ChatGPT Voice Mode Runs on April 2024-Vintage GPT-4o Model

What Happened

OpenAI's ChatGPT Advanced Voice Mode runs on a GPT-4o-era model with an April 2024 knowledge cutoff — not the company's current frontier models — according to a note published April 10, 2026 by Simon Willison, who surfaced the detail by querying the voice interface directly. Willison flagged the finding as "non-obvious to many people," given the intuitive assumption that a conversational interface would surface the most capable available model.

The observation was prompted by a post from Andrej Karpathy, who framed the divergence in blunt terms: OpenAI's free Advanced Voice Mode "will fumble the dumbest questions in your Instagram Reels" while the company's paid Codex model "will go off for 1 hour to coherently restructure an entire code base, or find and exploit vulnerabilities in computer systems."

Why It Matters

The capability gap between OpenAI's consumer voice surface and its developer-facing API products is now significant enough that users interacting exclusively through voice may be forming inaccurate mental models of AI capability — in both directions. They may underestimate what current frontier models can do in code and reasoning tasks, and overestimate what voice interfaces will reliably deliver.

Karpathy identified two structural reasons the coding and agentic product lines have pulled ahead:

• Verifiable reward functions: Code domains offer binary feedback signals — unit tests pass or fail — making them directly amenable to reinforcement learning. Natural conversation and voice quality are substantially harder to score automatically.
• B2B revenue concentration: The highest-value enterprise contracts center on coding assistants and agentic workflows, directing a disproportionate share of engineering headcount toward those products.

This creates a compounding dynamic: the modalities that are easiest to train with RL are also the ones with the largest paying customer base, accelerating their improvement relative to consumer voice features. Advanced Voice Mode, which Willison describes as "slightly orphaned," appears to be on the losing end of that prioritization calculus — at least for now.

For CTOs evaluating AI vendor capability, the practical implication is that product demos conducted through voice interfaces may systematically underrepresent what the underlying API can deliver. Procurement decisions and internal capability assessments should specify the exact model endpoint being tested, not the interface layer.

The Technical Detail

OpenAI's real-time voice pipeline uses a native multimodal model that encodes and decodes audio directly, rather than chaining speech-to-text → LLM → text-to-speech. This architecture reduces latency and preserves prosodic information, but it also means the voice-optimized model is a distinct artifact from the text-only or vision-capable frontier checkpoints.

Querying the voice interface for its knowledge cutoff date returns April 2024 — consistent with the original GPT-4o release window. OpenAI's current text API, by contrast, exposes models with knowledge cutoffs in early-to-mid 2025 depending on the variant. The delta is at minimum 12 months of training data and an unknown number of post-training alignment iterations.

Karpathy's framing implies the gap is not merely about knowledge cutoff but about RLHF signal quality: tasks with hard-to-specify reward functions (voice naturalness, conversational coherence) receive less RL-driven improvement per training cycle than tasks with automated verifiers (code execution, test suites, formal proofs).

What To Watch

• OpenAI voice model refresh: No public roadmap commitment exists for upgrading Advanced Voice Mode to a post-April-2024 checkpoint. Watch for any changelog entries or model card updates on the OpenAI platform status page within the next 30 days.
• Competitive pressure from Google: Google's Gemini Live uses Gemini 2.0 Flash as its voice backbone — a more recently trained model. If OpenAI's voice gap becomes a visible marketing differentiator for Google, that could accelerate an internal prioritization shift.
• Karpathy's broader thesis: His framing of RL-amenable domains as the primary driver of capability improvement has implications for which product categories will see the next step-changes. Watch agent and code-execution benchmarks (SWE-bench, Aider leaderboard) for continued divergence from voice and creative writing evals.
• Enterprise voice adoption: If the B2B revenue argument holds, enterprise voice assistant use cases — call center automation, meeting summarization — may eventually generate the verifiable signal needed to close the training gap. Vendor announcements in this space in Q2 2026 will be a signal of strategic intent.