LLM News: Kimi K3.1 Pro Goes Open, GPT-6 Cyber Stays Gated

Moonshot AI's biggest change this week was a license, not a launch.

It published the full Kimi K3.1 Pro weights on Hugging Face. Kimi K3.1 Pro now ranks 12th on the BenchLM leaderboard with an estimated 75.03, the highest of any open-weight model we track.

This issue covers November 2 to 7, 2026. We read scores and ranks from our data on November 7, and prices are per million input and output tokens.

Kimi K3.1 Pro is now the strongest model you can download

Moonshot released Kimi K3.1 Pro on October 23 as an API model. Its Hugging Face repository, dated November 2, now holds the full MXFP4 weights. The model card lists 2.8 trillion total and 104 billion active parameters. You need a rack of B300 or H200 GPUs to run it, not a workstation, and we added it to the self-host cost calculator on November 4.

Its license is modified MIT with two conditions. Model-as-a-service businesses with more than $20 million in trailing 12-month revenue need a separate agreement with Moonshot. Products with more than 100 million monthly users or $20 million in monthly revenue must display the model name. Below those lines, it works much like MIT.

Moonshot's own API still lists it at $2/$10. If you were waiting for a frontier-class model you could host yourself, this is the strongest one so far. Our open-weight models explainer covers what license terms like these mean in practice.

Two smaller labs shipped weights too. Ricursive Intelligence, the chip-design lab behind AlphaChip, released Ripple-65B Preview under Apache 2.0 on November 4. It is a 65-billion-parameter dense, text-only model trained from scratch, and Ricursive says the whole program cost $32 million. OpenRouter serves it through GMI Cloud at $0.20 input and $0.60 output, but only with a 16,384-token context, not the full 1M.

We also added three decision models: Celeris-1.1 Decision, GLiDE v1.2 and Drex Max. They answer typed questions with a probability per option instead of generating text.

GPT-6 Cyber is real, and most teams can't use it

OpenAI released GPT-6 Cyber on November 5. It is GPT-6 Astra post-trained for vulnerability discovery and exploit development, and for staying inside the scope a researcher authorizes. OpenAI rates it Critical in cybersecurity under its Preparedness Framework.

Only approved Daybreak Red customers get access, through the API and Codex. That program requires identity verification, SSO, MFA and IP allowlisting. List price is $20/$100.

OpenAI reports 48.5% on ExploitGym, against 42.4% for GPT-6 Astra. On its honeypot test, GPT-6 Cyber touched out-of-scope services in 0.07% of runs without deployment safety instructions and in none with them. OpenAI ran every one of these evaluations itself. We show them on the cybersecurity page and keep them out of rankings, along with five new cyber benchmark pages such as RemoteExploitBench.

Inkling 2 Small became a base other labs build on

Thinking Machines Lab launched Inkling 2 Small on November 5. It is a mixture-of-experts model with 594 billion total and 17 billion active parameters, image input, and a 1M-token context. On Tinker it costs $0.3/$1, and it enters our leaderboard at 56th with an estimated 59.06. Thinking Machines says full weights follow under Apache 2.0 once external safety testing ends.

Cognition had already shipped a product on the earlier Inkling-Small base. SWE-2 Flash, released November 4, is its small, fast coding model. Cognition says it costs a fifth of SWE-2 and runs about three times faster, and it is free in Devin Desktop and Devin CLI through November 30. There is no standalone API or list price yet. Cognition published its own benchmarks without harness details, so SWE-2 Flash has no rank.

Tencent released Hy4 the same day, adding image input and computer use. List price is $0.834/$2.501, and Tencent bills half that through December 31. Tencent says it will open-source the weights, but no checkpoint was public on November 7, and its launch benchmarks are all Tencent-run.

New benchmarks crown whoever ran them

We added four benchmark pages this week. Read them together, because the leader depends on who ran the test.

OpenAI's two benchmarks only test OpenAI models, so "GPT-6 Astra leads" there means it beat its siblings. LifeVal is still useful inside the family. GPT-6 Astra scores 80.6 and GPT-5.6 Sol scores 80.0, so on everyday tasks the newer, pricier model barely moves the result.

Outside benchmarks test across labs, and they split. On ScarfBench, Gemini 4 Argon leads at 75.2, Claude Opus 5.5 follows at 70.5, and GPT-6 Astra sits at 65.4. On Clocktower Bench, a social deduction game that rewards bluffing and reading other players, GPT-6 Astra leads Claude Fable 5.1 by 264 rating points.

No lab wins both.

Our November 6 data refresh also updated human-preference ratings. Claude Opus 5.5 and Gemini 4 Argon now sit within each other's margin on creative writing, a near tie worth knowing if you pick a model for prose.

Release dates and sources for every model above are on model updates. We'll be back next week with the next issue.