Text-to-speech costs: budget for finished minutes and retakes
For our standard 60-minute narration benchmark, generating 54,000 characters consumes $2.16 of ElevenLabs v4 Turbo API capacity. A fresh Cartesia Pro subscription is $5, while Google Cloud Neural2 can cost $0.00 if your monthly free allowance is intact. Every provider structures quotas and retakes a little differently, so the most reliable voice budget is one you build around your own scripts and the accounts you will actually use. Let's walk through it together.
Count the narration you deliver and the speech you generate
There's a gap between what your audience hears and what your speech engine actually generates. A finished minute is the polished, published audio. Getting there usually means generating takes you'll throw away, testing alternate pronunciations, and re-reading copy after an edit. If you buy only enough characters for your raw script, your production will stall the first time you need a retake. Ask me how I know.
We use 750 script characters per finished minute and a 1.2 generation factor as editable planning assumptions, not an immutable industry rule. That works out to 900 generated characters for every 60 seconds of finished audio. Treat these as a starting point, and calibrate them with your own voice talent tests, reading pace, and script density.
Small production: sixty finished narration minutes
For an hour of finished audio, our planning model needs 54,000 generated characters (60 × 750 × 1.2 = 54,000). At public pay-as-you-go developer rates, here's how that looks:
- ElevenLabs v4 Turbo API: 54,000 characters at $0.04 per 1,000 characters = $2.16 ($0.036 per finished minute).
- ElevenLabs v4 API: 54,000 characters at $0.08 per 1,000 characters = $4.32 ($0.072 per finished minute).
- Google Cloud Neural2: 54,000 characters sits comfortably inside Google's monthly free quota of one million characters, so you'd pay $0.00 in synthesis charges if your account has an unused free tier.
- Cartesia Pro: The base plan is $5.00 monthly with 100,000 included credits, which easily covers the 54,000-character requirement.
Pay-as-you-go developer accounts come with no required Creator fee or recurring seat commitments, which makes them a nice fit for one-off or sporadic projects. If you want a closer look at ElevenLabs subscription tiers versus character meters, our ElevenLabs finished minute guide has the full breakdown.
Pricing Google Cloud Neural2 with and without the free allowance
Google Cloud Text-to-Speech gives you 1,000,000 characters of Neural2 synthesis each month at no charge, per billing account. The practical catch is that the free tier is shared across every service, developer key, and automated workflow in your Google Cloud organization. If another team's pipeline eats the allowance first, that lovely $0.00 bill disappears, so I wouldn't count on it forever.
At the standard listed rate of $16.00 per million characters ($0.000016 per character), generating 54,000 characters with no free pool left costs exactly $0.864. That's roughly $0.014 per finished minute. Neural2 is reliable and predictable for programmatic audio, though creative directors often find ElevenLabs delivers more emotional inflection for narrative work.
High volume: 10,000 finished minutes across ElevenLabs, Google and Cartesia
Once you scale up, say to voicing a big catalog of 10,000 finished audio minutes, the math changes quickly. At 750 script characters per minute and a 1.2 generation factor, the workload comes to 9,000,000 generated characters. Here's how the providers stack up at this volume:
| Provider & Tier | Pricing Formula | Total Production Cost |
|---|---|---|
| Google Cloud Neural2 (with free pool) | (9M − 1M free) × $16/M | $128.00 |
| Google Cloud Neural2 (no free pool) | 9M × $16/M | $144.00 |
| ElevenLabs v4 Turbo PAYG | 9M × $40/M | $360.00 |
| ElevenLabs v4 PAYG | 9M × $80/M | $720.00 |
| Cartesia Pro + overage | $5 + (9M − 0.1M) × $65/M | $583.50 |
| Cartesia Startup + overage | $49 + (9M − 1.25M) × $45/M | $397.75 |
| Cartesia Scale + overage | $299 + (9M − 8M) × $38/M | $337.00 |
Cartesia's current monthly plans include 100K, 1.25M and 8M credits at the base fees shown. What to keep in mind: overages have to be explicitly enabled on the billing dashboard, otherwise API requests get rejected once your allowance runs out. Cartesia's credit pool is also shared across speech-to-text and text-to-speech, so any transcription workloads you already run will shrink your available synthesis capacity. Commercial deployment requires an active paid tier.
At 9M characters, Cartesia Scale is the lowest-cost of the three Cartesia options at $337.00, compared to $397.75 on Startup and $583.50 on Pro. That said, unit rates shouldn't be the only thing driving your stack choice. Run blind listening tests with your real terminology, proper names, and emotional registers, and make sure the generated speech clears your quality bar.
Where OpenAI fits now: deprecations and real-time audio
OpenAI's official deprecation schedule lists tts-1, tts-1-hd and gpt-4o-mini-tts for removal on January 6, 2027, with active migrations pointing to gpt-realtime-2.1-mini and newer audio endpoints. So comparing against legacy text-character pricing (like the historical $15 per million character models) isn't a solid foundation for a brand-new build.
Modern multimodal audio endpoints are billed on a mix of input text tokens, audio input tokens, and audio output tokens under current model pricing. To find your real-world cost, run a representative pilot, record the exact token consumption across all modalities, and divide the total billing charge by the number of accepted finished audio minutes.
For example, if a pilot run incurs a $1.80 token charge and yields twelve minutes of broadcast-ready audio, your measured generation cost is $0.15 per finished minute. Benchmark your own script and caching behavior before you stretch that number across a large project budget.
The one-billion-character planning check
At enterprise scale, such as a big library migration of 1,000,000,000 generated characters, public list prices give you a clear picture of relative vendor economics. At this volume, Google Cloud Neural2 totals $16,000 (or $15,984 if a single 1M free pool is credited), ElevenLabs v4 Turbo totals $40,000, and full ElevenLabs v4 totals $80,000.
Extending a fresh Cartesia Scale pool at its documented overage rate gives: $299 + (1,000M − 8M) × $38/M = $37,995. These are public-rate calculations, so at this scale you'll want to talk to vendor sales about custom volume commitments, concurrency guarantees, and dedicated SLA contracts.
At an average of 900 generated characters per finished minute, one billion characters corresponds to approximately 1,111,111 finished minutes (over 18,500 hours of audio). One more reality check: at enterprise volume, audio storage, network egress, quality control, and human editorial oversight can easily outweigh the raw API synthesis bill.
Before you buy
Which provider is the cheapest option?
Going purely by raw character pricing, Google Cloud Neural2 gives you the lowest invoice. Voice realism and listener preference vary a lot, though, so run blind listening tests with your team before you decide on unit price alone.
Does the free quota mean my project is free?
Only if your monthly character volume stays under 1,000,000 characters and no other application in your Google Cloud organization has already used up that quota. If you've already used your free pool, enter zero in the calculator to see your true billing rate.
How should I configure the worksheet?
Start with your target finished minutes, script character density, and a realistic revision factor. The interactive worksheet below defaults to the 10,000-minute Neural2 scenario with a fresh 1M free pool, which yields $128. If you're modeling paid credit subscriptions, adjust the plan fee and included capacity to match.
Building interactive voice agents or telephony stacks? Take a look at our Vapi vs Retell vs Bland voice agent benchmark.
Recalculate the Neural2 character example
10,000 finished minutes × 750 script characters × 1.2 generation factor = 9M generated characters. Assumes the full eligible 1M Neural2 free pool remains, giving $128. Set free characters to zero for $144. Excludes tax, review, storage and delivery.
Compare the model rates for your workload →