Preface¶

On July 24, 2026, Anthropic officially launched Claude Opus 5 and made it the default model for Claude Code. The official positioning is straightforward: this is not the company’s most powerful model—the top-tier Claude Fable 5 still handles extreme tasks—but Opus 5’s capabilities are nearly on par with Fable 5 in programming and knowledge work scenarios, while costing only half as much.

On the same day, GitHub Copilot also announced integration with Opus 5. In the programming agent track, Anthropic raised the default bar for daily development toolchains once again with its combination of “stronger, more affordable, and more practical”.

Release Highlights: Who Should Use It, and How Much It Costs¶

Opus 5 targets long-duration, multi-step agentic programming—fixing bugs, cross-file refactoring, continuous calls within the toolchain, rather than one-off code completions. The official pricing is the same as Opus 4.8:
- Input: \(5 / million tokens** - Output: **\)25 / million tokens

The API model ID is claude-opus-5. Its main specifications are as follows:

If you are already using Claude Max/Pro or Claude Code, you will get Opus 5 without extra configuration in most cases. You can also explicitly specify it in Claude Code:

claude --model claude-opus-5

Or add this line to your shell configuration:

export ANTHROPIC_MODEL="claude-opus-5"

Benchmark Performance: Frontier-Bench and SWE-bench¶

Anthropic emphasized several real-world software engineering-focused benchmarks in its announcement, rather than just one-off function completion scores.

Frontier-Bench v0.1 (74 tasks, considered the successor to Terminal-Bench 2.1) is one of the most watched metrics this time. According to official and public interpretations:
- Opus 4.8: Approximately 18.7%
- Opus 5 (max effort): Approximately 43.3%
- Opus 5 (xhigh effort): Approximately 44.4% (in official internal benchmarks, xhigh even slightly outperforms max)
- Reference comparison: Fable 5 at ~33.7%, GPT-5.6 Sol at ~37.5%

In other words, Opus 5 outperformed competing models of the same era on Frontier-Bench, more than doubled its performance compared to its predecessor Opus 4.8, and the official emphasized lower per-task cost.

On the SWE-bench series:

Additionally, in knowledge work and IDE-focused benchmarks such as GDPval-AA and CursorBench 3.2, Anthropic stated that Opus 5 achieved a new SOTA. In CursorBench 3.2 under max effort, the gap with Fable 5’s peak score is within 0.5%, while the cost is about half. The cybersecurity direction is still led by Mythos 5, and Opus 5 does not cover this top-tier scenario.

Adaptive Thinking and Effort Parameter¶

The biggest API behavioral change in Opus 5 compared to 4.8 is that adaptive thinking is enabled by default.

On Opus 4.8, requests did not think by default, and you had to explicitly set thinking: {"type": "adaptive"} to enable it. Opus 5 reverses this: the same request will automatically enter thinking mode, with the model deciding the depth of each round of reasoning on its own. Developers use the effort parameter to control “how deeply to think”.

Available tiers: low, medium, high (default), xhigh, max.
- high: The default value for API and Claude Code, balancing quality and cost.
- xhigh: For long-duration agent tasks over 30 minutes; sometimes outperforms max on Frontier-Bench.
- max: Pursues ultimate quality, with the highest token usage and latency.

The official recommendation: Prioritize adjusting effort for cost control in daily development, rather than turning off thinking directly. In most tasks, thinking enabled + low effort often outperforms thinking disabled.

Breaking Changes to Note During Migration¶

If you still want to disable thinking, you can only set effort to high or lower:

{
  "model": "claude-opus-5",
  "thinking": {"type": "disabled"},
  "output_config": {"effort": "high"}
}

Pairing thinking: {"type": "disabled"} with xhigh or max will return a 400 error from the API. In addition, max_tokens limits both thinking tokens and reply tokens, so you need to re-evaluate your budget when migrating from 4.8.

GitHub Copilot Integration¶

On July 24, GitHub simultaneously announced that Claude Opus 5 is now available in GitHub Copilot for coding tasks that require complex reasoning, tool calls, and multi-step execution.

Coverage includes:
- Visual Studio Code, Visual Studio, JetBrains, Xcode, Eclipse
- Copilot CLI, Copilot cloud agent, Copilot app
- github.com and GitHub Mobile

For subscriptions, Copilot Pro+, Max, Business, and Enterprise users can select Opus 5 in the model selector; Business/Enterprise administrators need to enable the corresponding model permissions in Copilot policies. The rollout is gradual, and some accounts may see the access point slightly later.

For teams, this means Copilot and Claude Code can share the same Opus 5 capability baseline, making it easy to run A/B comparisons between IDE plugins and terminal agents.

How to Choose: Opus 5 or Fable 5¶

Anthropic’s own tiered strategy is clear:
1. Daily programming, enterprise knowledge work, predictable agent pipelines → Opus 5. High cost-performance ratio, with Frontier-Bench/SWE-bench data supporting the narrative of “near top-tier capability at half the cost”.
2. Highest-risk, longest autonomous runtime tasks that require maximum capability → Stick with Fable 5. Fable 5 still has a slight edge in individual metrics such as SWE-bench Pro.
3. Cybersecurity red teaming and exploit chains → Exclusive to Mythos 5; Opus 5 is not in this category.

Implementation recommendations:
1. First switch Claude Code/Copilot’s default model to Opus 5, and run a round of regression tests on your own repositories (builds, tests, typical issues).
2. Use effort to optimize costs: Start with medium or high, then gradually lower it until quality drops below your threshold.
3. Try xhigh for ultra-long tasks (>30 minutes), don’t default to max—Frontier-Bench data shows that “the highest tier ≠ always optimal”.
4. When migrating from Opus 4.8, check all code paths that disable thinking to avoid conflicts with xhigh/max.

Summary¶

The launch of Claude Opus 5 is not about hitting another benchmark score, but about bringing agentic programming capabilities nearly matching Fable 5 into the Opus price tier, and rolling it out across both Claude Code and GitHub Copilot’s main distribution channels. The default enablement of adaptive thinking, five adjustable effort tiers, and the significant improvements on Frontier-Bench and SWE-bench all point to the same thing: programming agents are shifting from a “trial feature” to the default way of working in IDEs.

If you are already using Claude Code or Copilot Pro+, there is only one thing worth doing right now—open a real project, run Opus 5 through an entire multi-file task chain, and use effort to check your billing. This will tell you more than any benchmark number whether it should become your team’s new default.