Kimi K3.1's 'Chasing Astra' Headlines Meet a 13% Terminal-Bench Reality

K3.1 Chases Astra, But First Fix the 13%

On October 22, an anonymous model codenamed Kibitz appeared in Arena's code arena. That day, several media outlets ran nearly identical headlines: performance chasing GPT-6 Astra, official announcement expected by end of month.

On the same day, on Artificial Analysis's intelligence index, Kimi K3 scored 44, GPT-6 Astra scored 53.

9 points. Doesn't sound like much.

But K3 has another metric that fell harder: Terminal-Bench 4.0, 13%. On the same benchmark, GPT-6 Astra got 59%.

And this round's leaked upgrade direction says "long-horizon autonomous tasks, game and 3D code enhancement, computer operation."

Computer operation. Terminal. 13%.

Set that aside. We'll come back to it.

What's solid, what's "netizens say"

The headline says "chasing fast," but the real big claim sits in the final line of the body text: "the first domestic large model to break 50 points on the AA leaderboard."

The news itself isn't false. Four outlets reported it on October 22: Kuaikeji,站长之家, Sina Finance republishing机器之心, and Bloomberg wire. The core fact: two unclaimed models, Kibitz and Kalmia, showed up in Arena.ai's (formerly LMArena) Code Arena.

But "Kibitz is Kimi K3.1" — the official side hasn't said a single word. The community's evidence is threefold:

The first is the hardest: tokenizer fingerprint. Account @aitrackerbot ran two rounds of tests, having models repeat 11 special tokens designed to identify Moonshot models. Kibitz garbled all 11, matching every Kimi model that account has tested; meanwhile the 7 other labs' models tested all reproduced these strings accurately.

The second is an official API leak. Someone captured a response from a config endpoint on Kimi's official domain, containing a session-switch target like "to": "k3.1", showing the product side is already paving the way for K3.1. The poster added that night: the config has been pulled.

The third is rhythm. Max For AI says both companies are in the final pre-release stage, expected to launch within the week. By Arena's convention, anonymous models usually go from testing to official release in about a week.

Also, the other model in the batch: Kalmia. The community reckons it's Alibaba's Qwen 4. Someone who saw its 3D drivable car demo said it "can rival GPT-6.1 Sol Ultra." If both judgments hold, the end of this month is a three-way collision — Kimi K3.1, Qwen 4, and Zhipu GLM-5.5.

This is the first time domestic flagships aren't queuing up to launch, but colliding in the same month. Good for readers, not for the vendors.

Combined, the three pieces are fairly credible. But remember their nature: three pieces of indirect evidence, not one official confirmation.

The real missing piece is the score. Arena's anonymous blind test doesn't publish Elo, only records wins and losses in the battle backend. In other words, as of today, no public number supports "chasing Astra."

So what can be verified?

Between 9 points stand 20 configuration tiers

From 53 to 44, those 9 points squeeze in 7 models and 20 configuration tiers.

Artificial Analysis is one of the few outfits that compress "capability" into a single number and publish all raw scores. Its intelligence index currently runs v4.3.2, weighted across ten benchmarks: Agent 30%, General 30%, Programming 20%, Scientific reasoning 20%.

Look at the chart first, then the table.

AA tested K3 at two tiers, Astra at five. K3 at max settings scores lower than Astra at its lowest tier.

On this ruler:

Model	Intelligence Index	Cost per task	Output speed
Claude Opus 5.5 (max)	58	$5.98	96 tok/s
Claude Sonnet 5.5 (max)	56	$5.46	133 tok/s
GPT-6 Astra (max)	53	$3.26	56 tok/s
Gemini 4 Argon (high)	53	$1.99	79 tok/s
MiMo-V2.6-Pro	46	$0.13	52 tok/s
GPT-6 Astra (low)	46	$0.82	53 tok/s
GLM-5.3 (max)	45	$2.01	78 tok/s
Qwen3.8 Max	45	$5.41	40 tok/s
Step 5 Preview	44	$1.03	180 tok/s
Kimi K3 (max)	44	$2.00	37 tok/s

Read the table carefully before moving on.

Above 50 points sit 19 configuration tiers, all from four companies: Anthropic, OpenAI, Google, Meta. Not a single Chinese company.

The highest domestic score is Xiaomi MiMo-V2.6-Pro at 46.

So "K3.1 is expected to become the first domestic model to break 50 points," translated into plain language: it needs to climb 6 points from where it is now, crossing all 20 configuration tiers ahead of it, becoming the only Chinese model standing above 50.

This isn't "domestic first" scale. It's "sky-drop into the global top 10, and the only Chinese one" scale. From 44 to above 50 in a single version, I haven't seen a precedent.

Also, the expectation of "hoping to reach 55 or even 60 points" — 60 doesn't exist right now. The global #1 Opus 5.5 sits at 58.

The 13% is the one to watch this time

Same page, same test. Terminal-Bench 4.0 column: 59% vs 13%. On SciCode, K3 actually beat Astra by 3 points.

This chart carries the most information in the whole story.

What does Terminal-Bench 4.0 test? Give an agent a terminal, 66 tasks across seven domains — software, machine learning, science, operations, security, hardware, media — and see whether it can actually finish the job. Run three times, average pass@1. This is currently the public benchmark closest to "operating a computer."

Kimi K3 scored 13% on it. For context: GPT-6 Astra is 59%, Claude Fable 5.1 is 52%, GLM-5.3 is 42%, even Xiaomi's MiMo-V2.5-Pro has 35%.

Astra is 4.5 times K3.

And K3.1's upgrade list has "long-horizon autonomous tasks" first, "computer operation capability" third.

If those two statements are true, the first number that should move is 13%. Not pelicans, not 3D animation — the 87% of unfinished tasks in the terminal.

Now look at SciCode — K3 scores 59%, 3 points higher than Astra. That shows it's not comprehensively behind, it's lopsided: single-shot code writing isn't bad, but chaining a bunch of steps and running them to completion in a real environment is weak.

That lopsided spot happens to be the frontline battlefield of the Agent era.

57 and 44 are two different rulers

On the day of the revision, the 44-point tier had two entries: GLM-5.3 and K3, then tied for open-source top.

Need to interrupt here, or the numbers will mislead.

If you search "Kimi K3 Artificial Analysis," you'll find plenty of reports saying "57 points, global third." That's the number from K3's July release. Correct at the time.

But today, on AA's official site, K3 is 44.

The model didn't change. The ruler did.

On September 7, AA released intelligence index v4.3: Terminal-Bench upgraded from 2.1 to 4.0 (questions comprehensively harder), τ³-Banking replaced with AutomationBench-AA built with Zapier (657 private tasks, testing cross-application business flows), and the weight of evaluations with "private question sets/answers" raised from 40% to 45%.

The result: the whole industry dropped points. Claude Fable 5.1 dropped 3.39, GPT-6 Astra dropped 1.85, Grok 4.6 dropped 6.17. Claude Fable 5 was 60 in July, now 49.7.

The significance: any claim citing an absolute score must carry the version number. A "57" or "44" without a version number is a number without a scale.

So the "first to break 50" claim in that news — which ruler does it use? Probably still the July one, because under the new one, 50 means matching GPT-6 Astra's mid-tier.

A small reversal while we're at it. On the day v4.3 came out, AA's own words were "GLM-5.3 and Kimi K3 continue to lead open-weights models," both at 44. Two months later, the open-source leader is Xiaomi MiMo-V2.6-Pro at 46.

Leaderboards change, and faster than press conferences.

Its biggest rival isn't across the ocean

AA wrote "notably slow" right on the page: 37.1 tok/s, ranked 72nd of 78 models in the same batch.

Back to the table. Look at three columns.

Cost: $2.00 per task. Same-score MiMo-V2.6-Pro is $0.13, same-score Step 5 Preview is $1.03, the 45-point DeepSeek V4.1 Flash is $0.26. K3 is 15x more expensive than Xiaomi.

Speed: 37 tok/s. Among domestic models at this tier, the fastest is Step 5 Preview at 180, nearly 5x difference. AA directly wrote "notably slow" on the page.

Wall-clock time: 1231 seconds per task, 20 minutes 31 seconds. Running the entire intelligence index, K3 emits 160 million tokens.

These three together line up with the media's understated "the previous K3 was dragged down by compute, slow, and has had less presence than DeepSeek V4.1 Flash these two months."

The numbers are uglier than that sentence: K3, in its own country's first tier, ranks tied 5th in intelligence index, dead last in speed, and is the most expensive at the same score — twice Step 5 Preview.

So K3.1 has two courses to make up, not one. The 13% is the capability course; the $2.00 and 37 tok/s is the usability course. The latter is harder — it has nothing to do with model architecture, everything to do with chips.

A company preparing for IPO can't afford a second rate-limit

There's another layer here, unrelated to scores, but it will shape what K3.1 looks like.

After K3 launched in July, Moonshot AI was overwhelmed by request volume within 48 hours. The official announcement said requests "far exceeded predictions," suspended new user subscriptions, kept only existing members, then reopened in batches. The same announcement mentioned splitting Kimi main membership and Kimi Code into two plans.

The business backdrop: per China Daily reporting at the time, Moonshot's annualized revenue grew from $200 million in April to $300 million in June, then multiplied several times after K3's release; the company has been distributing resolutions to shareholders, preparing for a Hong Kong listing, possibly completing an IPO within half a year.

A company preparing for IPO fears a user who paid but can't use the product more than dropping a spot on a leaderboard.

So you'll see a contradictory pricing strategy: K3 prices output at ¥100 / million tokens, more than ten times DeepSeek V4-Pro and Xiaomi MiMo-V2.5-Pro, 3.5 times GLM-5.2, 2.8 times Qwen3.7-Max. Analysts explain it as "pricing on capability, ending domestic low-price involution."

But the user-side experience is another story: expensive, slow, and rate-limited.

These two contradictions are the real test for K3.1 this time. Not "can it match Astra," but "can it hold up after matching." 37 tok/s and 20 minutes per task are evidence it can't hold up.

Don't rush to believe the 1/14 number

The most widely circulated community number: Kibitz does the same task at 1/14 the token cost of GPT-6 Astra.

Two problems with this claim.

First, the basis is unclear. If calculated by per-token price, K3's current official price is $3 input, $15 output per million tokens; Astra is $10 / $50. The ratio is about 1/3, not 1/14. To get to 1/14, you'd have to assume Kibitz's pricing dropped dramatically again — that's speculation, not an announcement.

Second, AA officially calculates per-task cost, not per-token price. By that standard, K3 is $2.00, Astra is $3.26, only 39% cheaper. The reason: K3 spits out tokens one at a time, a single task needs 80K tokens (48K output + 32K reasoning), so a cheaper unit price can't fill the gap.

Cheap unit price + verbose output = not-cheap actual bill. In the Agent era, burning money per token, you count the total, not the unit price.

A cold shower, and how to verify yourself

The last paragraph of 机器之心's piece deserves to be pulled out: actual results are uneven. In some tests, Kibitz can reach the level of GPT-6 Astra Medium and Opus 5.5 Medium; in others, quality drops noticeably, like two different models.

The more reasonable explanation: what's on Arena is probably a checkpoint from mid-training, not the final version. Checkpoints have their ceiling showing but their floor not yet grown in.

So the only thing we can confirm now: the ceiling is high. The floor, speed, price, license — none of it is at a point where you can draw conclusions.

If you want to verify yourself, here's the path: open Arena.ai's Code Arena, throw a long, hard coding or reasoning problem in, vote, and see who's on the other side. Kibitz and Kalmia appear at random; unlucky people may need several tries. Free until official release. First come, first served.

The input box's placeholder text has already changed to "or do an Agent task" — the product narrative moved one step first.

If K3.1 really launches at month-end, three numbers matter most, in order: whether Terminal-Bench 4.0 jumps from 13%, whether per-task cost drops from $2.00, whether output speed rises from 37 tok/s.

If even one of the three doesn't move, then "chasing Astra" is still just a netizen sentiment.

Reply "k31" in the background and I'll send you the raw leaderboard data, comparison page links, and the three numbers to watch after K3.1's release.

Do you have real work that makes an agent run a terminal? Leave the most troublesome one in the comments; at month-end I'll pick a few and rerun them on K3.1 and Astra side by side.