Meet the New Claude Opus 5: Frontier-Class Agentic Coding and Computer Use at Unchanged Opus Pricing
Anthropic released Claude Opus 5, replacing Opus 4.8 as the flagship model with unchanged pricing ($5 input/$25 output per million tokens) and a 1M-token context window. "Thinking" is enabled by default, controlled by an effort parameter, with breaking changes that prevent disabling thinking at high or max effort levels. The model demonstrates significant leaps in agentic coding and computer use, scoring 70.57% on OSWorld 2.0 and 96.0% on SWE-bench Verified, outperforming previous versions and c
Analysis
TL;DR
- Anthropic released Claude Opus 5, replacing Opus 4.8 as the flagship model with unchanged pricing ($5 input/$25 output per million tokens) and a 1M-token context window.
- "Thinking" is enabled by default, controlled by an effort parameter, with breaking changes that prevent disabling thinking at high or max effort levels.
- The model demonstrates significant leaps in agentic coding and computer use, scoring 70.57% on OSWorld 2.0 and 96.0% on SWE-bench Verified, outperforming previous versions and competitors like GPT-5.6 Sol.
- Reasoning capabilities reached gold-medal level on IMO 2026 problems and achieved a verified 30.16% on ARC-AGI-3, roughly four times the prior best score.
- Safety measures include reduced false positive rates for prompt injection and a nuanced cybersecurity policy that allows vulnerability finding while blocking exploit generation.
Why It Matters
This release signals a strategic shift toward agentic workflows where deep reasoning ("thinking") is mandatory for complex tasks, forcing developers to adapt their prompting strategies and API configurations. The substantial performance gains in coding and autonomous agent benchmarks suggest that Opus 5 is becoming a primary candidate for automated software engineering and complex multi-step task execution. Additionally, the refined safety protocols regarding cybersecurity and prompt injection provide critical insights for enterprises deploying LLMs in sensitive or interactive environments.
Technical Details
- API Changes: Thinking is on by default; the
effortparameter controls depth. Settingthinking: {"type": "disabled"}withxhighormaxeffort returns a 400 error. Developers are advised to remove manual verification prompts to avoid over-verification. - Context and Limits: The model ID is
claude-opus-5. Context window is 1M tokens (default and max). Max output is 128k tokens via Messages API, extendable to 300k via Message Batches API beta header. Minimum cacheable prompt size reduced to 512 tokens. - Benchmark Performance:
- Coding/Agentic: 43.3% on FrontierBench v0.1 (vs 18.7% for Opus 4.8), 96.0% on SWE-bench Verified, 70.57% on OSWorld 2.0, and 26.0% on Zapier AutomationBench.
- Reasoning: 42/42 correct on IMO 2026 problems (gold-medal level), 30.16% on ARC-AGI-3 (verified), and 56.3%-64.7% on Humanity's Last Exam.
- Multimodal: Chartography score improved from 29.6% to 83.0% with tool use; BenchCAD Vision2Code voxel IoU increased from 0.366 to 0.821 with tools.
- Safety Metrics: Indirect prompt injection success dropped to 2.0% on Gray Swan benchmark. Browser attack success in Claude Cowork dropped to 3.70% without safeguards and 0% with auto mode. Cybersecurity classifiers intervene 85% less often than Fable 5, though exploit generation remains blocked.
Industry Insight
- Prompt Engineering Evolution: The mandatory nature of "thinking" at higher effort levels means developers can no longer rely on simple, fast inference for complex tasks. Optimization strategies must now focus on managing token costs associated with extended reasoning traces and adjusting
max_tokenslimits accordingly. - Agentic Infrastructure Demand: The superior performance in OSWorld and SWE-bench suggests a growing market need for robust agentic frameworks that can effectively orchestrate Opus 5's tool-use capabilities. Enterprises should prioritize integrating models with strong computer-use and coding agents into their CI/CD and automation pipelines.
- Security Posture Adjustment: With reduced false positives in safety classifiers and specific restrictions on exploit generation, security teams should update their threat models. The ability to find vulnerabilities but not exploit them creates a new dynamic for red-teaming and defensive coding practices, requiring careful handling of the Cyber Verification Program for advanced use cases.
Disclaimer: The above content is generated by AI and is for reference only.