks
A defined pull request or multi-file feature	Medium or high	Enough depth without automatically paying the x-high tax
Architecture, difficult debugging, or open-ended agent work	X-high	Anthropic’s recommended starting point for demanding coding and agentic work
A rare, high-value task with measurable acceptance criteria	Max	Use only when an effort sweep proves the extra time and cost improve outcomes
Latency-sensitive work with high value per minute	Fast mode	About 2.5× faster, but at twice the base token price

Anthropic’s Opus 5 prompting guide recommends giving complex coding tasks a complete specification up front. It also warns that the model already verifies and self-corrects: legacy instructions to “double-check everything” or spawn a verifier can multiply tokens without improving the result.

Two prompt changes are immediately useful:

1. Specify the desired output length. Lower effort reduces thinking, but does not reliably make the visible answer shorter.
2. Constrain scope and subagent use. Opus 5 can expand tasks and delegate more readily than earlier models, which raises both cost and elapsed time.

What is the real Opus 5 insight?

Aaron Levie described the release on X as a large jump over Opus 4.8 across many capabilities. Box’s results, quoted by Anthropic, put numbers behind that reaction: an 8% overall improvement, with 11% on data analysis and 17% on due-diligence workflows.

The deeper signal is not that Fable has become irrelevant. It is that model tiers are turning into routing policies. Opus 5 at high effort can beat a more expensive frontier model on one workflow; Sonnet 5 can be the better economic choice on another; a second reviewer can outperform either model reviewing its own work alone.

So, is Opus 5 the daily driver? Yes—if “daily driver” means the model you trust with difficult work. No—if it means sending every cheap, repetitive, latency-sensitive task to the same expensive endpoint.

The winning setup is a ladder: Sonnet by default, Opus for consequential ambiguity, Fable by measured exception, and independent verification wherever failure matters. Browse our AI model comparisons and Claude Code coverage for the next round of evidence.