Murf AI Falcon 2 Voice Model Ranks Above OpenAI Realtime on Price

What happened

Murf AI, a text-to-speech company based in Bengaluru, has released Falcon 2. The model ranked higher than OpenAI's Realtime API on Artificial Analysis, an independent platform that tracks how AI models perform.14 It became publicly available on August 20 after a limited release to selected clients.14 Murf says Falcon 2 places ahead of OpenAI Realtime and of ElevenLabs' Flash and Turbo models on the Artificial Analysis Speech Arena, which measures how natural voices sound.1

The benchmark result got the attention, but the pricing may matter more. Murf charges $0.01 per generated minute. Bloomberg's reporting, carried by the Economic Times, says that is one-fifth of what larger rivals such as ElevenLabs charge.14 At scale, a million minutes of speech costs $10,000 on Falcon 2 and $50,000 on ElevenLabs Turbo and Flash.14 Murf also says the model starts producing audio in under 100 milliseconds, offers more than 150 voices across 35 languages, and can handle up to 10,000 calls at once without its latency slipping.114

Murf was founded in 2020 by Sneha Roy, Ankur Edkie and Divyanshu Pandey, all graduates of IIT Kharagpur. Matrix Partners and Elevation Capital back the company. It says it serves more than 10 million users at 300 companies, including Cisco, Pfizer and Air France-KLM.14

What the claim covers, and what it doesn't

The headline that a Bengaluru startup beat OpenAI is accurate, but only for a narrow comparison. All the coverage agrees that Falcon 2 ranked above some of OpenAI's and ElevenLabs' products, not above every model on the leaderboard. Bloomberg's wording is careful: Falcon 2 scored above some platforms from better-funded companies.14 Analytics India Magazine went further and ran its story under the verdict that Indian voice models are getting better but are not yet the best.1

The leaderboards support that more cautious view. Cartesia's Sonic 3.6, released on August 27, took first place in the Artificial Analysis text-to-speech arena with an Elo rating of 1273.8 Another ranking puts Google's Gemini 3.8 Flash TTS close behind at 1260.15 Even on Hindi-specific tests, Sonic 3.6 led in customer support, media and general assistant tasks, and Gemini 3.1 Flash led in content creation and education.1 Falcon 2's achievement is clearing a particular group of established commercial products, which is still significant for a company from outside the usual US labs. It did not take the top spot.

The comparison with OpenAI also needs a caveat. The "OpenAI Realtime" entry Falcon 2 beat is scored on voice naturalness in a text-to-speech arena. OpenAI's newer speech-to-speech model, GPT-Live-1, is graded on a separate scale. There it ranks second with a score of 81.5, behind Gemini 3.8 Live with Extended Thinking at 82.6, and it costs a flat $0.05 per minute plus the cost of the reasoning model behind it.8 Falcon 2 and GPT-Live-1 solve different parts of a voice agent: one only generates speech, while the other listens, reasons and speaks within a single model. In practice, Murf beat OpenAI's speech-output product on naturalness and on price. It did not beat OpenAI's full voice-agent system.

Built around Indian call centres

Falcon 2's design tells you who Murf is selling to. The company says the training data included real customer calls, support flows and booking conversations, not just scripted narration.14 It also says the model reads out the details that trip up many speech systems: numbers, currencies, dates, addresses, PIN codes and vehicle registration plates.1 Murf is aiming at call centres, banks and airlines.14 It promises data residency in 11 regions and the option to deploy inside a customer's own cloud.14

That target makes sense. India handles more than a billion phone calls a day, much of it customer support, debt collection, logistics and recruitment.3 One detailed cost analysis shows how tight the economics are. A voice agent built from separate speech-to-text, language-model and text-to-speech components costs about ₹4.68 a minute at list prices. Adding orchestration and telephony pushes that to about ₹6.13, while a human telecaller costs about ₹5.14 a minute.10 In a market where automation is only just competitive with human labour, the cost of the speech layer often decides whether a deal makes sense. That is why Murf's chief executive, Edkie, says the hard part was improving quality and latency at the same time without raising compute costs.1

Part of a wider Bengaluru push

Falcon 2 is not the only recent case of a Bengaluru company claiming wins over US labs in voice. Sarvam AI says its Saaras V3 speech-recognition model achieved a word error rate of about 19.3% on the ten most widely used languages in the IndicVoices benchmark. It says that beats OpenAI's GPT-4o Transcribe, ElevenLabs' Scribe v2, Google's Gemini 3 Pro and Deepgram's Nova-3.11 According to Sarvam, the lead grows on lower-resource Indian languages, and the model also topped Svarah, a benchmark for Indian-accented English.1112 One outlet noted that these figures were published by the company through a co-founder's post on X and have not been independently audited.12 The same caution applies to Murf, although Falcon 2's ranking comes from a third-party arena rather than the company's own charts.

The two companies work at opposite ends of the pipeline. Sarvam's model turns speech into text, while Murf's turns text into speech. Together they show a local ecosystem building its own voice components instead of putting a front end on US models. One newsletter describes the change as a move from Bengaluru "wrappers" around models built in San Francisco to systems developed in India. It credits state-subsidised compute as much as clever engineering.2 Investment is following. Smallest.ai raised a $13 million Series A, Ringg AI added $10 million from Peak XV, and Gnani.ai opened a Mumbai office, all within about a month in mid-2026.10 Sarvam itself reportedly became a unicorn after a $234 million Series B in June.3

How this fits with OpenAI, Google and Anthropic

The big labs have been releasing voice products at a fast pace. In September, Google's Gemini 3.8 Live became the leader in speech-to-speech rankings, and OpenAI made GPT-Live-1 available through its API.8 Anthropic is also competing. One voice-platform vendor writes that OpenAI and Anthropic released rival voice updates on the same day in late August.7 The same vendor says Claude Voice uses a pipeline of separate speech-to-text, language-model and text-to-speech steps, while OpenAI uses a single multimodal model.7 This is one vendor's analysis rather than lab documentation, but the architectural point matters for Murf. Pipeline systems need a separate text-to-speech engine, which opens the door for specialist providers like Falcon 2 to compete for that slot.

Open-weight models add competition from below. Kokoro-82M is available under an Apache 2.0 licence, and one ranking calls it the cheapest model at its quality level, at about $0.65 per million characters.1518 Qwen3-TTS, also released under Apache 2.0, can clone a voice from three seconds of audio.8 In India, AI4Bharat at IIT Madras has released freely licensed speech-recognition and text-to-speech models covering many Indian languages.19 One integrator's assessment is that open-source Indian text-to-speech is good enough for routine transactional calls but not for premium brand voices.19

Our reading

Falcon 2's real achievement is price, backed by a benchmark result that is credible but limited. It shows that an Indian-built model can match established commercial products on naturalness while costing a fraction as much. It does not show that Murf has caught up with the leaders. Cartesia and Google are still ahead on the arena, and the speech-to-speech systems that may eventually replace pipeline architectures are run by Google, OpenAI and xAI.815

The bigger risk is pricing pressure. Gemini 3.8 Live charges $0.005 per minute for audio input and $0.018 per minute for output.8 At those rates, the big labs can match Falcon 2 on cost. One analyst also warns that most of the value in Indian voice AI tends to go to buyers, telecom carriers and model owners, not to application companies.10 Murf's competitive edge comes from handling code-mixed speech, Indian number formats and very large call volumes. Rankings change from week to week, so that edge has to stay ahead.16 For now, Falcon 2 shows that capable, low-cost voice models are no longer coming only from the US. Whether that turns into a durable business will depend less on any single leaderboard placement than on how quickly global labs make their own voice models cheaper and better at Indian languages.