Finance
Token usage & cost
Spend, throughput and forecast — attributed by channel, client and conversation. Every figure recomputes against the controls below, so this doubles as a planning tool.
Spend over time
21 days · last 30 days
By channel
Token usage and processing time per channel. Email and chat skip transcription entirely, which is most of why they cost a fraction of a call.
| Channel | Audits | Input | Cached | Output | Total tokens | Avg AI time | p95 | Avg cost | Total |
|---|---|---|---|---|---|---|---|---|---|
| call | 23 | 33,107 | 55,651 | 131,204 | 220.0k | 1m 13s | 1m 40s | $0.1190 | $2.74 |
| 8 | 7,644 | 20,000 | 41,171 | 68.8k | 1m 7s | 1m 32s | $0.0673 | $0.54 | |
| chat | 7 | 3,344 | 17,500 | 33,350 | 54.2k | 56.2s | 1m 22s | $0.0614 | $0.43 |
Budget
Monthly ceiling versus the run rate implied by this period.
99% headroom — roughly 2,525 more conversations per month.
Volume planning
Cost if throughput scales, at current per-conversation economics.
| Scenario | Conversations/mo | Monthly cost | Annual |
|---|---|---|---|
| Today | 38 | $3.71 | $44.48 |
| 2× | 76 | $7.41 | $88.95 |
| 5× | 190 | $18.53 | $222.38 |
| 10× | 380 | $37.06 | $444.76 |
| 25× | 950 | $92.66 | $1.1k |
Linear in volume — there are no step costs until transcription throughput needs more workers, which is the point to move to self-hosted ASR.
By client
What each programme costs to run — the number to price a contract against.
Where the money goes
- transcription100
- model271
- Input (uncached)
- 44,095
- Cached reads
- 93,151
- Output
- 205,725
- Cache hit rate
- 68%
Transcription: hosted vs self-hosted
Only applies to calls — email and chat skip transcription entirely.
| Hosted | Self-hosted | |
|---|---|---|
| Rate per hour | $0.54 | $0.06 |
| Last 30 days (1.9h audio) | $1.00 | $0.11 |
| Projected monthly | $1.00 | $0.111 |
| Annual difference | $10.68 saved |
Self-hosting is not only cheaper — it keeps audio inside infrastructure you control. With hosted transcription the raw recording leaves the country before redaction can run on it, which is the awkward part of the PDPA conversation.
What each tier would cost
Same conversations, same settings. Cheaper is only cheaper if agreement holds — check calibration first, since a model that disagrees with your QA lead costs more in disputes than it saves in tokens.
| Model | Rate in/out | Last 30 days | Projected monthly | vs current |
|---|---|---|---|---|
| Opus 5 | $5 / $25 | $3.71 | $3.71 | selected |
| Sonnet 5 | $3 / $15 | $2.62 | $2.62 | −29% |
| Haiku 4.5 | $1 / $5 | $1.54 | $1.54 | −58% |
Most expensive conversations
38 in period| Conversation | Client | Ch. | In / Cached / Out | AI time | ASR | Model | Total |
|---|---|---|---|---|---|---|---|
| Priya Ramanconv-0013 · 22 Jul | Meridian Telecom | 2,665 / 2,500 / 6,879 | 1m 26s | $0.0876 | $0.0933 | $0.1809 | |
| Daniel Ongconv-0033 · 4 Jul | Lumina Retail | 1,989 / 2,500 / 5,060 | 1m 11s | $0.0929 | $0.0688 | $0.1617 | |
| Nor Azlinaconv-0035 · 20 Jul | Lumina Retail | 2,084 / 2,500 / 4,260 | 49.9s | $0.0914 | $0.0591 | $0.1504 | |
| Ravi Kumarconv-0012 · 3 Jul | Straits Retail Bank | 2,570 / 2,500 / 6,889 | 1m 48s | $0.0570 | $0.0932 | $0.1502 | |
| Nor Azlina2026072300751 · 31 Jul | Lumina Retail | 2,821 / 3,151 / 8,207 | 1m 30s | $0.0370 | $0.1104 | $0.1475 | |
| Michelle Kohconv-0017 · 6 Jul | Meridian Telecom | 1,772 / 2,500 / 6,949 | 1m 40s | $0.0420 | $0.0919 | $0.1339 | |
| Farah Ismailconv-0008 · 31 Jul | Straits Retail Bank | 2,279 / 2,500 / 4,345 | 46.2s | $0.0707 | $0.0606 | $0.1313 | |
| Marcus Limcall-002 · 30 Jul | Straits Retail Bank | 1,430 / 2,500 / 6,167 | 1m 22s | $0.0495 | $0.0813 | $0.1308 | |
| Nurul Aisyahcall-006 · 23 Jul | Meridian Telecom | 1,040 / 2,500 / 7,147 | 1m 35s | $0.0360 | $0.0926 | $0.1286 | |
| Marcus Limcall-009 · 22 Jul | Straits Retail Bank | 1,153 / 2,500 / 6,559 | 1m 21s | $0.0399 | $0.0855 | $0.1254 | |
| Kausalya Rameshconv-0005 · 3 Jul | Lumina Retail | 1,790 / 2,500 / 4,563 | 1m 8s | $0.0630 | $0.0621 | $0.1251 | |
| Priya Ramancall-005 · 27 Jul | Meridian Telecom | 823 / 2,500 / 7,268 | 1m 38s | $0.0285 | $0.0935 | $0.1220 | |
| Michelle Kohconv-0032 · 21 Jul | Meridian Telecom | 2,682 / 2,500 / 7,604 | 1m 26s | $0.0173 | $0.1024 | $0.1196 | |
| Siti Nurhalizaconv-0004 · 28 Jul | Anchor Insurance | 754 / 2,500 / 4,958 | 1m 13s | $0.0462 | $0.0645 | $0.1107 | |
| Kausalya Ramesh2026072300750 · 31 Jul | Lumina Retail | 1,344 / 0 / 5,263 | 57.9s | $0.0399 | $0.0691 | $0.1090 |
Rates are first-party Claude API list prices. Cached reads bill at 10% of input, cache writes at 125%, Batch API halves everything. Transcription is $0.54/hr hosted or $0.06/hr self-hosted. AI time is model wall-clock per audit — with Batch enabled these run concurrently, so it is a cost and capacity signal rather than a latency the user waits on.