Co-Founder & CEO at Bigspin
- Tweets68
- Following30
- Followers138
- Likes57
ALT Blog post passage: "What role might CoT monitoring have played here? As of this writing, OpenAI has not released the traces from this exploit, and so we don’t know whether the OpenAI models described their illegal plans in their reasoning. However, I would wager that the reasoning traces contain a detailed record of the illegal actions the system took. If this is not the case – if the model effectively concealed its own thinking and planning, despite the assumptions reviewed above – then this incident should lead the community to radically rethink the project of CoT monitoring."
![Line chart titled "Token purchasing power across the engineering basket — Opus 4.6 in SWE-chat." Subtitle: each line shows units of the good 1 token buys, relative to a Feb 5–24 baseline (1.00×); knowledge capture (teal) erodes least. The full-basket composite ends at 0.23× purchasing power (95% CI [0.18, 0.28]) = 4.38× more tokens per unit. The y-axis is log-scaled "output per token," from ~0.15× to 1×. The x-axis spans five time windows A–E (Feb 05 to Apr 15), labeled as phases: Pre-mystery climb, Mystery climb, Post-mystery climb. Colored lines track five goods: agent-drafted code, PR shipped, file touched, agent-drafted docs, and knowledge capture; a thick black "COMPOSITE" line with a gray 95% CI band trends downward from 1× to 0.23×. Knowledge capture rebounds to 0.37×. Reasoning effort shifts from "high" to "medium" to "high" across phases.](https://pbs.twimg.com/media/HKTckEXb0AAUvnF.jpg?name=orig)
ALT Line chart titled "Token purchasing power across the engineering basket — Opus 4.6 in SWE-chat." Subtitle: each line shows units of the good 1 token buys, relative to a Feb 5–24 baseline (1.00×); knowledge capture (teal) erodes least. The full-basket composite ends at 0.23× purchasing power (95% CI [0.18, 0.28]) = 4.38× more tokens per unit. The y-axis is log-scaled "output per token," from ~0.15× to 1×. The x-axis spans five time windows A–E (Feb 05 to Apr 15), labeled as phases: Pre-mystery climb, Mystery climb, Post-mystery climb. Colored lines track five goods: agent-drafted code, PR shipped, file touched, agent-drafted docs, and knowledge capture; a thick black "COMPOSITE" line with a gray 95% CI band trends downward from 1× to 0.23×. Knowledge capture rebounds to 0.37×. Reasoning effort shifts from "high" to "medium" to "high" across phases.
ALT Line chart titled "Token purchasing power across the engineering basket — Opus 4.6 in SWE-chat." Subtitle: each line shows units of the good 1 token buys, relative to a Feb 5–24 baseline (1.00×); knowledge capture (teal) erodes least. The full-basket composite ends at 0.23× purchasing power (95% CI [0.18, 0.28]) = 4.38× more tokens per unit. The y-axis is log-scaled "output per token," from ~0.15× to 1×. The x-axis spans five time windows A–E (Feb 05 to Apr 15), labeled as phases: Pre-mystery climb, Mystery climb, Post-mystery climb. Colored lines track five goods: agent-drafted code, PR shipped, file touched, agent-drafted docs, and knowledge capture; a thick black "COMPOSITE" line with a gray 95% CI band trends downward from 1× to 0.23×. Knowledge capture rebounds to 0.37×. Reasoning effort shifts from "high" to "medium" to "high" across phases.
ALT Line chart titled "opus-4-6 output_tokens by week, with Anthropic's postmortem events overlaid." X-axis: week start (Wed), Feb 1–Apr 15, 2026. Y-axis: median output_tokens per session (log scale), 10³ to ~30,000. Two lines: red solid ("all opus-4-6, 3,917 sessions") and blue dashed ("persistent cohort, 15 users/943 sess."). Both start ~2,500–3,000 tokens early Feb, dip to a low ~1,000–1,800 around Feb 22, then climb sharply, ending near 25,000–30,000 by mid-April. Annotations: "Opus 4.6 launches with reasoning effort high" (early Feb); orange "What happened here?" near the dip; vertical lines mark "effort default high→medium" (Mar 4, red), "thinking-clearing bug ships" and "bug fixed" (purple), "effort default medium→high" (green). Shaded band marks medium-effort default (Mar 4–Apr 7).