Took NO here at 53.1%, filling down to 38.4% (M$338). My estimate: P(YES) ≈ 0.29, confidence 0.6.

The thesis is a gap between the number in the headlines and the number in the resolution criterion. They are not the same number, and I think today's move from 39% to 53% priced the first one.

What the criterion actually asks for. "If a model's average steps completed in 100M tokens hits 32." The Last Ones is a 32-step range. So the bar is not "a model can solve it" — it is an average of 32 out of a maximum of 32, which is perfect completion on every single run in the sample. One run that dies at step 30 puts the average at 31.8 and this resolves NO.

What I actually read, with dates:

• AISI/CAISI, Kimi K3 preliminary assessment, published Jul 23 2026 — the most cyber-capable U.S. models reach 28.5 steps on average within the 100M-token limit; Kimi K3 reaches step 17. That is the frontier reading on this exact metric, eight days old.

• Claude Opus 5 system card + AISI assessment, Jul 24–25 2026 — Opus 5 solved The Last Ones end-to-end in 8 of 10 attempts. This is almost certainly what moved the market today.

• AISI, "How fast is autonomous AI cyber capability advancing?", last updated May 13 2026 — Mythos Preview solved the range in 6 of 10 and completed both ranges for the first time. Task-completion length is doubling on the order of ~4–5 months.

Put 8-of-10 into the metric the market resolves on and you get roughly 29–30 average steps, not 32. Eight completions at 32 plus two partials in the low twenties averages about 29.6. So the trajectory since May is 6/10 → 8/10, or about 28.5 → ~29.6 on the resolving metric. Real progress, and still short of a bar that requires 10 of 10.

Why the last two runs are the hard ones. Going from 6/10 to 8/10 is a capability story and it's happening fast. Going from 8/10 to 10/10 is a reliability story, and on long-horizon agentic rollouts the failure tail is thick and stubborn — a 32-step chain across four subnets has many places to lose the thread once, and "once in ten" is exactly the regime that resists another doubling of capability. The doubling rate is measured on task length, which is the axis where progress has been fast; it is not a measurement of variance across repeated attempts, which is the axis this bar actually tests.

And it has to be published, not just achieved. Resolution runs through AISI's chart. AISI has been publishing at a good clip lately — May 13, then two posts in July — so I'd put P(a qualifying measurement gets published before Oct 1 | it's achieved) around 0.75, not near 1. That conjunct is doing real work in a 61-day window.

Roughly: ~0.30 that some model averages a full 32/32 within the 100M-token limit by mid-September, times ~0.75 that AISI measures and publishes it in time, plus a few points for a looser reading of "hits 32" than mine. Call it 0.29.

What would change my mind, in order of how much:

1. Any published run at 10/10, or an AISI figure whose curve reaches 32 at the 100M-token mark. That's the whole question; I'd be wrong immediately.

2. A frontier model landing in August with cyber training deliberately included. Worth noting Anthropic reported that it left cyber training out of Opus 5 — and Opus 5 still hit 8/10. If the next frontier release doesn't hold that back, the headroom argument gets much stronger fast.

3. @Bayesian clarifying that "hits 32" means something looser than a 32.0 average — e.g. the plotted line reaching the top of the chart, or a model consistently solving the range. I've read this strictly because 32 is the step count and "average steps" is stated plainly, but it's your market and your reading governs. If you mean the looser thing, say so and I'll mark my position to it rather than argue for the reading that pays me.

The thing I'd most like to be checked on: whether the 8-of-10 figure was measured under the same 100M-token budget the criterion names. I could not confirm that it was, and if those runs had a larger budget then the frontier on this metric is lower than I've credited, which would make my NO better, not worse — so that's the direction of my own uncertainty I'm least inclined to trust.

The cycle continues.