Anthropic has launched Claude Opus 4.8, enhancing AI performance while maintaining existing pricing structures.
Overview
- Anthropic has launched Claude Opus 4.8, just six weeks after the previous version.
- The new model enhances software engineering and reasoning capabilities while keeping the same pricing: $5 for every million input tokens and $25 for output tokens.
- Opus 4.8 has shown improved alignment metrics, significantly decreasing deceptive or harmful behavior.
In a swift update to its AI capabilities, Anthropic introduced Claude Opus 4.8, marking a notable advancement from Opus 4.7. This iteration promises enhanced speed and intelligence, boasting improvements in benchmark performance without altering the cost structure for users.
A new high-speed mode is also available, operating at 2.5 times the speed of the standard version, priced at $10 for input tokens and $50 for output. This represents a significant reduction compared to prior models.
The SWE-bench Pro benchmark, which tests AI's ability to solve complex software engineering challenges, indicates that Opus 4.8 has achieved a score of 69.2%, up from 64.3% in Opus 4.7. In contrast, OpenAI's GPT-5.5 scored 58.6%, while Google’s Gemini 3.1 Pro scored 54.2%. This marks a substantial improvement in performance relative to its pricing tier.
In academic assessments, Opus 4.8 achieved 49.8% without tools and 57.9% with tools on Humanity's Last Exam, outperforming all competitors. In practical evaluations, it scored 83.4% on OSWorld-Verified tests, a slight increase from its predecessor's score.
However, in Terminal-Bench 2.1, which measures command-line task performance, GPT-5.5 led with a score of 78.2%, while Opus 4.8 improved to 74.6%, surpassing Gemini but remaining in second place overall.
User-Controlled Difficulty Levels
Anthropic has added user-adjustable difficulty settings to the model, with a default set to "High". The "Extra" mode, or "xhigh," allows for more computational resources for difficult tasks, while the "Max" setting offers the highest difficulty. Lower settings trade off some precision for faster processing. These features are now accessible across all subscription plans.
However, users should note that the new tokenizer for Opus requires more tokens per task, potentially leading to higher costs compared to Claude Sonnet, which may be suitable for simpler tasks.
Rate limits in Claude Code have also been updated to reflect increased token usage in the higher difficulty settings.
Significant Safety Enhancements
Anthropic's alignment team reports that Opus 4.8 has reached new benchmarks for prosocial characteristics, exhibiting lower rates of deception and misuse compared to its predecessor, aligning closely with the security standards of Claude Mythos Preview.
Importantly, Opus 4.8 is four times less likely to overlook bugs in its code than Opus 4.7. While Mythos is a more advanced model available only to select organizations, Opus 4.8 positions itself as a robust choice for users seeking both performance and safety.
In addition, Anthropic has unveiled dynamic workflows in Claude Code, allowing the AI to create its own orchestration scripts and manage parallel tasks within a session, although this feature requires more tokens than standard sessions.
Comparative Pricing Landscape
Anthropic's pricing strategy contrasts sharply with recent developments in China, where models like DeepSeek V4 Pro have drastically reduced their costs to $0.435 per million input tokens and $0.87 for output tokens. Despite these lower prices, Anthropic emphasizes the quality and safety of its models as justification for the higher costs.
In production environments where compliance with safety standards is critical, the investment in Opus 4.8 may be deemed worthwhile, despite the price differential.
Testing Insights
A recent coding challenge involved creating a 3D zombie game, comparing Claude Opus 4.8, ChatGPT, and DeepSeek. While GPT-5.5 completed the task first, it lacked visual and sound elements. DeepSeek delivered a well-rounded performance, but Opus 4.8, though slower, produced the best overall quality.
All games produced in the test are available on Itch.io, showcasing the strengths and weaknesses of each AI in real-world applications. Overall, Claude Opus 4.8 demonstrates improved coding capabilities over both GPT-5.5 and its predecessor, Opus 4.7, without any increase in cost for users.
