ElevenLabs v3 Prompts: Audio Tags, Examples & Credit Costs (2026)

ElevenLabs v3 prompts are plain-text cues — audio tags like [whispering], ALL CAPS for emphasis, and punctuation for pauses — that steer the model’s dramatic delivery. v3 supports 70+ languages with a 5,000-character (about 5-minute) limit per request. Every character bills 1 credit, so a 2,000-character prompted script costs 2,000 credits. Checked 4 Nov 2026.

What do v3 prompts actually control?

A v3 prompt is not a separate settings panel — it is the script itself, written with cues the model reads as delivery instructions. Audio tags in square brackets tell the voice how to perform a line ([whispering], [shouting], [sighs]); capitalization tells it which word to stress, because the official docs say capitalization increases emphasis; punctuation tells it where to pause and how a sentence should feel. Because the cues are plain characters, they work identically in the web app and through the API, and they cost nothing beyond the 1 credit per character the text already bills.

They do not control everything. v3 does not support SSML — the docs state plainly that Eleven v4 and Eleven v3 do not support SSML break tags, and phoneme-level SSML tags only work with the eleven_flash_v2 model. Prompts also cannot set exact pause lengths in milliseconds or lock a precise loudness level; those values are not published.

One honest note: ElevenLabs now lists v3 as its previous generation model, with Eleven v4 as the flagship. v3 (model ID eleven_v3) is still available and documented, and the prompting techniques on this page are the ones the docs describe for v3 — the same audio tags, punctuation, and dialogue formatting that v4 uses.

Audio tags: the v3 prompt vocabulary

Audio tags are the heart of v3 prompting. The official prompting guide groups them into voice-related tags, sound effects, and one-off unique tags. Voice-related tags shape how a line is delivered: [laughing], [chuckles], [whispers], [whispering], [shouts], [shouting], [sighs], [sarcastic], [curious], [excited], [crying], [nervously], [playful]. Sound-effect tags drop produced sounds into the read: [applause], [clapping], [explosion], [gunshot], [door slams], [footsteps], [phone ringing], [thunder], [triumphant music]. A few unique tags do stranger jobs, like [strong French accent] or [sings].

Tags sit inline in the script, right where the delivery should change: “We won. [whispering] At least, that’s what they told us.” The tag itself is billed like any other characters — [whispering] is 12 characters, so about 12 credits — which makes tags almost free next to the script around them.

Two cautions from the docs. First, tags work best when the voice matches them — a tag cannot force a calm narrator voice into a convincing scream. Second, professional voice clones are not fully optimized for v3; the docs recommend v4 for cloned voices, so run tag-heavy scripts on stock or designed voices for the most predictable results.

Write a v3 prompt in 5 steps

Five steps from a flat script to a performed one. Writing is free — only generation bills.

1. Step 1 — Pick the voice and draft the plain script. Choose a voice whose natural energy is close to the scene: a tag nudges delivery, it does not recast the actor. Write the scene as clean sentences first. Cost of this step: $0 — no credits are spent until you generate.
2. Step 2 — Insert audio tags where delivery should shift. Mark entrances and emotional turns: [sighs] before the bad news, [excited] before the reveal. Keep tags short and sparse; one tag per shift is the pattern the docs’ own examples follow. Cost of this step: about 12–15 credits per tag at generation — ten tags add roughly 130 characters to the bill.
3. Step 3 — Capitalize the words that must land. The docs say capitalization increases emphasis, so “this is HUGE news” hits harder than “this is huge news.” Reserve all-caps for one or two words per paragraph — a paragraph of shouting emphasizes nothing. Cost of this step: $0 extra — the characters were already in the script.
4. Step 4 — Punctuate the pauses. Ellipses add pauses and weight to the words around them; em dashes create interruptions; short sentences create urgency. Read the script aloud once and put a mark everywhere you naturally breathe. Cost of this step: a few characters per mark — a dozen ellipses add about 36 credits.
5. Step 5 — Generate and iterate. Send the script to v3 and listen for whether each cue landed. The 5,000-character request cap means the longest single v3 generation costs 5,000 credits; a 2,000-character scene with tags runs about 2,050 credits, roughly a fifth of the free plan’s 10,000 monthly credits. If a tag misfires, strengthen the cue — a clearer tag, cleaner punctuation — rather than stacking more tags on the same line.

What v3 prompts cost in credits

Prompting has no separate price — every character of script and markup bills 1 credit. Typical jobs work out like this:

The practical takeaway: a month of the free plan covers about five maximum-length v3 requests, or roughly twenty 500-character tag experiments. If you iterate a lot, keep experiments small — tags are cheap to test at 500 characters and expensive to discover at 5,000. For the plan quotas behind these numbers, see the ElevenLabs pricing breakdown.

5 v3 prompting rules that work

• One cue per shift. A line with [whispering], caps, and ellipses all at once gives the model conflicting directions — pick the strongest one.
• Tags follow the voice, not the other way round. Cast a voice close to the scene first; tags refine a performance, they do not create one.
• Anchor dialogue in layout. v3 supports natural multi-speaker dialogue — give each speaker their own line and keep tags inside the right speaker’s lines.
• Test at 500 characters. A short passage exercises every tag and emphasis trick for about 520 credits before the full script commits thousands.
• Format before generating. Every post-generation rewrite re-bills the whole character count, tags included.

Common problems and fixes

The tags are read aloud as words

The tag is malformed — v3 expects square brackets with no extra spaces inside, like [whispering] rather than [ whispering ]. Retype the tag exactly and regenerate a short test passage before the full script.

The emphasis lands on the wrong word

Something else in the line is stealing the stress — another capped word, an exclamation mark, or a tag pulling the other way. Strip the line to a single emphasis cue and rebuild; one cue per idea is the rule that fixes most misfires.

The scene sounds flat despite the tags

The voice is miscast. Tags nudge delivery but cannot turn a calm narrator into an action hero — recast with a voice whose natural energy matches the scene, then re-apply the tags. Our prompt formatting guide covers matching style to the job in more depth.

A long script gets cut off at the end

That is the v3 request cap, not a bug — v3 allows 5,000 characters (about 5 minutes) per request. Split the script at a scene break and generate in parts; v4 allows 10,000 characters per request if you need longer single takes.

Who should NOT prompt v3

Skip v3 prompting if your work is plain narration — articles, documentation, course lessons read straight. Clean text with honest punctuation needs no audio tags, and every tag you add is a character billed for no audible gain. Also skip it if you need SSML pipelines or single takes longer than 5 minutes — those are v3’s hard edges, and no clever prompting moves them. For straight narration, the text-to-speech settings page covers the sliders that matter more.

Frequently asked questions

What is an ElevenLabs v3 prompt?

An ElevenLabs v3 prompt is a script written with delivery cues — audio tags like [whispering], capitalization for emphasis, and punctuation for pauses — that steer v3’s dramatic delivery. There is no separate prompt box; the cues live inline in the text you generate. Every cue is billed as characters at 1 credit each, and the same techniques work through the API.

Which audio tags work with v3?

Voice-related tags ([laughing], [whispers], [sighs], [sarcastic], [excited], [crying]), sound-effect tags ([applause], [thunder], [door slams]), and unique tags like [strong French accent] all work with v3, per the official prompting guide. Place tags inline where the delivery should change. Tags are plain characters, so a 12-character tag costs about 12 credits at generation.

Does v3 support SSML?

No. The official docs state that Eleven v4 and Eleven v3 do not support SSML break tags, and SSML support overall is experimental. Phoneme-level SSML tags only work with the eleven_flash_v2 model. For emphasis on v3, use capitalization, ellipses, and audio tags instead — those are the documented mechanisms.

How many characters can a v3 prompt be?

5,000 characters per request, which the docs equate to about 5 minutes of audio. Longer scripts must be split at scene breaks and generated in parts. At 1 credit per character, a maximum-length request costs 5,000 credits — half the free plan’s 10,000 monthly credits, or about a sixth of Starter’s 30,000.

Do v3 prompts cost extra credits?

No — prompting has no separate price. Audio tags, caps, and punctuation are billed as ordinary characters at 1 credit each. A 2,000-character scene with ten tags bills about 2,050 credits. The real cost risk is iteration: every regeneration re-bills the full script, so test tags on a 500-character passage first.

Should I use v3 or v4 for prompted scripts?

v4 is the current flagship: it allows 10,000 characters per request, and the docs recommend it for professional voice clones. v3 remains available with the same prompting grammar — audio tags, caps, punctuation — and suits dramatic dialogue in 70+ languages. Pick v4 for longer takes and cloned voices; v3 is fine for short dramatic scenes on any plan.

Sources checked

Model facts, the 5,000-character limit, 70+ languages, and multi-speaker dialogue: official ElevenLabs models documentation (elevenlabs.io/docs/overview/models), checked 4 Nov 2026. Audio tags, emphasis rules, and SSML limits: official prompting best-practices guide (elevenlabs.io/docs/overview/capabilities/text-to-speech/best-practices), checked 4 Nov 2026. v3 per-character pricing: official ElevenLabs API pricing (elevenlabs.io/pricing/api), checked 4 Nov 2026.