Reasoning & Test-Time Compute

Reasoning is the clearest case of a technique dying by absorption. Chain-of-thought spent two years as a prompt trick, two more as a scaffolding industry, and then the scaffolding collapsed into training: the model that thinks by default replaced the prompt that asked it to.

The lineage

Figure 1. From prompt to parameter. Verifier-guided search and voting survive inside products as parallel test-time compute; the prompt-engineering layer died when the behavior moved into weights.

The test-time compute axis

Mechanism	Cost shape	Where it works
Longer chains (sequential)	Linear in thinking tokens	Math, code, anything checkable stepwise
Parallel sampling + selection	Linear in samples	Verifiable outputs; needs a checker or vote
Search over steps (PRM-guided)	Superlinear	Formal domains; heavy infrastructure
Tool calls inside reasoning	Per-call	Agentic tasks; the current frontier default

Two practical constants emerged: accuracy rises roughly with the log of thinking tokens while cost rises linearly (billed as output tokens; see our K3 cost measurements), and models overthink, spending tokens past the accuracy plateau unless budget-controlled. Effort tiers shipped in every frontier product by 2026.

Open questions the field is testing

• Elicitation vs creation: RLVR results on pass@k suggest much of trained reasoning re-weights paths the base model already contained; the boundary is unsettled.
• Transfer: gains concentrate in verifiable domains; transfer to open-ended judgment tasks is partial and disputed.
• Faithfulness: the visible chain is not a reliable account of the computation; monitoring research treats it as evidence, not transcript.
• Hallucinated confidence: reasoning models assert more; K2.6 to K3 measured hallucination rising with accuracy (audit).

Ledger

Verdict	Techniques
Dead	CoT prompting as a differentiator; prompt-level reasoning scaffolds (ToT-style) as products; “reasoning is emergent scale” as a plan
Current	RLVR-trained thinking, effort/budget controls, parallel sampling in high-stakes products, tool-interleaved reasoning
Contested	Elicitation vs creation; process rewards vs outcome rewards; how far test-time scaling extends beyond verifiable domains

Next: Multimodality, the other capability that moved from bolt-on to native.