Anthropic has officially launched Claude Opus 4.8, an advanced iteration of its large language model, Claude Opus 4.7. The company asserts that this upgrade delivers significantly improved performance across critical domains including coding, sophisticated agent work, complex reasoning tasks, and general knowledge work. Users can access the capabilities of Claude Opus 4.8 through claude.ai, a specialized environment called Claude Code, and via the Claude API, where it is identified by the API name claude-opus-4-8.

Alongside the release of Opus 4.8, Anthropic has refined various aspects of its product line-up. For instance, users interacting with claude.ai and Cowork now have the ability to modulate the 'effort' Claude dedicates to generating a response. This feature directly influences the number of tokens the model consumes, allowing for more granular control over resource usage. Claude Code introduces dynamic workflows, an innovative feature designed to plan tasks, execute parallel sub-agents, meticulously verify outputs, and report findings back to the user. Furthermore, the Messages API has been enhanced to accept live modifications to the messages array, empowering developers to update instructions during an ongoing task without disrupting prompt cache utilization or necessitating a new user turn.

Regarding the economic aspects of this release, Anthropic has stated that the pricing for Claude Opus 4.8, when not operating in 'fast' mode, will remain consistent at $5 per million input tokens and $25 per million output tokens. For users opting for the 'fast' mode, the cost escalates to $10 per million input tokens and $50 per million output tokens, though this mode offers a 2.5 times speed increase, as detailed in the company's official announcement.

Opus 4.8 is specifically engineered to excel in coding and agentic workflows, particularly in scenarios where the model needs to leverage internal tools and conduct self-correction. Anthropic highlights that Opus 4.8 demonstrates notable improvements over its predecessor, Opus 4.7, across various benchmarks for coding proficiency, agentic capabilities, reasoning, and office productivity tasks. A comprehensive System Card is also available for those seeking more subjective details and deeper insights into the model's characteristics. The platform has undergone rigorous testing by several companies spanning sectors like software development, legal services, finance, and research, prior to its broader release.

Feedback from early testers has largely commended Opus 4.8's agentic workflows. One tester remarked on achieving cost parity with GPT-5.5 during internal benchmark testing, while CursorBench specifically noted that Opus 4.8 required fewer tool steps to achieve comparable output quality. Anthropic emphasizes that Opus 4.8 is significantly less prone—four times less likely—to overlook or pass flawed code without comment compared to Opus 4.7. Moreover, the new model exhibits lower rates of deception or tendency to comply with misuse, a characteristic described as comparable to those demonstrated by Claude Mythos Preview.

The newly introduced effort control mechanism provides users with a flexible way to manage the inherent trade-offs between response quality, generation speed, and token consumption rates. While Opus 4.8 defaults to a 'high effort' setting, particularly for coding tasks, Anthropic notes that this higher default uses token numbers similar to Opus 4.7 but yields superior performance. For tasks demanding even greater computational intensity, users can select an 'xhigh' effort setting. To accommodate the resulting increased token usage, Anthropic has proactively raised the rate limits for Claude Code.

The dynamic workflows within Claude Code are particularly beneficial for managing large codebases, possessing the capability to migrate codebases encompassing hundreds of thousands of lines. These advanced features are currently in a research preview phase and are accessible to customers on Enterprise, Team, and Max plans. The Messages API's ability to update instructions dynamically during an agent's run—through edits within the messages array—allows for real-time adjustments such as modifying permissions, altering token budgets, or adjusting context without interrupting the agent's ongoing work.

Looking ahead, Anthropic also used this release to hint at its ongoing development of future models that promise to deliver the current levels of ability at a reduced cost for users. The company anticipates releasing a new class of models that will surpass the capabilities of the current Opus platform. Its strategic roadmap includes Project Glasswing, an initiative where a consortium of organizations is actively utilizing Claude Mythos Preview for advanced cybersecurity scanning. Anthropic clarifies that models of this capability level necessitate more robust safeguards before a general release to all customers, with expectations to bring 'Mythos-class' models to market in the coming weeks. The additional controls integrated into Opus 4.8 are designed to transparently expose the cost and effort trade-offs to users as the company transitions from subscription tiers to a more granular token-based billing system.