Moonshot AI's newest model beat one of Anthropic's own Claude tiers this week1. Artificial Analysis scored Kimi K3.1 Pro at 55 points on its Intelligence Index. Claude Fable 5.1 scored 532.
Anthropic built Fable 5.1 to work safely for anyone with a Claude account. A stricter sibling, Claude Mythos 5.1, carries a different name and a different audience. Only a short list of vetted labs can run that one at all45.
The index, read straight
Artificial Analysis built its Intelligence Index to rank models on one scale. Reasoning, coding and knowledge tests all roll into a single score2.
Kimi K3.1 Pro lands fourth on that board, at 55 points. GPT-6 Astra sits beside Fable 5.1, tied at 53 apiece2. Claude Sonnet 5.5 and Meta's Muse Wonder Full tie at 56, one step above Kimi3. Anthropic's own Opus 5.5 still leads the whole board, at 58.
Kimi K3.1 Pro's 55 beats Claude Fable 5.1's 53, trailing Opus 5.5 by three
Artificial Analysis Intelligence Index scores, selected models, as scored Oct. 23, 2026. Higher is better.
Each bar is one model's score on Artificial Analysis's Intelligence Index: the magenta bar is Kimi K3.1 Pro, the blue bar is Claude Fable 5.1, and gray bars are the other models on the same board.
Sources: R&D World [2]; The Decoder [3]. Chart by Kelly Dennis.
The numbers behind this chart
| Item | Value |
|---|---|
| Claude Opus 5.5 | 58 |
| Claude Sonnet 5.5 | 56 |
| Meta Muse Wonder Full | 56 |
| Kimi K3.1 Pro | 55 |
| Claude Fable 5.1 | 53 |
| GPT-6 Astra | 53 |
Cost tells a sharper story than rank alone. Artificial Analysis found Kimi K3.1 Pro averaged $2.11 for each Intelligence Index task. Claude Fable 5.1 averaged $7.63 for the same work, more than three times as much2.
Yang Zhilin founded Moonshot AI in Beijing in 2023. His lab now sells a model that undercuts Anthropic's price on that single measure, whatever its final rank1.
What guarded actually means
A claim spread widely this week: Kimi K3.1 Pro beat the Claude model Anthropic once called too dangerous to release. Anthropic's own pages tell a different story45.
Fable 5.1 is the generally available Claude, the one anyone with an account can reach. Anthropic's own description explains the tradeoff plainly. "Claude Fable 5.1 includes robust safeguards for cybersecurity and biology. Many queries in these domains are automatically routed to less capable models if flagged by these safeguards," Anthropic wrote on its product page4.
That safeguard exists because the underlying model carries real risk. Anthropic's page names it directly: the model's skills in cybersecurity, biology and chemistry run advanced enough to enable wide cyberattacks or dangerous bioweapons4.
Claude Mythos 5.1 is the sibling that skips those safeguards. Anthropic limits it to "vetted cyberdefenders and life scientists" through what it calls trusted access programs5. That is the tier closest to what the claim describes. It stays outside Artificial Analysis's public board, because the firm scores only the models it can reach2.
The weights land in ten days
A second piece of the claim said anyone can download Kimi K3.1 Pro today. Moonshot AI's own blog says otherwise. "Full model weights will be released by Nov 4, 2026," the company wrote1.
Today the model runs through Moonshot's hosted apps and its API1. Renting access works right now. Installing it on a personal machine still waits on that November date.
Nathan Lambert, the machine-learning researcher who writes the Interconnects newsletter, flagged that gap the day Moonshot announced the model. "The benchmark scores look to be clearly the strongest current open weight model, if the weights are released as they claim by November 4th," Lambert wrote on X6.
Moonshot's own cut
Moonshot made the same tradeoff Anthropic made. It filtered its own model before opening it up. The company cut cyber-related material from two stages of training. It kept the coding skills the model needs for daily work1.
The filter carries a real cost on Moonshot's own test. Before the cut, the model solved 90.5 percent of tasks on Moonshot's internal hacking benchmark, xdev-bench-hard. After the cut, it solved 81.0 percent1.
Moonshot cut its own model's hacking score from 90.5 to 81.0 points before release
xdev-bench-hard, Moonshot's internal hacking benchmark, percent of tasks solved. Higher is better.
The blue dot is Kimi K3.1 Pro's score before Moonshot removed cyber-related training data, the magenta dot is the score after, and the rod is the nine-point cut Moonshot chose to make before release.
Source: Kimi.ai [1]. Chart by Kelly Dennis.
The numbers behind this chart
| Item | Before the cut | After the cut |
|---|---|---|
| Kimi K3.1 Pro, xdev-bench-hard | 90.5% | 81% |
Ana Maria Constantin reported the same figures independently at The Next Web. She wrote that Moonshot "removed cyberattack-related material from training before the open release, so it would not push the limits of what freely available models can do"7.
Kimi.ai@Kimi_MoonshotIntroducing Kimi K3.1 Pro: our most capable 3T-class open-weight model.
🔹 1M-token context, 2.8T total / 104B active params (MoE), Multimodal.
🔹 Full-project coding: games in Unreal, Unity & Godot.
🔹 Excels at parametric 3D modeling in Blender & CAD.
🔹 More reliable at… pic.twitter.com/klydcszv02
Both sides of the score
Rowan Sterling read the same numbers for OrcaRouter and reached a cooler verdict. "That inversion is the whole comparison, and it comes with an asterisk large enough to read from across the room: Fable 5.1 is sold as the tier you reach for when Opus is not enough, and the same index puts Opus at 58, five points above it," Sterling wrote8.
His caution lands on the comparison itself. "The one thing not to do is treat the index inversion as proof that the cheap model is better," Sterling added8.
Sterling named the tier the claim likely meant. "Claude Mythos 5.1 has the same capabilities as Fable 5.1 and is available only to organisations verified through Anthropic's programmes, such as its Cyber Verification Program," he wrote8.
Both readings start from the same board. Lambert calls the open model's climb the strongest showing an open model has managed. Sterling calls the Fable comparison one slice of Anthropic's lineup, since Opus 5.5 sits five points higher on the same board368.
What to do
Buyers choosing a model this month should weigh cost against the tier they actually need. Kimi K3.1 Pro costs a fraction of Fable 5.1 per task, and that gap matters more than one rank on one index2.
Security and biotech teams curious about Mythos 5.1 should apply through Anthropic's trusted access programs directly. A public leaderboard excludes that tier, so its real capability stays a private fact5.
By the numbers
- Kimi K3.1 Pro scores 55 points on Artificial Analysis's Intelligence Index, fourth among the models it tracks2.
- Claude Fable 5.1 and GPT-6 Astra tie at 53 points on the same board2.
- Anthropic's Opus 5.5 leads at 58 points, still its highest score on the index3.
- $2.11: Kimi K3.1 Pro's average cost per Intelligence Index task, against Fable 5.1's $7.632.
- Nov. 4, 2026: the date Moonshot has promised for Kimi K3.1 Pro's open weights1.
- Moonshot's own hacking benchmark score fell from 90.5 to 81.0 percent after the company filtered its training data1.
- Sept. 1, 2026: the date Anthropic shipped Fable 5.1 and Mythos 5.1 together4.
What to watch
Watch Hugging Face for Moonshot's promised weights on Nov. 4. Running a model locally is a different fact than renting it by the month. Moonshot's open-weight date will settle which claim holds1. Watch Anthropic's own site too, for any change to Mythos 5.1's access rules, since that tier sits closest to the viral comparison5.
Sources
- Moonshot AI, "Kimi K3.1 Pro," Kimi.ai, Oct. 23, 2026, https://www.kimi.ai/blog/kimi-k3-1-pro
- Brian Buntz, "How Kimi K3.1 Pro stacks up against Claude Opus 5.5 and GPT-6 Astra," R&D World, Oct. 23, 2026, https://www.rdworldonline.com/how-kimi-k3-1-pro-stacks-up-against-claude-opus-5-5-and-gpt-6-astra/
- Matthias Bastian, "Chinese AI model Kimi K3.1 Pro is closing in on Anthropic's Claude Opus 5.5," The Decoder, Oct. 24, 2026, https://the-decoder.com/chinese-ai-model-kimi-k3-1-pro-is-closing-in-on-anthropics-claude-opus-5-5/
- Anthropic, "Claude Fable," Anthropic, Sept. 1, 2026, https://www.anthropic.com/claude/fable
- Anthropic, "Claude Mythos," Anthropic, Sept. 1, 2026, https://www.anthropic.com/claude/mythos
- Nathan Lambert, "Kimi K3.1 is announced as 3.1 Pro...," X, Oct. 23, 2026, https://x.com/natolambert/status/2113682066374062528
- Ana Maria Constantin, "Moonshot releases Kimi K3.1 Pro and says it cut its hacking skills," The Next Web, Oct. 23, 2026, https://thenextweb.com/news/moonshot-kimi-k3-1-pro-open-weight-cyber
- Rowan Sterling, "Kimi K3.1 Pro vs Claude Fable 5.1: A Fifth of the Price," OrcaRouter, Oct. 25, 2026, https://www.orcarouter.ai/blog/kimi-k3-1-pro-vs-claude-fable-5-1