OpenAI rolled out two significant updates recently, upgrading ChatGPT’s default model and expanding its voice API with capabilities that move conversational AI closer to a real-time, multilingual work tool.

GPT-5.5 Instant is now the default ChatGPT model

OpenAI replaced GPT-5.3 Instant with GPT-5.5 Instant on 5 May as the default model for all ChatGPT users, citing meaningful gains in factual reliability. Internal evaluations showed the new model produced 52.5% fewer hallucinated claims on high-stakes prompts in domains such as medicine, law, and finance, and reduced inaccurate claims by 37.3% on conversations users had previously flagged for factual errors.

The update also tightens response style — fewer unnecessary follow-up questions, less overformatting, and more direct answers without sacrificing detail. OpenAI says response latency is unchanged from GPT-5.3 Instant.

Alongside the model swap, OpenAI is rolling out a new memory transparency feature called Memory Sources across all ChatGPT models. When a response is personalised using context from past chats, saved memories, or connected services such as Gmail, users can tap a Sources icon to see what information shaped the answer — and delete or correct anything outdated.

Enhanced personalisation — where the model draws more actively on past conversations and connected data — is currently available to Plus and Pro subscribers on the web, with rollout to Free, Business, and Enterprise tiers to follow. GPT-5.3 Instant remains accessible to paid users for three months via the model picker before retirement.

For enterprise deployments, the Memory Sources feature raises a practical consideration: it offers partial observability into model context but does not constitute a full audit trail. Organisations running ChatGPT alongside retrieval-augmented generation pipelines may need to reconcile OpenAI’s model-reported context with their own application logs.

Three new voice models bring reasoning and translation to the API

On 8 May, OpenAI announced three new models for its Realtime API, each targeting a different layer of live voice interaction.

GPT-Realtime-2 is the company’s first voice model built on GPT-5-class reasoning. Unlike its predecessor, GPT-Realtime-1.5, it is designed to handle complex requests mid-conversation — calling tools, managing interruptions, and maintaining context — without breaking conversational flow. Zillow, an early enterprise tester, reported a 26-point lift in call success rate on its hardest adversarial benchmark after prompt optimisation.

GPT-Realtime-Translate adds live spoken translation supporting more than 70 input languages and 13 output languages, processing speech in real time without perceptible delay. For markets across Southeast Asia — where customer interactions routinely span multiple languages — the capability removes a significant integration burden. Previously, developers building multilingual voice products needed to stitch together separate transcription, translation, and text-to-speech components. GPT-Realtime-Translate consolidates that stack into a single API session.

GPT-Realtime-Whisper rounds out the trio with streaming speech-to-text transcription, designed for low-latency captioning, meeting notes, and live documentation use cases.

Pricing for all three is available immediately through the Realtime API: GPT-Realtime-2 at $32 per million audio input tokens and $64 per million audio output tokens; GPT-Realtime-Translate at $0.034 per minute; GPT-Realtime-Whisper at $0.017 per minute.

A broader shift in how AI enters daily workflows

Taken together, these releases reflect a consistent direction from OpenAI: moving its models out of standalone chat interactions and into the tools, conversations, and workflows where work already happens. Whether through a spreadsheet, a voice call, or a multilingual customer service exchange, the underlying capability — a model that reasons, remembers, and responds in context — is the same. The infrastructure to deploy it is becoming more accessible and more granular with each release cycle.