OpenAI has released three new voice models that fundamentally change how developers build voice agents, bundling capabilities that previously required stitching together multiple vendors into single, cheaper tools. The release includes GPT-Realtime-2 with GPT-5-class reasoning, GPT-Realtime-Translate covering 70+ languages, and GPT-Realtime-Whisper for speech-to-text, priced aggressively enough to reshape the competitive landscape.

What Makes These New Models Different From Existing Voice AI?

For the past year, companies building voice agents have assembled their systems like a patchwork quilt. They would grab Whisper or Deepgram for transcription, ElevenLabs or Cartesia for text-to-speech synthesis, GPT-4 or Claude for reasoning, and then write custom code to handle turn-taking and interruptions. Each handoff between components introduced latency, complexity, and cost.

GPT-Realtime-2 changes this equation by handling audio in and audio out within a single model, with reasoning happening inside the audio loop rather than between separate steps. This architectural shift eliminates the delays that made voice interactions feel unnatural. The model can now say "let me check that" while calling tools, so users do not sit through awkward silence. It can fire multiple backend requests simultaneously and narrate which one is running. If something fails, it surfaces the error gracefully instead of freezing the conversation.

The context window has expanded to 128,000 tokens, up from 32,000, which means longer conversations and complex workflows no longer require external state management. Developers can also adjust reasoning effort on a knob with five settings: minimal, low, medium, high, and xhigh, with low as the default to keep response times snappy.

How Are Real Customers Actually Using These Models?

The performance gains are not theoretical. Zillow, the real estate platform, reported a 26-point lift in call-success rate on its hardest test cases, jumping from 69% on the prior model to 95% on GPT-Realtime-2. BolnaAI, a voice AI company building for Indian languages, reported 12.5% lower word error rates on Hindi, Tamil, and Telugu using the translation model. On OpenAI's own benchmarks, GPT-Realtime-2 at high effort scored 15.2% higher than its predecessor on audio reasoning tasks, with even larger gains at maximum effort.

OpenAI's launch customer list reads like a who's who of voice-agent deployment: Zillow, Glean, Genspark, Bluejay, Intercom, Priceline, and Foundation Health for the realtime model; BolnaAI, Vimeo, and Deutsche Telekom for translation. These are not startups experimenting in sandboxes; they are production systems handling real customer interactions.

Why Are Competitors Suddenly Worried?

The pricing is where the competitive pressure becomes undeniable. GPT-Realtime-2 costs $32 per million audio-input tokens and $64 per million audio-output tokens. GPT-Realtime-Translate runs at $0.034 per minute. GPT-Realtime-Whisper costs $0.017 per minute. That last number is roughly a third of a cent per minute for translation, undercutting most enterprise translation pipelines by a wide margin while bundling latency and language coverage that cost-conscious deployments have historically had to sacrifice.

ElevenLabs, the most-funded pure-play voice company in the market, prices its voice agents on a per-minute model that bundles synthesis with model inference. Deepgram sells streaming transcription as a standalone primitive. When OpenAI's bundled model is also doing the reasoning, the arithmetic for buyers becomes much harder. ElevenLabs raised its Series D in February at an $11 billion valuation explicitly on the agent thesis; Deepgram has been moving in the same direction. Both now face the question of whether they can hold their market position or must accelerate their own integrated stacks.

How to Evaluate Voice AI Models for Your Use Case

• Benchmark Performance: Test models on your specific language pairs and use cases rather than relying on generic benchmarks. Zillow's 26-point improvement came from real-world adversarial testing, not lab conditions.
• Total Cost of Ownership: Calculate the full cost of your current stack, including transcription, synthesis, reasoning, and custom integration work. Compare that to bundled pricing from single vendors.
• Latency Requirements: Determine whether your application needs minimal latency for natural conversation flow. Reasoning-effort settings let you trade speed for accuracy, so understand your tolerance for each.
• Language and Localization Needs: If you serve multiple regions, evaluate language coverage and error rates in your target languages. BolnaAI's results on Hindi, Tamil, and Telugu show that performance varies significantly by language.
• Integration Burden: Assess the guardrails, evaluation, escalation, and analytics work required before going live. OpenAI's models handle the audio reasoning, but compliance, brand voice, and tool-call observability still require developer work.

The competitive question now is which platform reduces integration burden fastest. OpenAI's bet is that doing audio reasoning inside one model is more defensible than stitching three vendors together. Whether ElevenLabs, Deepgram, and others can hold their wedge depends on how quickly they push their own integrated stacks. The next quarter will be the first time this comparison is made on production workloads rather than on demos.

For developers and enterprises evaluating voice AI, the immediate test is available in OpenAI's Playground tab and through SDK calls. The price card and the benchmarks suggest OpenAI is not waiting for the market to catch up.