the per-token comparison keeps missing that k3 spends way more tokens per task. if it burns 3x tokens to reach the same result as fable, cheap per-token stops mattering

They address that in the article :) to quote :-

"So where's this huge price gap coming from? token pricing, prompt caching, and effort-per-task. On SWE for example, K3 works much harder than Fable: roughly 55 turns and 1.3M tokens a task versus 21 turns and 130K. On the long terminal tasks it's the other way around: Fable is the one that spirals, running up 64 turns and 1.5M tokens (sometimes straight into a timeout).

Prompt caching does most of the work of turning that effort into K3's price advantage: even when K3 reads ten times the tokens, with cache hits that means that SWE runs still come in lower cost than Fable. There’s a tradeoff. Tasks with extra turns generally mean more wall-clock time per run i.e. slower runs. If you need an answer in two seconds, that matters; if you're running agents in the background at scale, a bill that's a fraction of the size matters a lot more."

So this makes sense for a standard SaaS app - but given that models in general perform much better with low context window usage, it probably also means that Fable is still significantly better at 'frontier-level tasks' -- hard research problems, complex geometric rendering algorithm optimization, etc., no?

does it separate flick errors from tracking errors, or is it just angular delta over time? valorant aim usually feels dominated by confidence-to-fire more than path smoothness.

why Zmmul and not full M — is it just fpga area, or did the doom port let you drop div/rem?

I'm kinda lazy lol. We are also implementing out-of-order rn, so we want to get a really solid rv32i setup going before adding multiplication, floating point, vector, etc.

is the target activation actually measured against per-subject fMRI, or optimized only against the encoder model? the landing page is ambiguous and it changes what the whole result means.

cost per task > opus-low is a weird place to land. is there a specific task shape where sonnet 5 medium actually wins?

are there existing ant-keeping trackers he compared against, like the AntsCanada community tools? curious what gap he saw.

if fugu really is an orchestrator dispatching to opus/gpt under the hood (as the openrouter page suggests), the $20-in-one-prompt complaints actually start making sense — you're paying api markup twice.

neat. when you say production-scale traces don't fit in context — does halo cluster failures first or just stream-summarize each trace?

It's more-so recursive decomposition rather than clustering failures or stream summarizing. You can look at this repo if you want to learn more about how RLM's work: https://github.com/alexzhang13/rlm