Moonshot AI launches Kimi K3.1 Pro, a 2.8T-parameter open-weight model

On paper, Kimi K3.1 Pro looks like the most serious open-weight challenge yet to the closed frontier. On Terminal-Bench 4.0, a benchmark that tests coding agents on real terminal-based tasks, it scored 65.7%, clearing GPT-6 Astra's 58.2% and landing within a point of Claude Opus 5.5's 66.4% ^[1][2]. On the Artificial Analysis Intelligence Index, it placed 4th overall with a score of 55, an 11-point jump from K3's 44, trailing only Claude Opus 5.5, Claude Sonnet 5.5, and Muse Wonder Full ^[3].

The more interesting detail is where those gains concentrate. The Decoder's analysis found K3.1 Pro's strongest improvements show up specifically in agentic tasks, not as a uniform lift across every capability, which narrows its gap with Claude Opus 5.5 to just 3 points on the Intelligence Index ^[4]. That pattern holds on agentic-work benchmarks: K3.1 Pro posted 1,846 Elo on GDPval-AA (real-world agentic work tasks), up 312 Elo from K3 and within 20 points of Opus 5.5's 1,866 ^[5], and it actually surpassed Opus 5.5 on SWE Marathon v1.1 (51.7% versus 46.3%) ^[4]. Crucially, this isn't a new model built from scratch - it's a post-trained version of K3's existing architecture with a far larger post-training compute budget, which explains why the gains cluster in agentic and tool-use skills rather than raw reasoning ^[6].

Ahead of releasing K3.1 Pro's weights, Moonshot says it identified and removed cyber-related data from both of the model's training stages, a deliberate attempt to curb misuse once the weights can no longer be revoked. The company reports this dropped its score on an internal exploit test, xdev-bench-hard, from 90.5% in an unfiltered build to 81.0% in the released model ^[7]. That is a meaningful self-reported reduction, but TheNextWeb is blunt about the caveat: the figures are Moonshot's own and have not been checked independently ^[8].

The transparency gap goes deeper than a single disputed number. Implicator.ai points out that xdev-bench-hard itself is not public, and Moonshot's blog post never addresses whether fine-tuning after release could restore the capability it just stripped out ^[9]. That question matters precisely because open weights, once downloaded, are permanent - safety researchers cited by TheNextWeb note that open models can't be recalled, and their safeguards generally lag behind those shipped in closed, API-gated systems ^[8]. Moonshot is effectively asking the public to trust a safety intervention it can neither inspect nor undo.

Hallucination claims tell two different stories depending on who's measuring. Moonshot's own internal Kimi Hallu Bench Hard shows claim-level hallucination falling sharply, from 23.4% for K3 to 8.3% for K3.1 Pro ^[7]. Artificial Analysis, using its own AA-Omniscience methodology, measured a smaller but still real improvement - hallucination falling from 53% to 44% as its Omniscience Index rose from 20 to 30 ^[5][9]. Both show progress, but the gap between an 8.3% internal figure and a 44% third-party figure is large enough to invite skepticism about how either test is constructed.

That skepticism runs through the model's reception across every platform where people are actually testing it. On Reddit, enthusiasm over the raw benchmark numbers was tempered almost immediately by accusations of 'benchmaxxing' - tuning specifically to score well on known test suites - and by a LiveBench update that reportedly places K3.1 Pro below Astra Max and Sonnet Max, directly contradicting the launch framing; separately, early adopters flagged aggressive usage-limit exhaustion on paid subscriptions, a practical friction that undercuts the headline pricing story. On X, independent testers ran their own head-to-head generative comparisons against Gemini 4 Argon and Fable 5.1 using identical prompts, framing the results as uncomfortable for at least one of the three models tested. Separately, a finance-and-crypto-adjacent account summarized the raw spec sheet for its own audience rather than testing the model directly. YouTube commentary leaned into a bigger-picture framing, situating K3.1 Pro's launch inside a wave of near-simultaneous releases from competing labs and reading the timing itself as evidence of an accelerating competitive race between Chinese and US model developers; hands-on reviewers testing it directly were split between treating it as a credible step toward frontier parity and pushing back hard on that framing.

Moonshot is pricing K3.1 Pro aggressively: $2.00 per million input tokens and $10.00 per million output tokens, a 33% cut from K3, matching GPT-6.1 Sol, and roughly half what Claude Opus 5.5 costs ^[6]. That pricing pressure is paired with real enterprise distribution - Tencent's WorkBuddy platform integrated both K3.1 Pro and its Highspeed variant, though access is initially rationed to Enterprise Flagship Edition customers and individual subscribers because of resource constraints ^[12][13]. Tencent's involvement isn't incidental: alongside Alibaba's 36% equity stake, it underscores how deeply the model's commercial reach depends on China's largest tech platforms ^[11].

That commercial momentum sits awkwardly next to unresolved geopolitical pressure. In July 2026, the White House's Office of Science and Technology Policy accused Moonshot of distilling Anthropic's model and of accessing export-controlled Nvidia GB300 chips through Thailand to train its models, and Treasury Secretary Bessent warned that labs found stealing US IP could face sanctions ^[10]. Community discussion of K3.1 Pro has leaned into a counter-narrative - that Moonshot's roughly 300 engineers are competing with Google and OpenAI's vastly larger compute budgets through talent density rather than hardware access made scarce by export controls. CNBC's reporting on the preceding K3 release captured the same framing from Wall Street: Morgan Stanley's Gary Yu described it as signaling an all-round catch-up of Chinese LLMs with US leaders ^[14]. Whichever story is more accurate, Moonshot's low prices and open-weight strategy are now operating directly in the shadow of an active US sanctions threat.