Partly releases repair AI, claiming gains over general models with a data advantage
Interpreter 3 beat six general-purpose models in Partly’s tests. But it also had licensed vehicle-parts catalogues its rivals could not access.
Loading page…
Interpreter 3 beat six general-purpose models in Partly’s tests. But it also had licensed vehicle-parts catalogues its rivals could not access.
Listen to this story
Partly is pitching Interpreter 3 as a repair workflow system that combines reasoning with licensed OEM catalogues, not as a general model that can match it using web access alone. In company-run tests, it led six general-purpose rivals on three tasks, but the comparison gave Interpreter exclusive access to catalogue data, so it does not isolate model quality. The practical proposition is faster parts lookup and repair suggestions for shops, with an estimator retaining the final decision.
Interpreter identified the exact vehicle variant for 19 of 20 cars, versus four of 20 for the best general model; it routed uncertain cases to a person.
For parts lookup, Interpreter put the exact OEM part first 75% of the time, compared with 63% for Gemini 4 Argon and 62% for GPT-6 Astra.
Interpreter’s repair suggestions matched work later added by the shop 37% of the time, versus 17% for Argon; Partly says estimators still decide what work to approve.
Partly released Interpreter 3 on November 17, pairing its vehicle-repair reasoning model with company-run tests claiming wins over six general-purpose rivals. The tasks were concrete: identify the exact car, find its parts and spot missing repair work. A key difference sits behind the scores: Interpreter had licensed catalogue data the other models did not.
Partly ran the benchmarks in November 2026, using 250 vehicles, 257 parts and 100 repair jobs from a U.S. shop whose data had not been used to train Interpreter. Its competitors had web search. The comparison covered OpenAI’s GPT-6 Astra, Google’s Gemini 4 Hydrogen, Argon and Carbon, and Anthropic’s Claude Opus 5.5 and Claude Fable 5.2.
Interpreter drew on licensed original-equipment manufacturer, or OEM, catalogue data spanning more than 50 brands. The general models did not have that resource, Automotive World noted. The results therefore compare systems with different information access, rather than isolating the effect of repair training or reasoning alone.
That distinction is especially relevant to exact matching. Partly says one car model can have thousands of variants, while the correct brake pads may be one of hundreds of options. In its tests, general models given a vehicle identification number and the open web usually failed to name the exact variant.
Partly reports that Interpreter led every general model tested across vehicle identification, parts lookup and repair planning. The measures differ, however: naming a vehicle, ranking a part first and suggesting work a shop later adds are separate kinds of success.
Interpreter 3 adds reasoning to earlier versions, according to Partly. The company says it works through repair problems, calls specialized models and tools, and asks for clarification when a request does not add up. Its description puts those steps alongside catalogue lookup, rather than presenting the model as a parts-search box alone.
For repair shops, the model’s proposed role is to handle lookup and suggest missing work. The estimator still makes the final call.
Loading discussion...
Join the conversation
Explain whether fair comparison or real-world usefulness matters more.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.