Replika says it dropped OpenAI's GPT-3 for its own smaller model

The design

Replika's own account of a major technical transition appears in an October 2021 post from the company's AI team, archived here after the live page was later removed. The post states that 'at some point in Replika's history, we used the well-known GPT-3 as a generative model,' identifying it as OpenAI's large neural network, and that 'Replika became one of the first partners of OpenAI back in March 2020.' It says the company 'fine-tuned the GPT-3 model with 1.3B parameters on our dialogs' and 'deployed the model for millions of our users,' reporting that by February 2021, 71 percent of conversations using that model 'made people feel better.'

What the evidence says

OpenAI's own June 2020 announcement of the GPT-3 API states it entered private beta that month; Replika's post places its own access 'back in March 2020,' three months earlier, a discrepancy neither document explains. The 2021 post is otherwise Replika's own uncorroborated account of its engineering history; no OpenAI statement here names Replika as an early partner. The post goes on to say the company 'decided to move from GPT-3 to our own generative model,' a smaller network of 'only 774M parameters' that it says 'exceeded GPT-3 in terms of the positive session fraction.' Retrieved 16 September 2026, the current help-center article 'How does Replika work?' states the app now runs on 'a sophisticated system that combines our own and third party Large Language Models,' consistent with moving away from sole reliance on one vendor, though it does not name GPT-3 or confirm the 2021 figures.

What it asks of people

The 2021 post asks a technically literate reader to accept one self-reported metric, the positive session fraction, as the measure by which one model is judged better than another, without describing how the underlying feedback was collected or validated. Neither document discloses what proportion of a user's conversation today runs on Replika's own model versus a third-party one.

Privacy and safeguards

The 2021 post does not address what happened to conversation data processed through the OpenAI-hosted GPT-3 model during the period described, a question the company's current privacy policy, covered elsewhere on this site, addresses only for its present-day third-party arrangements. Moving generation in-house, as described, would reduce the number of external parties handling conversational content for that portion of processing, though the post frames the change entirely as a product and control decision, not a privacy one.

• What happened to the earlier GPT-3-processed conversation data once Replika's own model took over?
• Does the current mix of in-house and third-party models, as the help center describes it, still include any GPT-3 successor?
• Would the positive session fraction metric described in 2021 be considered a rigorous measure by outside researchers studying companion chatbots?

The move described here predates the site's coverage of Replika's later in-house model work and shows a company narrating its own dependency on a third-party API as a limitation it engineered its way out of, in its own words, without independent confirmation of the metrics it cites.