Deel Pangaea: small HR and payroll AI models vs. frontier LLMs
Deel's Pangaea HR and payroll AI models claim near-frontier accuracy at a fraction of the cost. The savings figure has no clear baseline. Here's what to ask.
Deel launched Pangaea, a pair of specialist AI models for HR and payroll, on October 21. The pitch: a model built on a 9-billion-parameter base answers employment-law questions almost as well as the frontier models, about twice as fast, for a fraction of the cost. It's the clearest sign yet that the vendors in your back office will stop renting the biggest model for every question. Read the cost claim carefully, though.
What actually happened
From VentureBeat's report:
- Two models: Pangaea K1 handles employment rules and procedures. It's adapted from an unnamed 9B open-weights model, trained on 43,500+ pages of reference material kept by Deel's specialists. Pangaea P1 handles payroll and builds on Nvidia's Nemotron family, with extra training by AI21 Labs.
- Data: Deel says neither model learns from identifiable employee information or customers' actual payroll records.
- Benchmarks (Deel's own): On Deel's 601-question EmBench Global Work test, K1 scored 70. The top score was 75; Claude Opus 5.5 scored 74. K1's median response time was 7.7 seconds, against 14.8 to 28.4 seconds for the four models above it.
- The cost claim: Deel's materials say K1 costs about one-forty-ninth as much per answer, roughly 98% less. VentureBeat found Deel didn't clearly identify the comparison model or say whether the figure includes training, hosting, or expert review.
- Availability: Pangaea runs inside Deel AI, included in existing subscriptions with no per-question charge. No downloadable weights, public API, or public pricing.
Why specialist HR and payroll AI models matter for your business
You won't pay for this directly, but you'll feel it. Pangaea is bundled. What changes is your vendor's margin, and so whether "AI included" stays included at renewal.
The benchmark is the vendor's own. Deel built the test and the model. A 70 against a 74 might be fine for "how many PTO days in Ontario." It's not fine for a termination letter. Ask which questions route to Pangaea and which go to a general model.
"98% cheaper" needs a denominator. Cheaper than what, per what, including what? If a vendor can't answer, the number is marketing.
Routing is the real product. Deel picks the model per question. That's the pattern we build for clients: small model for lookups, big model for reasoning, a human for anything that ends up in a contract.
Key takeaways
- Deel's Pangaea K1 and P1 are specialist HR and payroll models inside Deel AI
- K1 scored 70 vs. a top 75 on Deel's own benchmark, at roughly half the latency
- The ~98% cost-savings claim has no clearly stated comparison or scope
- Ask vendors which questions hit which model, and keep humans on anything legal
Paying frontier-model prices for lookup questions? We build model routing that sends simple questions to small models and hard ones to big ones, with logs that show what each answer cost. Run the numbers on your AI spend, or see how we build it.
Sources: VentureBeat: Deel launches Pangaea AI models for HR and payroll, Deel AI.
- #deel
- #payroll
- #hr-software
- #specialist-models
- #ai-costs
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Stripe Treasury for agents: ChatGPT and Claude can pay invoices
Stripe Treasury for agents lets ChatGPT and Claude check balances, pay invoices, and move money over Stripe's MCP server. Approval is on by default. Keep it on.
Read itKognitos AI Employees: 61 finance roles run on 515 SOPs
Kognitos launched AI Employees for finance and banking: 61 roles, 515 approved SOPs, 133 monitors, and MCP access. The lesson: write the procedure first.
Read it