Anthropic Unleashes Claude Opus 5: Frontier AI for Coding, Reasoning, and Agentic Use at Unchanged Pricing

A New Era of Agentic AI: Introducing Claude Opus 5

At a glance, Anthropic has officially launched Claude Opus 5, positioning it as the new flagship model in its Opus tier. This release marks a significant leap forward in frontier-class AI, particularly in agentic coding, computer use, and advanced reasoning. Remarkably, this substantial upgrade comes with no change to the existing Opus pricing structure, maintaining $5 per million input tokens and $25 per million output tokens.

Meanwhile, Opus 5 is touted by Anthropic as approaching the intelligence level of Claude Fable 5 but at half the cost. It is now the default model on Claude Max and stands as the strongest model available on Claude Pro, making advanced AI capabilities more accessible to a wider range of users and developers.

Key API-Level Enhancements for Developers

The release of Opus 5 brings several critical changes at the API level that developers should be aware of:

• Default Thinking Enabled: Unlike Opus 4.8, where “thinking” needed to be explicitly set (e.g., thinking: {“type”: “adaptive”}), Opus 5 enables internal thought processes by default. The new effort parameter now controls the depth of this thinking. Developers must review existing max_tokens values as they now cap both thinking and response text.
• Breaking Change for Thinking Control: Attempting to disable thinking (thinking: {“type”: “disabled”}) with an effort level of xhigh or max will now result in a 400 error. The restriction is enforced per request; users must either cap effort at high or omit the thinking field entirely.
• Streamlined Verification Prompts: Anthropic advises developers to remove explicit “final verification step” instructions from their prompts. Opus 5 inherently verifies its own work, and such instructions can lead to over-verification, potentially impacting efficiency. The updated Opus 5 prompting guide provides detailed tuning patterns.

In practical terms, Beyond these functional changes, Opus 5 uses the model ID claude-opus-5. It features a robust 1 million token context window, serving as both the default and maximum, with no smaller variants available.

Maximum output is 128k tokens on the synchronous Messages API, extending to 300k with the output-300k-2026-03-24 beta header for the Message Batches API. Furthermore, the minimum cacheable prompt size has been reduced to 512 tokens, down from 1,024.

Unleashing Agentic Power: Coding and Computer Use

Claude Opus 5 demonstrates significant advancements in agentic capabilities, particularly in coding and computer interaction:

• FrontierBench v0.1: On this 74-task successor to Terminal-Bench 2.1, Opus 5 achieved an impressive 43.3% at max effort, peaking at 44.4% mean reward at xhigh effort. This is a substantial leap from Opus 4.8’s 18.7%, and it outperforms Fable 5 (33.7%) and GPT-5.6 Sol (37.5%). Notably, Opus 5’s safety classifiers flagged only 5% of API calls, compared to Fable 5’s 42%.
• SWE-bench Performance: Opus 5 scored 96.0% on SWE-bench Verified and 79.2% on SWE-bench Pro, nearly matching Fable 5’s 80.0%. The multimodal variant of SWE-bench saw an even larger jump, from 38.4% to 59.4%.
• OSWorld 2.0: Agentic numbers show clear wins, with Opus 5 reaching 70.57% on OSWorld 2.0, a significant improvement over Opus 4.8’s 55.7%.
• Zapier AutomationBench: Opus 5 scored 26.0%, surpassing Opus 4.8 (17.0%) and Fable 5 (17.4%). Even at medium effort, it maintains a 24% score at a cost-effective $0.89 per task.
• Artificial Analysis GDPval-AA v2: Opus 5 secured the top two leaderboard spots with ELO 1861 and 1827, with the xhigh setting outperforming all other models while using 25% fewer output tokens than max.

Advanced Reasoning and Problem-Solving

Opus 5 exhibits gold-medal level reasoning capabilities:

• IMO 2026: Without tools or an agent harness, Opus 5 was prompted on all six IMO 2026 problems. A three-model judge panel scored all 24 generated solutions as correct, with human experts independently grading one pre-specified solution per problem at a perfect 7/7. The final 42/42 score is considered gold-medal level, exceeding the 29/42 cutoff.
• ARC-AGI-3: The ARC Prize Foundation verified a remarkable 30.16% on ARC-AGI-3 at high effort, approximately four times the best previously reported leaderboard score. For context, GPT-5.6 Sol reached 7.78%, and Opus 4.8 scored 1.52%.
• Humanity’s Last Exam: Opus 5 achieved 56.3% without tools and an even higher 64.7% with tool assistance.

The Strategic Advantage of Multimodal Tool Use

A practical lesson from the multimodal section highlights that agentic tool use scales test-time compute more cost-effectively than adaptive thinking alone.

• Chartography: Opus 5 scored 29.6% without tools, but this jumped to an impressive 83.0% with a container and an image-cropping tool.
• BenchCAD Vision2Code: Voxel IoU improved dramatically from 0.366 to 0.821 with tools, significantly outperforming Claude Mythos 5 at 0.678.

Cybersecurity Capabilities and Relaxed Safeguards

That said, While Anthropic did not explicitly train Opus 5 on cybersecurity tasks, its general capability gains led to an incidental rise in cyber prowess.

• ExploitBench: Opus 5 captured 10.14 mean capability flags and produced 99 full arbitrary-code-execution exploits. This is close to Mythos 5’s 132.
• OSS-Fuzz: Opus 5 scored non-zero on 79.4% of targets, comparable to Mythos 5’s ~80%, though it completed 4 full exploits compared to Mythos 5’s 13.

This gap—Opus 5 being strong at finding vulnerabilities but less proficient at exploiting them than Mythos 5—informs its safeguard design. Anthropic has cautiously unblocked vulnerability finding in source code.

However, binary-based scanning, penetration testing, and exploit generation remain blocked. Classifiers are expected to intervene around 85% less often than on Fable 5, and defenders can apply to the Cyber Verification Program.

Interestingly, In independent testing, the UK AISI evaluated early checkpoints of Opus 5 on three cyber ranges. Opus 5 successfully solved ‘The Last Ones’ end-to-end in 8 out of 10 attempts and reached step 22 of 23 on the harder ‘Doing Life’ range, further than any other model tested. Under the Responsible Scaling Policy (RSP), Opus 5 is classified as having CB-1 capabilities and maintains the same ASL-3 protections as Opus 4.8, not crossing the AI R&D threshold.

Robust Prompt Injection Defenses

One of the standout safety improvements is in prompt injection resistance:

• Gray Swan Benchmark: Attacker success within 15 attempts plummeted from 5.5% on Opus 4.8 to a mere 2.0% on Opus 5. This is superior to Mythos 5 (2.6%) and significantly better than GPT-5.6 Sol (20.0%).
• Claude Cowork (Browser Environments): In browser environments run through Claude Cowork, attack success dropped from 31.5% on Opus 4.8 to 3.70% with no safeguards. With auto mode enabled, it achieved a perfect 0% across all 129 environments.

Conclusion: A Powerful and Accessible AI Upgrade

However, Claude Opus 5 represents a substantial evolution in AI capabilities, delivering frontier-class performance in agentic coding, computer use, and complex reasoning. Its enhanced safety features, particularly in prompt injection, and strategic relaxation of cybersecurity safeguards for vulnerability finding demonstrate Anthropic’s commitment to both power and responsibility. All these advancements, offered at unchanged pricing, make Opus 5 a compelling choice for developers and organizations looking to harness the cutting edge of AI.

Expert Perspective

From an industry angle, the clearest signal around Claude Opus 5 is how it may influence opus. The story reads less like a one-day spike and more like a marker of broader movement.

The next phase will depend on how quickly teams, regulators, or customers react. In practice, that gives Claude Opus 5 room to reshape expectations across thinking over the near term.

For readers focused on practical impact, the best next step is to watch what changes around claude once attention turns into execution.

Frequently Asked Questions

Why does Claude Opus 5 matter right now?

A New Era of Agentic AI: Introducing Claude Opus 5 At a glance, Anthropic has officially launched Claude Opus 5, positioning it as the new flagship model in its Opus tier.

What broader change could Claude Opus 5 signal?

This release marks a significant leap forward in frontier-class AI, particularly in agentic coding, computer use, and advanced reasoning.

What should the market watch next around Claude Opus 5?

Remarkably, this substantial upgrade comes with no change to the existing Opus pricing structure, maintaining $5 per million input tokens and $25 per million output tokens.

Source: https://www.marktechpost.com/2026/07/24/meet-the-new-claude-opus-5-frontier-class-agentic-coding-and-computer-use-at-unchanged-opus-pricing/