On October 15, 2026, OpenAI introduced GPT-Live-1.1 and a smaller companion model, GPT-Live-1.1 Quip, together with 13 new voices in ChatGPT, according to The Neuron's October 16 review of the release. The update refines the voice model that OpenAI first announced in July and put into its API in September. It is an incremental release, and the interesting part is less the voices than the gap between what OpenAI describes and what it has published numbers for. This article separates the two.
OpenAIWhat was released
Per The Neuron, GPT-Live-1.1 replaces GPT-Live-1 as the default voice model for Plus and higher tiers, including Codex and Dots. Quip replaces GPT-Live-1 mini for other users. Availability varies by account and region. The new voices include options for French, German, Hindi, Japanese, Korean and Mexican Spanish, plus an English voice called Bubbie; the article does not list all 13.
OpenAI's stated changes are qualitative:
- Pronunciation is more natural across several languages.
- The model responds better to spoken instructions, which OpenAI calls steerability.
- Search status messages are more varied.
- Delegation, where a demanding task is handed to another model while the conversation continues, feels more natural.
OpenAI said API access to GPT-Live-1.1 would follow. For developers, that matters because GPT-Live-1 was launched in the API on September 10, 2026 at $0.05 per minute for the voice layer, as covered in our GPT-Live-1 analysis. The Neuron's article gives no pricing for 1.1, so any assumption that the rate carries over is unconfirmed.
What OpenAI has and has not quantified
| Claim | Published evidence, per The Neuron |
|---|---|
| More natural pronunciation | Not quantified; no language-specific recognition or error rates |
| Better response to spoken instructions | No comparative results on how often revised instructions are followed correctly |
| More natural delegation | Described, not measured |
| Spoken approvals | Not supported; actions needing approval still require an on-screen response |
| Safety, real-world audio | No statistically significant regressions versus predecessors |
| Safety, synthetic tests | Significant regressions in sexual content, mental health, emotional reliance and gore |
A release note can say a voice feels more natural. Only a measured error rate can say for whom.
Analysis
The safety rows are the most informative. OpenAI's real-world audio tests found no statistically significant regressions compared with earlier models, but its synthetic tests did show significant regressions in four categories: sexual content, mental health, emotional reliance and gore, and OpenAI says more research is needed. The Neuron notes that neither set of evaluations shows how often a user would meet a harmful response in everyday use. Two reasonable readings coexist: synthetic tests are built to stress a model with difficult prompts, so a regression there does not translate directly into everyday frequency; but the categories named, mental health and emotional reliance in particular, are areas where voice can feel more personal than text, which is why the question is worth tracking.
Why voice upgrades are hard to compare
Speech models resist the head-to-head tables that text models get. A spoken exchange involves recognition accuracy across accents, background noise, turn-taking, latency and the quality of the delegated answer, and each is measured differently. Without published error rates by language, a listener's impression of a new Hindi or Korean voice is the only comparison available. That is a transparency gap rather than a criticism, and figures such as the Full Duplex Bench gain quoted at the GPT-Live-1 launch remain vendor-reported.
The cost arithmetic is simple and worth doing, with the caveat that it uses the GPT-Live-1 API rate and not a 1.1 price.
Delegation is the architecture to watch
The detail in OpenAI's changes with the longest shelf life is delegation. Voice front ends that hand hard tasks to a stronger model while keeping the conversation going separate how a system sounds from how well it reasons. That split lets a product upgrade either side independently, and it is the same pattern covered in the GPT-Live-1 launch analysis. It also explains one limitation in this release: spoken approval is not supported, so an action that needs a go-ahead still needs a tap. For anything that moves money or sends messages, a confirmation step outside the voice channel is a sensible control rather than a gap, and teams building voice agents can adopt it regardless of vendor.
Reading the release
- What is solid: the release date, the model names and replacements, the 13 new voices, and the safety test outcomes as OpenAI reports them.
- What is unquantified: pronunciation, recognition and instruction-following gains, and the real-world frequency of harmful responses.
- What to watch next: API availability and pricing for 1.1, publication of per-language recognition rates, and the follow-up research OpenAI says is needed on the synthetic-test regressions.
Voice quality will keep changing quickly across vendors. Using a workspace that lets you compare voice and text models side by side, as Metir does for chat models, is one practical way to judge an upgrade on your own content rather than on release notes.
Sources:
- ChatGPT adds 13 voices with GPT-Live-1.1 (The Neuron, Oct 16, 2026)
- Everything That Happened in AI Today, Monday, October 19 (The Neuron)
- OpenAI launches GPT-Live-1 for full-duplex voice agents (TestingCatalog)
- OpenAI GPT-Live-1 API: Full Duplex Voice, Telephony, Pricing (DataStudios)
Image credits
Header and in-body image: The Pioneer Building, San Francisco (2019), by HaeB, Wikimedia Commons, CC BY-SA 4.0 (source); illustrative.