Claude Opus 5: What Actually Changes for Developers and Agents

I recently wrote about how I divide work among GPT-5.6 Sol, Terra, and Luna: one plans, another executes, and the third handles repetitive tasks. Opus 5 takes a different approach: sustaining long-running work, using tools, and verifying its own results within a single session.

I still have not used it enough to publish a verdict. This article brings together what has been confirmed about the launch, compares it with independent testing, and explains where it might make sense in a real development workflow.

---

Anthropic introduced Claude Opus 5 on July 24, 2026, just two months after Opus 4.8. It describes the model as being built for agentic coding, long-running tasks, and complex professional work.

That aligns with a concrete need in my current workflow. Separating planning, execution, and repetitive work helps control costs, but it introduces handoffs: one model designs the plan, another receives the context and executes it, and then I have to check that both interpreted the same intent.

Opus 5 interests me because it could reduce that coordination cost on large tasks. Not as an automatic replacement for Sol, Terra, and Luna, but as an option for work where preserving continuity may be more valuable than separating every role.

The Comparison I Care About

Today, I divide the work roughly like this:

Type of work	Current workflow	Potential role for Opus 5
Architecture and review	Sol	I would keep a specialized model
Scoped implementation	Terra	Opus 5 may be overkill
Repetitive work	Luna	Opus 5 would probably be overkill
Long, multi-file feature	Sol plans → Terra executes	The most interesting space for Opus 5

I do not see this launch as a head-to-head competition in which one model must replace the other three.

The Sol, Terra, and Luna workflow optimizes for specialization and cost. Opus 5 emphasizes continuity and autonomy: taking in more context, working for longer, and completing a deliverable with fewer handoffs.

If I test it, it will be when a feature is large enough to justify that approach. For now, this comparison helps explain what changes and where it could add value.

What Claude Opus 5 Brings

The official specifications are:

Feature	Claude Opus 5
API ID	claude-opus-5
Context	1 million tokens
Maximum output	128K tokens
Thinking	Adaptive and enabled by default
Effort	low, medium, high, xhigh, max
Input price	USD 5 per million tokens
Output price	USD 25 per million tokens
Fast mode	Approximately 2.5× faster, USD 10 / 50
Batch API	50% discount: USD 2.50 / 12.50
Cache hit	USD 0.50 per million tokens

The one-million-token context window is both the default and the maximum. There is no smaller variant. The model also enables adaptive reasoning without requiring thinking to be declared in every request.

The Most Important New Feature Is effort

Opus 5 responds better than earlier versions when inference compute is increased. Anthropic has made effort the main control for deciding how much work the model should invest:

low      → simple tasks, lower cost and latency
medium   → everyday work where quality holds up
high     → default value
xhigh    → demanding coding and agent tasks
max      → tasks where achieving the best possible result matters

This is not a response-length selector. effort controls how much the model reasons, not how much visible text it writes. Anthropic's documentation warns that Opus 5 tends to produce longer responses and documents than Opus 4.8, so concision must be requested explicitly.

A basic TypeScript call might look like this:

import Anthropic from '@anthropic-ai/sdk';

const client = new Anthropic();

const response = await client.messages.create({
  model: 'claude-opus-5',
  max_tokens: 8_192,
  output_config: {
    effort: 'high',
  },
  messages: [
    {
      role: 'user',
      content: 'Review this change, find the root cause, and propose a fix.',
    },
  ],
});

console.log(response.content);

For xhigh or max, Anthropic recommends increasing max_tokens and using streaming. The output limit includes both reasoning and the visible response.

Thinking Is Enabled by Default

This change may affect existing integrations.

In Opus 4.8, a request without a thinking configuration ran without extended reasoning. In Opus 5, the same request enables thinking adaptively.

In addition:

• Thinking can only be disabled with high effort or lower.
• Attempting to disable it with xhigh or max returns HTTP 400.
• With thinking disabled, the model may occasionally write tool calls as text or leak internal tags.

The practical recommendation is to keep thinking enabled and lower effort when you want to reduce token use or latency.

Official Performance: Strong in Agents, Coding, and Professional Work

Anthropic presents Opus 5 as its new state of the art on Frontier-Bench and GDPval-AA. The official summary reports:

Evaluation	Opus 5	Fable 5	Opus 4.8	GPT-5.6 Sol
Frontier-Bench v0.1	43.3%	33.7%	21.1%	34.4%
GDPval-AA v2	1861	1747	1593	1736
ARC-AGI-3	30.2%	—	1.5%	7.8%
OSWorld 2.0	70.6%	66.1%	55.7%	62.6%
AutomationBench	26.0%	17.4%	17.0%	18.1%

Table published by Anthropic. The results should be interpreted as vendor benchmarks, not as a guarantee for every workload.

On Frontier-Bench, which measures agentic coding in a terminal, Opus 5 more than doubles Opus 4.8's maximum score and outperforms Fable 5 at a lower cost per attempt.

Frontier-Bench v0.1 results published by Anthropic. Each point represents an effort level.

The most interesting data point is not just the maximum. The curve shows that the effort level substantially changes the tradeoff between cost and results. There is no universally correct configuration.

The First Independent Results Are Good, but Not Uniform

Artificial Analysis evaluated the model before launch. In its Intelligence Index, Opus 5 at maximum effort scores 61 points, virtually tied with Fable 5 at 60 and ahead of GPT-5.6 Sol at 59.

It also reports:

• First place on GDPval-AA v2 with an Elo score of 1861.
• Joint leadership in its Coding Agent Index.
• An average cost of USD 2.03 per task in the Intelligence Index.
• A 26% lower cost per task than Fable 5 in that evaluation.
• Less factual knowledge than Fable 5.
• A 50% hallucination rate on AA-Omniscience, 14 points higher than Opus 4.8.

Artificial Analysis results. The arrows identify Claude Opus 5 configurations at different effort levels.

This supports a more useful reading than “it is the best model”: Opus 5 appears especially competitive when it can act, use tools, and complete deliverables. It does not necessarily dominate at pure factual recall.

More Convincing as a Builder Than as the Sole Reviewer

CodeRabbit published an independent evaluation focused on code review. Its conclusion is more nuanced.

At xhigh effort, Opus 5 produced actionable comments with greater precision than its production baseline:

Actionable precision
Baseline    35.2%
Opus 5      39.3%

But it detected fewer known issues:

Issues detected
Baseline    61.1%
Opus 5      55.2%

It also generated four times as many nitpicks, and its overall precision fell when those comments entered the workflow.

CodeRabbit evaluation: Opus 5 x-high versus its production baseline.

Its final assessment is clear: Opus 5 works better as a builder for ambiguous, design-heavy, and long-running tasks than as the sole reviewer responsible for catching every error.

For code review, it could serve in a high-precision path alongside another model or process focused on coverage.

Practical Changes to Prompt Writing

Migrating prompts from Opus 4.8 without reviewing them can waste tokens.

1. Remove Redundant Verification Instructions

Opus 5 verifies its own work without being asked. Phrases such as:

Double-check every result.
Use another agent to verify your answer.
Include a final verification step.

can lead to over-verification.

2. Control Length Explicitly

Lowering effort does not guarantee a shorter visible response.

A useful instruction is:

Keep responses focused and concise.
Explain at a high level unless depth is requested.

3. Limit Subagent Use

Opus 5 delegates more readily. That helps with genuinely independent tracks, but it multiplies cost and context on small tasks.

Define when it may delegate and set a limit.

4. Provide the Complete Specification

For large features, refactors, and multi-file work, Anthropic recommends explaining the full objective and letting the model execute. Opus 5 is designed to finish the task, not leave stubs or placeholders.

Platform Updates Accompanying the Model

The launch also includes useful changes for agents:

Tools That Can Change During a Conversation

In beta, an application can add or remove tools between turns without invalidating the prompt cache.

This allows an agent to start with a minimal toolset and receive additional capabilities as the task progresses.

Automatic Fallback

The API can automatically redirect a request blocked by safety classifiers to another model.

Instead of returning a refusal or an error, it attempts to obtain a useful response from a compatible model.

Cache for Shorter Prompts

The minimum cacheable length drops from 1024 to 512 tokens. Systems with medium-length prompts can take advantage of caching without changing their architecture.

Fast Mode

Opus 5 can run approximately 2.5 times faster for twice the base price. It is available through the Claude API, but not on Bedrock, Vertex AI, or Microsoft Foundry.

Safety and Data Retention

Anthropic states that Opus 5 does not expand the frontier of risky dual-use capabilities.

The model:

• Can search source code for vulnerabilities.
• Blocks binary scanning intended to find vulnerabilities.
• Blocks penetration testing and exploit generation outside verified programs.
• Remains behind Mythos 5 in offensive exploitation and autonomous biological research.

According to Anthropic, its cybersecurity classifiers should intervene around 85% less often than with Fable 5.

Opus 5 also does not have the 30-day data retention requirement applied to Fable 5 and Mythos 5 for general access.

Where It Could Fit in My Workflow

Based on the announcement and the first independent evaluations, the use cases that interest me most are:

• Implementing multi-file features.
• Refactors that require planning and verification.
• Agents with tools and long sessions.
• Technical research that produces deliverables.
• Spreadsheets, slides, and complex documents.
• Code review as a second, precision-oriented layer.

I would not assume it is automatically the best option for:

• Small tasks where Sonnet or a fast model can solve the problem at lower cost.
• Factual retrieval, where Fable 5 retains an advantage.
• Code review as the only safety barrier.
• Penetration testing or offensive research.
• Workflows that cannot tolerate lengthy responses without prompt adjustments.

I do not have a formal benchmark or test suite prepared. I would use it when a task large enough to justify it appears, then compare the result with the workflow I already know.

A Reasonable Starting Configuration

An initial configuration might be:

{
  "model": "claude-opus-5",
  "max_tokens": 8192,
  "output_config": {
    "effort": "medium"
  }
}

I would use medium for everyday work, move up to high or xhigh for large features, and reserve max for tasks with an evaluation that justifies the cost.

The official starting point is high, but there is no reason to pay for that level on every request if your internal evaluations show that medium maintains quality.

Conclusion

Claude Opus 5 does not automatically replace Sol, Terra, and Luna simply because it is the latest release.

I recently published that workflow precisely because separating responsibilities helps match the right cost and capability to each task. Opus 5 introduces a different proposition: concentrating more planning and execution within a single long session.

The official benchmarks are strong, and the first independent evaluations confirm that there is a real improvement. They also reveal limitations: higher consumption, longer responses, lower coverage on some review tasks, and factual knowledge below Fable 5.

For now, I do not have a verdict. I have a candidate for a specific kind of work:

Large features where preserving continuity may be worth more than dividing the work among several models.

The question is not which model wins. It is when the continuity of Opus 5 is worth more than the specialization of Sol, Terra, and Luna.

Sources

• Introducing Claude Opus 5 — Anthropic
• What's new in Claude Opus 5 — Claude Platform
• Claude Opus pricing — Anthropic
• Prompting Claude Opus 5 — Anthropic
• Claude Opus 5 System Card
• Artificial Analysis: Opus 5
• CodeRabbit: Opus 5 model review
• Anthropic launches Opus 5 — TechCrunch
• Anthropic announces Claude Opus 5 — CNBC