Kimi K3.1's 'Chasing Astra' Headlines Meet a 13% Terminal-Bench Reality
Choosing a model on launch-week headlines is the trap: today's K3 fails 87% of Terminal-Bench 4.0 tasks that require finishing work in a real terminal, and the $2.00 per-task cost with 37 tok/s output fails any cost or latency bar in production. The reporting also shows why benchmark versions matter — AA's v4.3 re-run dropped unchanged models by up to 6 points, so any team comparing absolute AI scores must pin the version or the comparison is meaningless.
On October 22, unclaimed models Kibitz and Kalmia appeared in Arena.ai's Code Arena, and four outlets ran headlines saying Kibitz's performance chases GPT-6 Astra. The community's case for Kibitz being Moonshot's Kimi K3.1 relies on three indirect trails: a tokenizer fingerprint test where Kibitz garbled all 11 Moonshot-identifying tokens, an official Kimi API config leak containing a 'k3.1' session target that was later pulled, and release-timing signals. No Arena Elo is public, so no hard number supports the claim yet.
The measurable numbers come from Artificial Analysis' intelligence index v4.3.2. K3 scores 44 against Astra's 53, a 9-point gap that actually spans 7 models and 20 configuration tiers, with no Chinese model above 50 points. The sharper problem sits in Terminal-Bench 4.0, where K3 lands at 13% against Astra's 59% — 4.5x worse — while beating Astra by 3 points on SciCode. That lopsided profile means the weakness is chained, real-environment execution, exactly the capability agent products need. The same run exposes K3 at $2.00 per task (15x more than same-score MiMo-V2.6-Pro) and 37 tok/s, ranked 72nd of 78; AA calls it 'notably slow.'
The 'chasing Astra' framing rests on three indirect evidence trails, one of which was withdrawn, and no Arena Elo is public; the gap between the press headline and what any grader can verify is the actual story — the model may be excellent, but the claim isn't supported yet.
K3's lopsided profile — ahead of Astra on SciCode, 4.5x behind on Terminal-Bench — places the deficit exactly where agent work happens: chaining steps through a real environment. That makes the 13% the number to watch, not the 9-point index gap.
The $2.00 per-task cost and 37 tok/s output are chip problems, not model problems, so a software update won't fix them. That connects directly to Moonshot's IPO prep: a company that already rate-limited users once after K3 launched cannot afford to repeat that while raising capital.
AA's v4.3 score drops across every lab show that benchmark revisions can wipe out a version's worth of apparent progress; any absolute score quoted without a version number is a number without a scale.
This whole article is AI-generated, right? The AI flavor is extremely strong (facepalm)
A big spoon of beef bone broth paste, brewed with 3L of plain water, so it's extremely strong (grin)