Descriptions:
Alexander Bricken from Anthropic’s Applied AI research team delivers a detailed technical session on test-time compute — what Anthropic calls ‘the thinking lever’ — explaining how Claude models can allocate more tokens at inference time to solve harder problems. The talk traces the evolution from early reasoning models through interleaved thinking (thinking steps after every tool call) to Adaptive Thinking, Anthropic’s latest approach that gives Claude full control over when, whether, and how much to think at each step of a task.
Bricken provides a live demonstration using Claude Opus 4.7 running a traffic simulation at low, high, and max effort levels. The progression is striking: the low-effort run (~4,600 output tokens) produces a functional but basic simulation, while the max-effort run uses roughly 10x the tokens and delivers physically accurate traffic light placement, multi-vehicle type rendering, and emergent car-following behavior. Benchmark results across DarcQA, OS World (computer use), and Humanity’s Last Exam all show consistent performance gains as token budgets increase.
The session also explains the key insight behind Adaptive Thinking: rather than classifying incoming requests and hard-coding thinking behavior, Claude is given a thinking tool it can invoke freely — or skip entirely for simple questions. Bricken notes that Anthropic runs all published benchmarks on Adaptive Thinking and has found it to be Pareto-efficient relative to the previous interleaved approach. Developers building with Claude Opus 4.7 who want to tune reasoning depth will find this a foundational reference.
📺 Source: Claude · Published May 20, 2026
🏷️ Format: Keynote Launch
