What is extended thinking in Claude Platform? — Transcript

Full transcript

1. 0:03Some tasks need more than a quick
2. 0:04answer, and Claude can work through the
3. 0:06problem before responding. And this is
4. 0:08called extended thinking.
5. 0:12Extended thinking lets Claude reason
6. 0:14step-by-step before producing a final
7. 0:16response. When enabled, Claude generates
8. 0:19internal reasoning tokens, often called
9. 0:21a chain of thought, and then delivers
10. 0:23the answer.
11. 0:24You can see the reasoning in the
12. 0:25response alongside the final text. So,
13. 0:27with Opus 4.7, thinking is adaptive. You
14. 0:31don't pick a token budget. You turn it
15. 0:33on, and Claude decides dynamically when
16. 0:35to think and for how much. Now, to
17. 0:37control how much Claude thinks, use the
18. 0:39effort parameter. But, it goes inside
19. 0:41output config, not next to the thinking
20. 0:43block. The levels are low, medium, high,
21. 0:47extra high, or max. The default is set
22. 0:49to high.
23. 0:50Extended thinking helps with math,
24. 0:52multi-step logic, code debugging,
25. 0:54regulatory analysis, anything that
26. 0:57involves tradeoffs or comparing options.
27. 0:59Skip it for simple classification,
28. 1:01extraction, or boilerplate. It just adds
29. 1:03latency and cost without actually
30. 1:05improving those results.
31. 1:07So, let's see thinking in action. I have
32. 1:09this agent loop with one weather tool,
33. 1:11and I'm going to ask Claude to plan a
34. 1:13road trip out of San Francisco.
35. 1:16Two stops, weighing weather and drive
36. 1:18time. That's a real tradeoff, the kind
37. 1:19of question where thinking earns its
38. 1:21keep. So, let's run it.
39. 1:25The output is more interesting than
40. 1:26usual. You'll see some thinking blocks
41. 1:28where Claude works through the
42. 1:29tradeoffs, followed by tool calls to
43. 1:31check each city, and finally a text
44. 1:33block with the actual recommendation.
45. 1:35The reasoning is visible. That's the
46. 1:37whole point. So, in a production app,
47. 1:40this is the difference between an agent
48. 1:41that finds problems one at a time and an
49. 1:44agent that connects them. So, in this
50. 1:45compliance review app, toggling adaptive
51. 1:48thinking on the auto review call lets
52. 1:50the agent reason across report sections,
53. 1:53catching things like a wind load spec in
54. 1:55section three that conflicts with the
55. 1:56material spec.
56. 1:58>> [music]
57. 2:02>> Extended thinking gives Claude room to
58. 2:04reason before it even answers. With Opus
59. 2:064.7, turn on with thinking type adaptive
60. 2:09and dial the depth with output config
61. 2:11with your effort level.
62. 2:13Use it for hard trade-off heavy problems
63. 2:16and skip it for simple ones. It just
64. 2:18costs latency and tokens.