Eleven v4 Turbo tops the TTS leaderboard at half the flagship price

ElevenLabs’ lower-cost Eleven v4 Turbo ranks first in the Artificial Analysis Provider Voice TTS Arena, ahead of the full-size Eleven v4. Turbo costs half as much, processes text faster and earned a higher rating in blind listening tests.

ElevenLabs released both models in late September, positioning Turbo as the low-latency option for phone agents, live assistants and other conversational systems. The standard v4 targets production work such as audiobooks, dubbing and games, where generation speed carries less weight. A release overview covers the shared expression controls and expanded language support.

Turbo takes a narrow lead

Artificial Analysis calculates its ratings from blind, pairwise votes. Listeners hear two samples without seeing the model names and choose the better result. An Elo system then adjusts each model’s score according to its wins, losses and opponents, using the same general approach found in chess rankings and model arenas.

Eleven v4 Turbo leads the cited snapshot with an Elo score of 1,334, followed by Eleven v4 at 1,321. The 13-point margin is small, so the ordering does not establish a large or statistically significant quality difference by itself. It does show that listeners found the faster model competitive with its more expensive sibling.

Provider prices can vary by plan, volume and hosting arrangement, so teams should confirm current rates before calculating production costs.

Latency changes the agent loop

Turbo processes 96 characters per second in the benchmark, compared with 84 for v4 and 59 for Eleven v3. Higher throughput shortens generation time for longer responses, which can reduce the pause between a user finishing a sentence and an agent beginning its reply.

Characters per second and inference latency measure different stages of the request. ElevenLabs reports median inference latency of about 100 ms for v4 Turbo, down from roughly 280 ms for v3 Conversational. Perceived delay also includes network transit, prompt processing, text generation, audio buffering and playback, so developers should measure time to first audio from the regions where their applications run.

Both v4 models track context across longer passages to adjust tone, pacing and emotion. They also build on v3’s inline expression tags, support more than 90 languages and can create an instant voice clone from about 10 seconds of reference audio. ElevenLabs describes the models and intended workloads in its v4 announcement.

The gains vary by task

Artificial Analysis also reports category, language and pronunciation results. Its Controlled Voice tests use the same cloned voice across models, reducing the influence of each provider’s default voice selection.

• Provider Voice: Turbo ranks first in Customer Service and Entertainment, and second in Assistants and Knowledge Sharing.
• Controlled Voice: It ranks first in Arabic, German, Hindi and UK English. Its Arabic, German and Hindi scores exceed those of Eleven v4.
• Pronunciation robustness: Turbo scores 90.1%, behind Eleven v4 at 91.7% and ahead of Gemini 3.8 Flash TTS at 89.5%.
• Exact character sequences: Turbo scores 73.9% on strings such as product codes and license plates, compared with 85.7% for SpaceXAI TTS.

Applications that read account numbers, confirmation codes or license plates should validate those inputs separately. Text normalization, controlled formatting and application-level repetition can reduce errors, but each strategy needs testing with the selected voice and language.

A practical shortlist for builders

Turbo offers ElevenLabs users a strong starting point for interactive agents because it combines the provider’s highest arena rating with lower latency and a 50% price reduction from v4. Standard v4 remains relevant for teams that prioritize its slightly higher pronunciation score or need to compare both models on long-form production audio.

Gemini 3.8 Flash TTS costs $16.50 per million characters, while Qwen-Audio-3.1-TTS-Plus costs $19.30. Their Elo scores sit within 50 points of Turbo, making them credible candidates for high-volume workloads with tight budgets. Language coverage, voice availability, deployment regions, rate limits and licensing terms can outweigh a modest leaderboard difference.

Production evaluations should use representative conversations rather than isolated demo lines. A useful test plan includes:

• Measuring time to first audio and total response time through the complete application stack.
• Testing names, dates, abbreviations, addresses, codes and domain-specific terminology.
• Comparing models with the same voice, output format and sampling settings where possible.
• Running multilingual tests with native speakers from the application’s target regions.
• Calculating costs from real character counts, including retries, fallbacks and repeated prompts.
• Reviewing consent, retention and usage rights for cloned voices and reference recordings.

Turbo’s leaderboard position puts pressure on flagship pricing because the $40 model currently outranks its $80 sibling while delivering higher throughput. Rankings will shift as models and vote totals change, but the current data gives developers a concrete reason to benchmark lower-latency tiers before paying for the largest model.