On Artificial Analysis’s HLE-Diamond test, the best model with released weights, DeepSeek V4.1 Flash, scores 34.8% on reasoning questions, according to the benchmark’s results page. Gemini 4 Carbon scores 79.6%.
That is a 44.8-point gap, by our arithmetic. On Artificial Analysis’s Terminal-Bench 4.0, the gap between Gemini 4 Carbon and the best open-weights model is 23.3 points, about half as wide.
Business Pill · OPEN WEIGHTS EXPLAINED
A short explainer on open-weights models: renting a model through an API versus running it yourself. It teaches the general idea only and says nothing about these benchmark results.
The key insight: On Artificial Analysis’s HLE-Diamond reasoning questions, Gemini 4 Carbon shows 79.6% and the best open-weights model, DeepSeek V4.1 Flash, 34.8%, a 44.8-point gap by our arithmetic; on Terminal-Bench 4.0 the gap is 23.3 points.
Artificial Analysis describes HLE-Diamond as “A refined and updated 1,000-question set of Humanity’s Last Exam, split evenly between reasoning and knowledge questions, with 272 questions that include an image”.
Its page says Gemini 4 Carbon “scores the highest on HLE-Diamond Reasoning with a score of 79.6%, followed by Claude Opus 5.5” at 78.0%. GPT-6 Astra also shows 78.0% on reasoning in the page’s data.
Among models Artificial Analysis flags as open weights, the page’s data shows DeepSeek V4.1 Flash at 34.8%, MiniMax-M3.1-Flash at 34.6%, Muse Spark Open at 33.0% and Qwen3.8 27B at 22.0% on reasoning.
Kimi K3.1 Pro shows 44.8% on reasoning. Artificial Analysis does not flag it as open weights, and the post that drew attention to the results notes its weights are not released yet.
Best open-weights model against Gemini 4 Carbon on Artificial Analysis’s HLE-Diamond Reasoning, HLE-Diamond Knowledge and Terminal-Bench 4.0, as shown on artificialanalysis.ai on 31 October 2026. Gaps are our arithmetic; Kimi K3.1 Pro, which Artificial Analysis does not flag as open weights, is not counted.
Knowledge and Agent Tasks
On HLE-Diamond’s knowledge questions, Gemini 4 Carbon shows 58.6% and the best open-weights score is Step 5 Preview at 22.0%, a 36.6-point gap by our arithmetic. The post that drew attention to the results compared Gemini with MiniMax-M3.1-Flash at 21.6%.
On Artificial Analysis’s Terminal-Bench 4.0 results, Gemini 4 Carbon shows 65.2% and the best open-weights model, GLM-5.3, shows 41.9%, according to the page’s data. Kimi K3.1 Pro, not flagged as open weights, shows 67.2%, above Gemini.
The results drew attention through a 30 October post by @TeksEdge, which argued that the test exposes “a very large frontier-model reasoning advantage when you strip away the harness, tools, retrieval and retries”. That framing is the poster’s.
The Business Lever: Where Closed Models Still Charge for Reasoning
Model makers sell access by the token, and open-weights models can be run by anyone. The scores show where a buyer still pays for a closed model: on these figures, the gap is widest on hard reasoning and narrower on agent tasks in a terminal.
As we read the page’s data, the gap between Gemini 4 Carbon and the best open-weights model is about twice as wide on HLE-Diamond reasoning as on Terminal-Bench 4.0.
As we read the page’s data, Kimi K3.1 Pro shows the highest Terminal-Bench 4.0 score among the models listed, 67.2%, and Artificial Analysis does not flag it as open weights.
As we read the page’s data, three closed models show 78.0% or more on HLE-Diamond reasoning and no open-weights model shows more than 34.8%.
Artificial Analysis, HLE-Diamond results page
“A refined and updated 1,000-question set of Humanity’s Last Exam, split evenly between reasoning and knowledge questions, with 272 questions that include an image.”
Three Implications
FOR MODEL BUYERS The page’s data shows the closed-model lead at 44.8 points on HLE-Diamond reasoning and 23.3 points on Terminal-Bench 4.0, by our arithmetic.
FOR OPEN-WEIGHTS DEVELOPERS The page’s data shows DeepSeek V4.1 Flash at 34.8% on reasoning and GLM-5.3 at 41.9% on Terminal-Bench 4.0 as the best open-weights scores.
FOR THE EVIDENCE The scores are Artificial Analysis’s own evaluations as shown on its page on 31 October; we did not rerun them.
The framework says “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize)”.
As we read it, the framework is about which layer of a stack to keep proprietary; the benchmark measures model scores, not business strategy, and the framework predates these results.
What Is Not Established
These are Artificial Analysis’s own evaluations; we did not rerun them. Its results page does not say in the text we read how tools were configured for HLE-Diamond, so the framing about stripped-away tools stays the poster’s. Scores can change as Artificial Analysis adds or retests models.
This Week In The AI Supercycle
Execution Fault Lines, Concentrated Demand, and Geopolitical Firewalls
This week’s issue: the AI build is not short of demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.
On HLE-Diamond reasoning, the best open-weights model trails Gemini 4 Carbon by 44.8 points, by our arithmetic from Artificial Analysis’s figures. On Terminal-Bench 4.0, the gap is 23.3 points.
A note on sourcing. Scores, open-weights flags and the benchmark description come from Artificial Analysis’s HLE-Diamond results page as read on 31 October 2026. Gaps are our arithmetic. The framing about tools is the lead post’s. We did not rerun the evaluations or contact Artificial Analysis.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
