A fine-tuned Nvidia Nemotron 3 Nano model with 4 billion parameters scored about 73% at sorting construction firm McCarthy’s requests for information (RFIs), according to a chart in Palantir’s Q3 2026 business update, published with its results on 2 November. On the same chart, Claude Sonnet 5.5 scored about 66% and GPT-6 Astra about 62%. The chart plots the Nemotron model at about $0.10 per million tokens and GPT-6 Astra at $20, a 200-fold price gap by our arithmetic.
Update, 3 Nov: what the earnings call added. Chief revenue officer Ryan Taylor cited Nvidia on the call at [06:23]: by post-training its open models, Nvidia “was able to get its smallest, most efficient model to outperform its largest model by a large gap, and can now retrain that model on just a single GPU in a couple of hours”. He quoted Nvidia as saying: “we captured our alpha, and we trained it into a model”, according to our machine transcription of the webcast. Taylor gave no benchmark, model name or cost figure for that example.
It is Palantir’s own customer test, on one task, and not an independent benchmark. The slide prints no data labels, so the accuracy figures are our reading of the plotted points, accurate to about one point.
Business Pill · Fine-Tuning
An evergreen Business Engineer explainer on fine-tuning: how a general AI model is adapted to one job. It teaches the general idea only and does not discuss Palantir or McCarthy.
The key insight: By our arithmetic, on Palantir’s own chart the fine-tuned 4-billion-parameter Nemotron model sits about 11 points above GPT-6 Astra at one two-hundredth of its blended price, and about 20 points above the same model untuned.
The chart is titled “Fine-Tuned Model Performance in Automating RFI Classification” and carries the McCarthy and Palantir logos. Its vertical axis is “Balance Accuracy (%)” and its horizontal axis is “Cost per 1M Tokens (USD)” on a log scale, with the note “Blended price · 3:1 input-to-output tokens”.
Five points are plotted. By our reading of the vector PDF, Nemotron 3 Nano 4B, marked fine-tuned, sits at about 73% and $0.10. The same model, marked untuned, sits at about 53% at the same price. GPT-6 Luna sits at about 59% and $0.20, Claude Sonnet 5.5 at about 66% and $4, and GPT-6 Astra at about 62% and $20.
By our arithmetic, the fine-tuned model is about 11 points above GPT-6 Astra at one two-hundredth of its blended price, about 7 points above Claude Sonnet 5.5 at one fortieth, and about 14 points above GPT-6 Luna at half. Fine-tuning moved the Nemotron model about 20 points, from about 53% to about 73%, with no change in its plotted price.
Wall St Engine, which cited the business update in a post on X at 21:57 UTC, gave the Nemotron figure as about 75%. Our measurement of the plotted point on the slide puts it below the 75% gridline, nearer 73%. Its Claude and GPT-6 Astra readings match ours.
Accuracy at classifying McCarthy’s construction requests for information (Palantir’s ‘balance accuracy’) against blended price per million tokens at three input tokens to one output token, log scale, from slide 6 of Palantir’s Q3 2026 Business Update. Values are our reading of the plotted points, approximate to about one point. Frontier prices match OpenAI and Anthropic list prices at that blend (our arithmetic); the $0.10 Nemotron price has no stated basis. One task, vendor-reported, not independent.
What Was Posted on X
PALANTIR SHOWS FINE-TUNED NVIDIA 4B AI MODEL BEATING GPT-6 ASTRA & CLAUDE SONNET 5.5 AT A FRACTION OF THE COST
In its Q3 business update, $PLTR revealed that a fine-tuned $NVDA Nemotron 3 Nano model with just 4 billion parameters outperformed much larger frontier models on a real-world customer task.
The results, based on classifying requests for information for construction firm McCarthy:
โข NVIDIA Nemotron 4B: ~75% accuracy at $0.10 per million tokens.
โข Claude Sonnet 5.5: ~66% accuracy at $4 per million tokens.
โข GPT-6 Astra: ~62% accuracy at $20 per million tokens.
Palantir says the fine-tuned models are expected to save McCarthy approximately 30,000 drafting hours annually.
Separately, NVIDIA itself is now training open AI models on its own supply chain data inside Palantir’s Ontology, spanning millions of parts and thousands of suppliers.
Jensen Huang says the partnership combines NVIDIA’s Nemotron models with Palantir’s Ontology to:
โReason, plan and orchestrate the journey from wafer to token.โ
Wall St Engine summarised the McCarthy chart from Palantir’s Q3 business update on X at 21:57 UTC on 2 November. It gave the fine-tuned Nemotron figure as about 75%; our reading of the plotted point on the slide is nearer 73%, and Palantir’s press release ties the 30,000 hours to a separate drafting project.
The Frontier Prices Match the List Prices
Palantir’s prices for the three frontier models reconcile with the published list prices. OpenAI’s pricing page lists GPT-6 Astra at $10 per million input tokens and $50 per million output tokens; at three input tokens for every output token, that blends to $20, by our arithmetic. Anthropic’s pricing page lists Claude Sonnet 5.5 at $2 and $10, which blends to $4. GPT-6 Luna, at $0.10 and $0.50, blends to $0.20.
The $0.10 Nemotron price is the one figure with no stated basis. Nvidia releases Nemotron as open models, and the slide does not say whether $0.10 is a hosting cost, a hardware cost or a provider’s price, or whether it includes the cost of fine-tuning.
Where the 30,000 Hours Come From
Next to the McCarthy logo, the slide reads: “Fine-tuned open models beat frontier models at a fraction of the cost and will save 30,000 drafting hours annually.” The Palantir press release linked from the slide splits McCarthy’s work into two projects.
The first is RFI classification, where McCarthy “fine-tuned a Nemotron 3 Nano 4B model in Foundry, and trained on labeled RFI data in the Ontology”. The second is subcontractor drafting: turning scopes of work into “Exhibit 1 contract language”. The release attaches the hours to the second project: “McCarthy delivered better performance than the frontier, and at significantly lower cost – saving 30k hours annually”. On the slide’s chart, only the RFI task is shown.
The release says McCarthy earlier benchmarked Nemotron against other models, “training on 100,000+ production RFIs to predict cost and schedule impacts before they hit a job site”. Rico Bertucci, McCarthy’s vice president of strategic initiatives, said: “We built this model on 160 years of McCarthy experience, and it outperforms the frontier at a lower cost”.
The Nvidia Supply-Chain Work Was Announced in September
The slide before it covers Nvidia’s own supply chain, which it says “spans millions of parts and thousands of suppliers”, with Nvidia “training open models on its own supply chain data inside the Ontology”. It quotes Jensen Huang, Nvidia’s founder and chief executive: “NVIDIA and Palantir are transforming this vast operational graph into sovereign intelligence — combining NVIDIA Nemotron models with Palantir’s Ontology to reason, plan and orchestrate the journey from wafer to token.”
That quote comes from a joint press release whose Business Wire item ID dates it to 10 September 2026, so the partnership itself is not new this quarter. The release says each Nvidia Vera Rubin rack has 1.3 million parts, and that the system starts with “materials allocation decisions”.
How Each Side Reads It
Palantir. The slide’s headline is “The Sovereign AI revolution is in full force.” Its subtitle reads: “We are helping our customers own their data, their models, and their advantage.” In the September release, Alex Karp, Palantir’s chief executive, said the stack is “delivering capabilities that exceed the frontier while providing alpha protection qualities unavailable otherwise”.
Nvidia. In the October release, Jeff Whitmer, Nvidia’s vice president of system supply chain, said post-trained Nemotron models help “spot potential problems earlier, act on them sooner”. The September release says the stack runs on Nvidia reference architectures, on premises with Cisco and Dell or in the cloud with Rackspace and Nebius.
The other customers. The same slide says L3Harris’s “sovereign model outperformed the frontier model and cut costs by 95% in preliminary results”, and that Tampa General Hospital’s fine-tuned models “improved performance by 7% and cut latency by 22x”. Neither has a chart.
The frontier labs. OpenAI and Anthropic are not quoted in the deck or the releases. On their own price lists, GPT-6 Astra is the dearest of the three frontier models on the chart, and GPT-6 Luna, the cheapest of them, scored about 59% at $0.20. The slide does not say how the frontier models were set up for the test.
As we read it, the business lever here is the price of a classification call: Palantir is selling the Foundry and AIP work of tuning a small open model on a customer’s own records, and the chart puts that model at the bottom of the price axis.
The Structural Read
As we read it, the chart’s largest gap is the one fine-tuning made: about 20 points on the same model at the same plotted price, larger than the gap between any two frontier models on the slide.
As we read it, the three frontier prices on the slide are list prices anyone can check, while the $0.10 Nemotron price is the one figure the deck does not explain.
Palantir press release, October 2026, quoting McCarthy’s Rico Bertucci
“No one else can pick it up, because it is tuned to how we build.”
Three Implications
FOR ENTERPRISE AI BUYERS Palantir’s chart compares a tuned open model with three frontier models on a single task; the slide does not say how the frontier models were set up, so it does not show how a tuned frontier model would have scored.
FOR OPENAI AND ANTHROPIC On the slide, the frontier entries are priced at their public list rates; GPT-6 Luna, OpenAI’s lowest-priced entry on the chart, scored about 59% at $0.20, by our reading.
FOR NVIDIA The September release says the stack runs on Nvidia reference architectures and that customers post-train Nemotron with their own data in Palantir Foundry and AIP; the deck gives no figure for the compute that uses.
The framework states the pattern as: “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize)”.
As we read it, Palantir’s slide puts that pattern in one picture: the model is an open Nvidia release, and the part kept proprietary is McCarthy’s labelled RFI data in Palantir’s Ontology. The framework was published in March 2026 and does not discuss Palantir or McCarthy.
The test is Palantir’s and McCarthy’s, published in Palantir’s investor presentation, and has not been reproduced. The slide does not give the number of RFIs tested, the definition of “Balance Accuracy”, whether the frontier models were prompted, given examples or tuned, or the basis of the $0.10 Nemotron price. Accuracy values are our reading of the chart’s plotted points, not figures Palantir printed. The slide says the work “will save 30,000 drafting hours annually”; Palantir’s press release ties the 30,000 hours to the drafting project, and neither shows how they were measured.
The slide does not say whether Palantir’s software or McCarthy’s tuning work is reflected in the chart’s prices. We did not contact Palantir, Nvidia, McCarthy, OpenAI or Anthropic.
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 one McCarthy task, Palantir’s chart puts a fine-tuned 4-billion-parameter Nemotron model at about 73% accuracy and $0.10 per million tokens, against about 66% for Claude Sonnet 5.5 at $4 and about 62% for GPT-6 Astra at $20, whose prices match the list prices. The 200-fold price gap is our arithmetic; the 30,000 hours belong to a separate drafting project; and the test is the vendor’s own.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. The chart and quotes come from Palantir’s Q3 2026 Business Update PDF and the two press releases it links, read on investors.palantir.com on 2 November 2026; list prices from OpenAI’s and Anthropic’s pricing pages the same day. Accuracy values are our reading of the vector chart; price blends and ratios are our arithmetic. We did not contact the companies.
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.
