GPT-Live-1 vs Fish Audio Drama 3: One You Can Download, One You Can Only Rent

Fish Audio Drama 3 Preview ships as 41.5 gigabytes of downloadable weights that you can point at your own GPU, inspect, fine-tune and keep running after the vendor loses interest. GPT-Live-1 ships as a URL. There are no weights to download, no local inference path, and no way to hear it at all without opening a session against OpenAI's v1/live/sessions endpoint at $0.05 a minute.

That sounds like a procurement detail. It is the decision. One of these models is an asset with a licence attached; the other is a service with a meter attached, and the two behave completely differently when your product outlives a pricing page.

What Drama 3 actually is

Fish Audio released Drama 3 Preview in mid-October 2026: a 35-billion-parameter mixture-of-experts text-to-speech model with roughly 3 billion parameters active per token, built on a Qwen3.5-35B-A3B backbone. It covers 83 languages, holds up to 16 distinct speakers inside a single request, clones voices from short references, and accepts direction in plain language — tell it a line should sound resigned, add a sound, or mark which words change, and it performs that rather than reading it flat.

The multi-speaker ceiling is the feature that matters commercially. A two-hander audio drama, a podcast with a host and a guest, a training video with a narrator and a customer — those are one API call each, not sixteen stitched-together files with mismatched room tone.

On Artificial Analysis' Provider Voice Arena, captured 29 October 2026, Drama 3 Preview sits at 1240 Elo, eleventh overall and first among open-weights entries, ahead of Breeze TTS 2 at 1228. Fish reports 93.2% pronunciation robustness, the highest claimed at launch and above the 91.7% ElevenLabs reports for Eleven v4.

What GPT-Live-1 actually is

OpenAI's model is the opposite kind of object. Since 10 September 2026 it has been callable at $0.05 per voice minute, billed by the second, with audio and text in and out, streaming and function calling supported, and structured outputs explicitly not. It is full-duplex: it hears you while it talks, and it hands heavy reasoning to a text model in the background.

It has no voice-cloning product, no multi-speaker mode, and no documented way to take a script and return a deterministic audio file. What it has is a conversation. For a phone agent that is exactly right. For a scene with four characters it is not a worse tool, it is the wrong category of tool.

The number both sides print

Here is a detail worth flagging because it will confuse anyone comparing spec sheets: Fish Audio's Drama 3 and Ant Group's Ming-3.0-TTS both report 93.2% pronunciation robustness on the same Artificial Analysis measurement. Two different labs, two different architectures, the same to one decimal place.

That is not a copy-paste error and it is not a tie. It is what a ceiling looks like — the remaining 6.8% is the hard tail of English text: alphanumeric sequences, unusual proper nouns, numbers read in context. Anyone claiming a decisive pronunciation win between the top open-weights models is reading noise. The useful question is which failure cases your content actually contains, and that requires testing your own text, not reading a leaderboard.

The licence is the part nobody reads

Drama 3 is open weights, not open source in the permissive sense. It ships under the Fish Audio Community License v1.0: free for research and non-commercial work, free for commercial use below $500,000 in annual revenue, and requiring a separate agreement above that. Third-party API hosting needs its own licence, and attribution is required.

That is a reasonable licence and a real constraint. Three consequences follow:

• Below the revenue line, self-hosting is genuinely free — you pay for electricity and GPU time and nothing else.

• Above it, you are back in a negotiation, and the "we own the weights" advantage narrows to "we can still run them while we talk".

• Reselling access to your own customers almost certainly needs the separate deal, whatever your revenue.

GPT-Live-1's licence is simpler in the way that a rental agreement is simpler: you pay per minute, you use it, you stop. No revenue threshold, no attribution clause, no obligation to disclose your customer count. That simplicity is worth something, and it is worth more the smaller you are.

What the hardware actually costs

Self-hosting Drama 3 needs roughly 48GB of NVIDIA memory in FP8, and about 41.5GB of that is weights. An H100 or a rented A100-class instance carries it; a consumer card does not, comfortably. Two framings follow.

If you already own the GPU, the marginal cost of synthesis is close to zero and the hosted price looks absurd: Fish lists $50 per million characters, is running a launch promotion at $10 per million through 12 November 2026, and Artificial Analysis records the board price as $15. A ten-minute scene of about 9,000 characters costs between nine cents and forty-five cents depending on which of those three numbers you are actually charged.

If you do not own the GPU, you are renting one anyway, and the comparison becomes GPU-hour against character-million. At a typical $2 per GPU-hour, a model that synthesises ten times faster than real time produces about ten hours of audio per hour of rental — roughly 540,000 characters at a normal speaking pace. That works out to something near $3.70 per million characters, well under every hosted price on the board, but only at sustained utilisation. Below about 30% duty cycle the rental wins nothing, and the assumption about throughput is exactly that: an assumption you have to measure on your own instance, not a figure Fish publishes.

Against that, GPT-Live-1's cost is linear and boring: $0.05 a minute, every minute, no fixed cost, no utilisation risk. Ten minutes of conversation is $0.50. Ten minutes of Drama 3 narration is roughly $0.135 at the recorded board price. Different jobs, different meters, and one of them has a fixed cost you can amortise.

Where the rental still wins

Compliance is the quiet argument for the hosted model. Drama 3's weights can be copied, fine-tuned into a competitor's product, or left running in a jurisdiction your legal team has opinions about. GPT-Live-1 gives you an endpoint, an invoice and a vendor who carries the model-side risk. For a regulated enterprise, that is not laziness; it is an insurance premium.

Latency policy is the second. Full-duplex turn-taking is a behaviour, not a checkpoint: a 0.8-second turn and 87% tool-calling accuracy are what OpenAI reports for this model, and both are vendor figures rather than independent measurements. You cannot download that behaviour with Drama 3, because Drama 3 is not trying to hear you.

And failure handling is the third. A downloadable model fails when your GPU fails. A hosted model fails when someone else's does — which is a different risk, not obviously a worse one, and one you can hedge.

Where to put the hedge

The hedge belongs behind the voice layer, not inside it. Neither of these models is served by OrcaRouter: Drama 3's weights are not on our catalogue, and GPT-Live-1 has exactly one endpoint and no compatible surface. What is routable is the reasoning leg a conversational agent depends on — GPT-6 Astra, GPT-5.6 Terra, Claude, Gemini, Qwen — reachable through a single key with automatic failover, at provider list prices passed through at 0% markup.

That is the right place to spend your redundancy budget. Voice output failing is loud and obvious; the text model behind it failing is the failure that silently ruins call quality for a week. Keep the ear locked to whichever model you chose. Keep the brain swappable.

The decision, in one paragraph each

Choose Drama 3 if you produce audio rather than conversations: scenes, podcasts, dubbing, multi-character training content, and if you are either under the revenue threshold or prepared to negotiate above it. The 35B MoE, the sixteen-speaker ceiling and the 1240 Elo are a genuinely strong package, and the weights mean nobody can take it away from you.

Choose GPT-Live-1 if the product is a live conversation and the value is in interruption handling. It is more expensive per delivered minute, it gives you nothing to own, and it does the one thing Drama 3 structurally cannot: listen while it speaks.

The two are adjacent on a spec sheet and nowhere near each other in use. Anyone who buys one expecting the other's job will conclude the model is bad, when the model was simply never asked to do the thing.