Last month, I explored how Google’s Gemini 3 compresses scientific visualization workflows, generating interactive molecular models in minutes. But more experience with the model has revealed more uneven performance.
In any case, visualization is just one piece of the R&D puzzle. The bigger bottleneck may be coordinating the dozens of discrete tasks that comprise a research project: literature synthesis, protocol optimization, data QC, documentation. Anthropic’s Claude Opus 4.5, released today, appears designed precisely for that orchestration challenge.
The new model follows in the footsteps of the Claude for Life Sciences platform Anthropic announced in October, which already integrates with Benchling, 10x Genomics and PubMed. What Opus 4.5 adds is stronger multi-agent coordination, the ability to manage teams of AI “subagents” working in parallel. Anthropic claims the model excels at handling ambiguity and reasoning through tradeoffs without hand-holding.
Opus 4.5 also supports a new “effort” parameter that lets developers dial capability up or down. At medium effort, Opus 4.5 reportedly matches its predecessor’s performance while using 76% fewer tokens. At maximum effort, Anthropic claims a 4.3 percentage point improvement on software engineering benchmarks.
In one internal test, Anthropic gave the model a customer service scenario where policy appeared to block a solution. Rather than refusing, Opus 4.5 found a legitimate two-step workaround the benchmark designers hadn’t anticipated.
The launch also comes on the heels of a rocky late summer for Claude. From early August through early September, Anthropic says three infrastructure bugs intermittently degraded response quality across several models, including its coding tools, after weeks of developer complaints and user reports of weaker answers. In a September technical postmortem, the company framed the episode as an infrastructure failure rather than deliberate throttling of quality, and promised internal process changes to reduce the risk of a repeat.
