anthropic has introduced Close Job 4.8an update to Claude Opus 4.7 that the company says improves coding performance, agency tasks, reasoning, and knowledge work.
The model can be accessed via claude.ai, Claude Code, and Claude API, where it is identified as claude-opus-4-8. The release also introduces new platform controls, including user-adjustable effort settings, dynamic workflows in Claude Code, and live updates to the Messages API.
Price and effort controls in Claude Opus 4.8
Claude Opus 4.8 maintains the same pricing structure as Claude Opus 4.7. Standard mode costs five dollars per million tokens in and twenty-five dollars per million tokens out.
Fast mode is priced at ten dollars per million tokens in and fifty dollars per million tokens out. Additionally, Opus 4.8’s fast mode runs at 2.5 times the speed of standard mode.
Claude.ai and Cowork users can now adjust how much effort Claude puts into a response, which directly affects token usage. Opus 4.8 defaults to a high effort setting.
Anthropic claims that even at this default level, encoding tasks use approximately the same number of tokens as Opus 4.7, while providing better performance.
An xhigh setting is available for tasks that require more calculations. Claude’s code rate limits have been increased to support the increased token consumption allowed by these effort controls.
These new options are part of Anthropic’s shift toward token-based billing rather than fixed subscription tiers.
Dynamic Workflows in Claude Code and Live Updates to the Messages API
Claude Code now supports dynamic workflows, allowing users to plan work, run parallel subagents, verify results, and report results.
Anthropic explains that this feature is targeted at large codebases and can handle codebases with hundreds of thousands of lines. Currently in research preview, dynamic workflows are available for Enterprise, Team, and Max plans.
The Messages API now supports real-time updates to the message array while an agent is active. Developers can change instructions, adjust permissions, modify token limits, or update the context during a task without interrupting message caching or requiring a new user turn.
Performance and security changes in Claude Opus 4.8
Anthropic claims that Opus 4.8 offers improvements over Opus 4.7 in several benchmarks, including coding, agent functions, reasoning, and office tasks.
The company also claims that the model is about four times less likely than Opus 4.7 to pass faulty code without commenting on it.
In terms of security, Anthropic reports that Opus 4.8 shows lower levels of deception and a lower tendency to comply with model misuse attempts compared to Opus 4.7. These security metrics are said to be similar to those of Claude Mythos Preview.
External evaluators mentioned by Anthropic include software development, law, finance and research organizations. CursorBench found that Opus 4.8 required fewer tool steps to achieve the same output quality as comparable models.
Another tester noted that the model’s profitability was on par with GPT-5.5 in internal benchmarks.
Looking to the future
Anthropic said it is developing models that offer current capabilities at a lower cost, as well as a new class of model that surpasses the existing Opus level.
The Mythos-class models, which are currently limited to Project Glasswing participants, are expected to be available to customers in the coming weeks once additional safety measures are implemented.
Users can now access Claude Opus 4.8 at claude.ai, Claude Code, and through the Claude API.
