Anthropic introduced the new high-end model of the Claude family, Claude Opus 4.8. The new model developed on Opus 4.7; It came with improvements to coding, agent-based tasks, logical deduction, and professional workflows. In the same announcement, the company also introduced effort control for claude.ai and Claude Cowork, dynamic workflows for Claude Code, and new API regulation for developers. What does Claude Opus 4.8 bring?
Claude Opus 4.8 has been released as the new flagship model in Anthropic’s Opus series. According to the company, the new version is built on Opus 4.7 and showed better results in coding, agent skills, reasoning and practical knowledge work tasks in benchmark tests. The model was made available with the announcement and was offered at the same price as Opus 4.7 in standard use. Anthropic stated that the new model is not limited to just the raw performance increase.
For Claude Opus 4.8, improvements have also been made in terms of reliability, error awareness and user collaboration. Feedback from early test users stated that the model asked more accurate questions in complex tasks, caught its own mistakes more often, and acted more cautiously in situations where there was not enough evidence. One of the most important innovations introduced with Claude Opus 4.8 was effort control.
The new control, located next to the model selector on claude.ai and Claude Cowork, allows users to determine how much processing power and analysis time Claude will allocate to a task. At lower effort levels, the model responds faster and consumes its usage limits more slowly. At higher effort levels, the model analyzes more frequently and deeper and produces more comprehensive answers. Anthropic announced that Opus 4.8 runs at high effort level by default.
The company has positioned this level as the most balanced option between quality and user experience. It was stated that in coding tasks, this setting consumes tokens similar to the default usage of Opus 4.7, but offers better performance. Users can choose the “extra” level for more challenging tasks, and the “xhigh” level or “max” option on the Claude Code side. A new feature called dynamic workflows by Claude Code has also been made available as a research preview.
This feature allows Claude to schedule very large tasks within a single session and run hundreds of parallel sub-agents simultaneously. With Opus 4.8, these agents can run for longer periods of time, and Claude goes through a verification step before presenting the outputs he creates to the user. One of the examples given by Anthropic was the ability to switch from beginning to end in code bases of hundreds of thousands of lines.
With Opus 4.8, Claude Code can drive large-scale code migrations from inception to assembly, using existing testing infrastructure as a benchmark. Dynamic workflows are currently offered in Enterprise, Team and Max plans on Claude Code. An API change that concerns developers has also been released with the new version. Messages API now accepts system input in the messages array. This arrangement allows the developers to update Claude’s instructions as the mission continues.
Thus, permissions, token budget or working environment information can be changed without corrupting the prompt cache and without passing the update through the user message. A price change was also made for Claude Opus 4.8 on the fast mode side. Anthropic announced that the use of fast mode, in which the model can operate at 2.5 times speed, has become three times cheaper than previous models. The standard usage price remains the same as Opus 4.7.
The standard API price for Claude Opus 4.8 was announced as $5 per 1 million entry tokens and $25 per 1 million exit tokens. The fast mode price was 10 dollars for 1 million entry tokens and 50 dollars for 1 million exit tokens. Developers can use the model via the Claude API under the name claude-opus-4-8. Anthropic announced that Opus 4.8 also provides better results on the honesty and security side. According to the company, the model is approximately four times less likely than Opus 4.7 to pass without detecting defects in its own generated code.
Anthropic noted that the model has a reduced tendency to make unsupported claims and flags ambiguities more frequently. The model underwent detailed alignment evaluation and security testing prior to release. According to Anthropic’s statement, Opus 4.8 achieved the highest levels the company has ever achieved in positive behavior metrics such as supporting user autonomy and acting in the user’s best interest. The company also shared that rates of non-compliant behavior, such as supporting cheating or abuse, have dropped significantly compared to Opus 4.7 and are approaching the Claude Mythos Preview level.
Anthropic is also working on models that will offer similar capabilities at a lower cost after the Opus series. The company also announced that it is preparing a new model class with a higher level of intelligence that will be positioned above the Opus. As part of Project Glasswing, a small number of organizations use the Claude Mythos Preview model for their cybersecurity efforts. Anthropic stated that models at this level need stronger cyber security measures before being made available for general use.
Claude Opus 4.8 is available to all users as of today. The model can be used via claude.ai, Claude Cowork, Claude Code and Claude API. With the new version, Anthropic has simultaneously updated many parts of the Claude ecosystem, from the user interface to developer tools.
Comments
You can write your views about this story. Comments may be moderated according to site settings.