Google’s Gemini 3.1 Pro Preview costs an average of $1.30 per task on the Artificial Analysis Intelligence Index, according to a chart of Artificial Analysis data published by I/O Fund founder Beth Kindig. On the same chart, GPT-5.5 (xhigh) costs $2.63 and Claude Sonnet 5.5 (max) costs $5.46.
By our arithmetic, $1.30 is about 49% of GPT-5.5’s cost and about 24% of Claude Sonnet 5.5’s. Kindig puts it as 51% and 76% cheaper. Claude models take the four highest-cost spots on the chart, at an average of about $6.26 per task, about 4.8 times Gemini 3.1 Pro by our arithmetic.
Business Pill · The Model and the Harness
A short explainer on how a model and the harness around it shape what AI work costs. It is a general explainer and is not about this story.
The key insight: On I/O Fund’s chart of Artificial Analysis data, Gemini 3.1 Pro Preview costs $1.30 per Intelligence Index task, about 49% of GPT-5.5 (xhigh) at $2.63 by our arithmetic, while the four Claude models average about $6.26.
The chart’s subtitle describes the metric as the weighted average cost per Artificial Analysis Intelligence Index task, in US dollars. Its footer says the figures include input, cache, reasoning and answer tokens, and that the Claude models were run with fallback.
Kindig wrote in her 23 October article that, according to Artificial Analysis, cost per intelligence task “reflects average prices for real-world evaluation scenarios, rather than just a simple per-million token price via APIs”.
Weighted average cost per Artificial Analysis Intelligence Index task for seven of the 17 models on I/O Fund’s chart, which cites Artificial Analysis. Costs include input, cache, reasoning and answer tokens; the Claude models were run with fallback.
The Spread Across 17 Models
The cheapest model on the chart is GPT-6 Luna (max) at $0.07 per task. The most expensive is Claude Fable 5.1 (max) at $7.63. By our arithmetic, that is a spread of about 109 times between the two ends of the chart.
The four Claude entries are Claude Sonnet 5.5 (max) at $5.46, Claude Fable 5.1 (xhigh) and Claude Opus 5.5 (max) at $5.98 each, and Claude Fable 5.1 (max) at $7.63. Kindig wrote that “Claude models rank as the most expensive to run, occupying the four highest-cost positions”.
Below Claude, the chart lists Meta’s Muse Wonder (max) at $4.52 and OpenAI’s GPT-6 Astra (max) at $3.26.
The Business Lever: Chips Behind the Price
Kindig ties the Gemini figure to Google’s own chips. She quotes her earlier article: “Google noted that TPU 8i delivers up to an 80% performance-per-dollar improvement over the Ironwood TPU, particularly at low-latency targets for large MoE models.”
She also writes that “Last Q4, Google’s CEO stated there was a 78% reduction in Gemini serving unit costs throughout 2025”. As we read it, the chart puts a price on the trade-off buyers face between top-end reasoning settings and the cost of each task.
As we read the chart, the metric prices a whole evaluation task, including reasoning tokens, rather than a list price per million tokens, which is how Kindig says Artificial Analysis describes it.
As we read it, the reasoning setting matters as much as the model: the chart lists most models at their max or xhigh settings, and the Claude models were run with fallback, per its footer.
As we read Kindig’s article, she links Google’s figure to its own chips, citing Google’s statement on TPU 8i performance per dollar and a 78% cut in Gemini serving unit costs in 2025.
Beth Kindig, I/O Fund, 23 October 2026
“Claude models rank as the most expensive to run, occupying the four highest-cost positions”
Three Implications
FOR AI BUYERS The chart prices one index task from $0.07 for GPT-6 Luna (max) to $7.63 for Claude Fable 5.1 (max), a spread of about 109 times by our arithmetic.
FOR GOOGLE Kindig cites Google’s statement that TPU 8i delivers up to an 80% performance-per-dollar improvement over Ironwood.
FOR ANTHROPIC The four Claude entries are the four most expensive on the chart, from $5.46 to $7.63 per task.
The framework puts it this way: “Three competitive categories within one layer: generalist substrate (NVIDIA), hyperscaler custom (TPU, Trainium, MTIA), merchant custom challengers (Cerebras, Groq, SambaNova).”
As we read it, Kindig’s argument places Gemini’s per-task price in the hyperscaler-custom category the framework describes, with Google’s TPUs under the model. The framework does not discuss the I/O Fund chart.
The figures are from a chart made by I/O Fund citing Artificial Analysis; we did not find the same per-task numbers on an Artificial Analysis page, and its model pages show other per-task figures for other index versions. Kindig’s post names GPT 5.5 (high), while the chart labels the bar GPT-5.5 (xhigh). The Claude models were run with fallback, which the chart does not explain. The article was published on 23 October; this piece follows Kindig’s 1 November post.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
On I/O Fund’s chart of Artificial Analysis data, Gemini 3.1 Pro Preview runs an index task for $1.30, about half of GPT-5.5 (xhigh) at $2.63. The four Claude models cost the most, at about $6.26 per task on average by our arithmetic.
A note on sourcing. The per-task costs come from a chart published by I/O Fund founder Beth Kindig, which cites Artificial Analysis; quotations come from her 23 October I/O Fund article and her 1 November post on X. The 49%, 24%, $6.26, 4.8 times and 109 times figures are our arithmetic.
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.
