LONG AI
Home › Blog › Kimi K3.1 Pro Independently Tested: 55,

Kimi K3.1 Pro Independently Tested: 55, Global #4, Open vs Closed Gap Down to 3 - but the 1/3 Cheaper Claim Didn't Hold

Category:行业资讯 · 11 min read · 2026-10-26

This column aggregates the latest large-model releases and open-source drops from the past one to two days, cross-checked against official announcements and multiple media outlets. Performance figures distinguish between vendor claims and independently verified third-party results; anything not yet replicated is clearly marked. Illustrations are original.

Two block towers with a narrow gap and a ruler between them illustration

Headline: Kimi K3.1 Pro independent test lands (Oct 24-25)

Yesterday's report left a note: pretty vendor numbers only count once a third party has touched them, so wait for independent replication. Today the answer arrived, and it's richer than expected - the test both confirmed its strength and debunked its marketing.

The real part: third parties confirmed "strongest open source ever"

Artificial Analysis added K3.1 Pro to its Intelligence Index (uniform environment across all mainstream models): 55 points, #4 out of 215 models. This is replicated data, not vendor claims:

The soft part: three things the vendor didn't make clear

First, "1/3 cheaper" didn't hold on the actual per-task bill. The test found K3.1 Pro is wordy: it used ~150M output tokens to finish the whole index, peer median ~87M - writing 14-70% more. Unit price dropped ($2/$10), but per task it costs about $1.98, nearly even with K3's $2.00. "Unit price down 1/3" and "your bill down 1/3" are different things - migrate on the latter.

Second, speed is below the peer median. Measured ~73.5 tokens/s, median ~82. To go faster you pay double for the Highspeed ($4/$20) - handing back the money you saved.

Third, the coding scores were almost all measured in someone else's harness. Reviewing the published Terminal-Bench 4.0, SWE-Marathon, and ProgramBench results, most ran in Anthropic's Claude Code toolchain - and the harness itself can materially change scores. The same model under a different harness may score quite differently; treat these as "directional," not "exact gaps."

Weak spots, honestly listed: ProgramBench 31.5%, Terminal-Bench Science 38.1%, both far behind GPT-6 Astra; HLE with tools 70.0 (above Opus), without tools 51.1 (Opus 59.8); long-context reasoning slightly down from K3 (88.7→86.3). The launch marketing also hit a snag: an overseas blogger posted a promotional brief that included "talk down competitors" wording, drawing criticism of the launch itself - a good model with bad marketing made some developers distrust it first.

What it means: the lesson for the industry matters more than the lesson for Kimi - the open-source camp touched within 3 points of closed-source frontier on an independent leaderboard for the first time, and the gap between "model unit price" and "task bill" got quantified and exposed. From now on, compare prices by "how much does finishing my task cost."

Other releases (7)

Model / ProductDateTypeHighlights
Hi3D V1.0 open source (formerly Sparc3D)10-23/243D generation (open, commercial)Code + weights on GitHub and Hugging Face; training assets grew from ~500K to 2M curated 3D assets, effective sparse-grid resolution up to 1536\u00b3 (in-house V3.0 already at 2048\u00b3). Vendor report: multi-view fidelity 0.75 and detail richness 0.841, both first among listed methods; 90% slicing success (3D-print ready). Three caveats: license is commercial but requires attribution with thresholds for revenue/volume; the official validation used a 96GB H20, so "runs on consumer cards" is overstated; independent replication is thin.
Dessa (Block, Square's parent)10-23Voice model (research preview, not open)Claims a "not a single word" voice model: trained from scratch on nothing but human conversation recordings - no text, no pretrained weights, no text tokenizer, audio in and audio out, full duplex, native emotion/laughter/sigh. Engineering highlight: the whole model trained on one GB200 NVL72 rack, faster than realtime, runs locally on a MacBook Pro M4 Max. Cold water: no params, data hours, scores, or license disclosed - just samples and a waitlist; "the GPT-3 moment for voice" is a claim, not a result.
WeVisOne-VL 2B/4B (Tencent)10-22Vision-language model (Apache 2.0)One model packs detection, counting, pointing, OCR, segmentation, and matting, with a 9.6M-param pixel decoder removing the SAM dependency. Vendor: 4B hits LVIS F1 55.3, 96.4% correct rejection on WeRef-Bench when the target doesn't exist, ReasonSeg gIoU 71.2. Suits document processing and visual search - "one model replaces a row of specialist models."
Nemotron 3.5 retrieval trio (NVIDIA)10-22Embed / Rerank / Parse (open weights)Embed 8B/1B, Rerank 8B, Parse 631M on Hugging Face for agent document work; vendor says Parse leads sub-billion-param models on ParseBench, and adding Rerank cuts per-query inference tokens ~13-14%. RAG teams can swap in directly.
Clef 9B/27B (Cloudflare)10-22Small decision models (open source)Built for "structured option classification": multimodal input, probabilistic output, low latency embedded in agent decision pipelines. Stripping "which path" off the big model and handing it to a small one is the standard agent cost-cut - same direction as Ant's LingDT this week.
Sage-A1.5 (AI Weekly)10-22Agent model (35B MoE)Only 3B active per token; vendor-claimed BrowseComp 81.7, SWE-Bench Pro up from 43.6 to 61.2, with the key change in a new orchestration layer and verification protocol - "orchestration + verification" is replacing "more parameters" as the agent score path.
Slot-JEPA (CAS team)10-22Robot planning (paper)Splits V-JEPA visual features into 16 object slots before planning arm motions; REALM-JEPA success up from 35.2% to 86.1% (paper). Structural insurance on the perception-planning interface, aimed at real assembly.

Industry news

The trend in one line

Today's news is one line: the open-vs-closed gap is down to 3 points, but "cheap" is being quietly eaten by token volume. K3.1 Pro's replication proved the post-training route and exposed the unit-price vs bill gap. For model users there's one conclusion: judge selection on two numbers - the independent leaderboard score, and the total cost of your real task; neither comes from the vendor's marketing page.

(Data from official announcements and public reports, 2026-10-25; vendor claims and third-party results are marked separately, with unverified items noted in the text. Model capability affects selection, but actual pricing follows the model plaza in real time.)

相关文章

GPT-6.2 Cuts Prices 20%: What the AI Price War Means

GPT-6.2 cuts list prices 20%, yet platform discounts run deeper: gpt-6-astra trades at under 1% of list. What this week's AI price war means for buyer

What Is Prompt Caching? How Cache Tokens Cut API Bills

Prompt caching stores a repeated prompt prefix and bills cache reads at 5-12% of normal input. Real prices, hit conditions, and how to cut your API bi

All 12 Claude Models Compared: Price and Best Use

Twelve Claude versions are on sale. The newest opus-5-5 costs $8.91 a month; the oldest opus-4-6 costs $49.50. Full table and a pick for each workload

Which AI Model Writes Code Best? Speed and Price Tested

gpt-5.6-terra and deepseek-v4.1-flash win on price and reliability; gemini-3.8-flash is fastest. Effective prices, measured speed, and a three-tier pi

One API Key, Every Model: How to Integrate a Multi-Model Marketplace

Point your base URL at the platform and one key covers all 138 on-sale models across 51 vendors, in both OpenAI and Anthropic request formats. Copy-pa