LafaekTOEFL AI CoachDocs

Cost per study plan

What one study plan really costs in AI calls: measured tokens, pictures and voice from a real plan, priced with Meta's and ElevenLabs' published rates.

4 min read 8 sectionsView on GitHub
On this page

How much does it cost to give one learner a full, personal TOEFL study plan? We measured it on the live app instead of guessing.

Answer: about $3.16 for a complete study plan, which includes:

  • 133 personal questions across all 12 TOEFL task types;
  • the AI study schedule;
  • 120 vocabulary words;
  • 25 pictures and 150 voice clips.

That is about $0.024 per question.

#How we measured it

On 2 November 2026 we created a brand-new account on lafaektoefl.tech and made one study plan, choosing every option:

  • all 7 study days, 90 minutes a day and a test on 25 February 2027;
  • all four skills as weak areas;
  • all five accents.

The plan was ready in 13 minutes. The account had done nothing else, so its usage totals are exactly the cost of one plan.

The app records every AI call in its usage_events table:

  • for Meta, the provider and model, plus the input and output tokens Meta reports;
  • for ElevenLabs, the characters sent.

The account's Usage page shows the totals:

Usage after one study plan

#What one study plan used

Item Amount
Meta text calls (questions, checks, vocabulary, schedule) 92 requests
Input tokens (Muse Spark) 89,399
Output tokens (Muse Spark, including reasoning) 356,228
Pictures (Muse Image) 25
Voice characters (ElevenLabs) 16,081 characters in 150 clips (about 16 minutes of audio)
Speech-to-text none (only used when a learner speaks)
Result 133 questions, 120 vocabulary words, 1 AI schedule, 122 finished jobs, 0 failed
Files stored in S3 175 files, 18.5 MB

#Prices used

Model Provider Price
muse-spark-1.3 (and the fallback muse-spark-1.2, same price) Meta $1.25 per 1M input tokens · $0.15 per 1M cached input tokens · $4.25 per 1M output tokens
muse-image-1.0 Meta $0.01 per image
muse-voice-transcribe-1.0 Meta $0.18 per hour of audio
eleven_multilingual_v2 ElevenLabs $0.08 per 1,000 characters

Sources:

  • Meta: prices from the Meta Model API model list (screenshots below), which match independent reports such as MindStudio.
  • ElevenLabs: the ElevenLabs API pricing page. Some third-party sites list $0.10 per 1,000 characters for the same model.
  • Cached input: the app does not track cached tokens, so all input is priced at the full rate. The real cost can only be the same or lower.

#The cost of one study plan

Part Calculation Cost
Muse Spark input 89,399 × $1.25 / 1M $0.112
Muse Spark output 356,228 × $4.25 / 1M $1.514
Muse Image pictures 25 × $0.01 $0.250
ElevenLabs voices 16,081 × $0.08 / 1,000 $1.286
Total $3.16
  • At $0.10 per 1,000 characters for voices, the total would be $3.48.
  • Storage is negligible: 18.5 MB in S3 costs well under one cent a month.

Where the money goes:

  • About half of the cost is Muse Spark's output tokens. Most of these are the model's reasoning while it writes and checks questions, about 2,700 output tokens per question.
  • About 40% is voices for the listening tasks.
  • Pictures are about 8%.

#Other AI costs

Action What it uses Cost
AI score and coaching for one email, discussion post or interview answer Measured: 935 input + 1,244 output tokens $0.0065
Speaking answer, plus speech-to-text A 45-second answer is 0.0125 hours × $0.18 about $0.002 more
Full mock test with AI rating (estimate) 11 transcriptions (about 3.5 minutes of speech) and 2 rating calls about $0.04
Word pronunciation Made once per word, the first time it is played about $0.001 per word

The mock-test figure is an estimate. We measured our test run, but its answers were left empty, so it used no AI calls.

#What this means for pricing

Plan Price to the learner AI cost to run it Difference
Starter: 1 study plan + 1 mock test Free about $3.20 −$3.20
Boost Pack: +1 study plan + 2 mock tests $5 about $3.24 about +$1.76
  • $150 of Meta Model API credit covers the Meta part of about 80 study plans ($1.88 each). With ElevenLabs included, $150 covers about 47 full plans.
  • Hosting is a fixed monthly cost and is not counted per plan. It covers the Vercel web app, the Render worker and database, and S3.
  • Compare this with prep materials: one official ETS practice test costs $49, and a writing or speaking evaluation from a prep site costs $27–$29 each.

#How to make it cheaper

  • Muse Spark 1.3 Contributor: $0.10 per 1M input and $0.20 per 1M output tokens.
    • It would cut the text part from $1.63 to about $0.08 per plan, so the whole plan would cost about $1.62.
    • The catch: inputs and outputs are used to train Meta's models, and it is available only in select countries. That needs the learner's consent.
  • ElevenLabs Flash or Turbo: $0.04 per 1,000 characters would halve the voice cost to about $0.64.
  • Fewer reasoning tokens: Build a Sentence already uses reasoning_effort: "low", which made it about 8× faster. Using low reasoning for more task types would cut output tokens further.

#Meta's published prices

Screenshots of the Meta Model API model list, with the prices used above:

Muse Spark 1.3, Muse Voice Transcribe and Muse Image prices
Muse Spark 1.2 and 1.1 prices
Muse Wonder and SAM 3.1 prices