XNews

Anthropic / Claude Opus 4.5 Release and Platform Progress

Anthropic / Claude Opus 4.5 Release and Platform Progress

Anthropic officially releases Claude Opus 4.5, focusing on coding and agent scenarios, and promotes it on the platform.

Key insights

Release and Platform Availability

Anthropic announces the release of Claude Opus 4.5, with the model identified as claude-opus-4-5-20251101 now available on the Claude application, API, and three major cloud platforms (Amazon Bedrock, Google Cloud Vertex AI, Microsoft Azure), targeting professional users and developers who require complex enterprise tasks and intensive agent workflows. The pricing is set at $5 per million input tokens and $25 per million output tokens, approximately one-third of Opus 4.1.

Leap in Coding and Agent Performance

Opus 4.5 shows significant performance in coding and agent tasks: scoring 80.9% in SWE-bench Verified, surpassing all human candidates in internal performance engineering tests, and achieving 66.3% in OSWorld computer usage tests. Anthropic reports a reduction of up to 65% in token usage for long-term coding tasks, with other tests indicating a 76% reduction in 'thinking tokens', enhancing efficiency and response speed.

Optimization of Advanced Tool Usage Architecture

The advanced tool usage described in Anthropic's engineering blog includes three key technologies: tool search tools (discovering and delaying the loading of tool definitions on demand to save tokens), programmatic tool calls (code generated by agents executed in a sandbox to coordinate multiple tools in parallel), and tool usage examples (providing calling examples for complex parameters). Internal benchmarks show these methods significantly reduce token consumption and improve calling accuracy and latency performance in multi-tool scenarios.

Impact of Pricing and Cost Efficiency

Opus 4.5 lowers the unit price to about one-third of Opus 4.1 (input $5/million, output $25/million), while the model itself can reduce output redundancy and inference rounds, achieving lower usage costs through decreased total token consumption in many tasks. Some users have noted the need to consider the initial peak performance at launch and subsequent quantization/downgrade risks.

Product Features and Ecosystem Integration

Updates at the product level for Claude include direct operations on browsers and Excel (automated web data scraping, bulk data organization, and spreadsheet modifications), as well as Claude Code desktop version supporting parallel tasks and collaboration with multiple sub-AIs. Examples demonstrate the ability to combine text, voice, and image links for short video generation and other multimodal pipelines.

Market Response and Industry Trends

Despite significant improvements in capability metrics with Opus 4.5, some users report low enthusiasm post-launch, as the industry shifts from 'tech worship' to 'pragmatism': users are increasingly asking whether it can generate revenue, enhance efficiency, and whether it is worth the cost. Some opinions suggest that Claude's positioning is too high-end, with its pricing and specialization making it hard for average users to perceive direct value, while more standout applications (like Nano Banana Pro, Veo, Sora) are more likely to excite the public.

Safety and Inference Control Features

Opus 4.5 introduces a hybrid reasoning model and 'effort parameter', allowing a trade-off between immediate responses and extended thinking; memory management has improved to support long-context agent scenarios; safety enhancements are reflected in stronger resistance to prompt injection and a lower rate of problematic behaviors.

Source posts