Anthropic's heads of engineering and product, Katelyn Lesse and Angela Jiang, used a recent Latent Space podcast to explain how the company launched Claude Opus 5.5 and how its platform is built. They described the launch as an API launch as much as a model release. Anthropic says Opus 5.5 costs less to run than its predecessor, though the actual saving depends on the workload and how much cached context it uses. The guests also acknowledged that earlier Opus versions were sometimes too verbose, and said the company is focused on getting more intelligence per dollar. A central theme was separating an agent's 'brain' from its 'hands'. In Anthropic's Managed Agents product, the session log and harness are kept apart from the sandbox that executes code, so a failed sandbox does not have to end the whole agent session. According to the guests, an earlier design ran the entire coding agent inside a sandbox behind an API, which added startup delay before the first response and meant a lost sandbox also lost the running agent. Separating the components reduced latency and improved reliability while keeping execution isolated. Credentials are held outside the sandbox and attached through a proxy, so generated code has no service credential to read. On the API side, Anthropic is also working to protect model reasoning from extraction that could be used for distillation. Preserving thinking blocks across turns is part of that effort, but the new requirement can break existing customer integrations. The company said its near-term attention has shifted toward safety as model launches become harder, with memory and learning still on the agenda. Future platform features were described as plans rather than commitments. For Australian businesses and professionals, this changes little day to day unless you build AI agents or call the Claude API directly. If you do, test the cost claim against your own workloads rather than accepting it at face value. The credential isolation approach is worth considering for any agent system that touches customer data, since the Privacy Act 1988 applies to personal information handled through such systems. For most people using Claude as an everyday tool, nothing in the episode requires action; the practical step is to run your own evaluations on real tasks before changing models or budgets.