AI spend by model, harness, and user. Unbounded consumption is OWASP LLM10.
Spend guardrail
Suggested: $400 a day for the fleet, over it on 1 of 30 days (Oct 21, $1,523). A ceiling makes a day like Wed Oct 21 a finding while it happens, instead of a line on next month's bill.
Spend by harness
Spend by user
Spend over time
Team · spend
Team | Tokens | Responses | Cost |
|---|---|---|---|
| Engineering est | 137.4M | — | $4835.00 |
| GRC est | 13.4M | — | $492.00 |
| Finance | 5.80M | — | $268.00 |
| Team B est | 9.00M | — | $140.00 |
| Team A | 7.00M | — | $130.00 |
Model · spend
Model | Tokens | Responses | Cost |
|---|---|---|---|
| claude-opus-5 | 42.7M | — | $4099.00 |
| claude-sonnet-5 | 49.0M | — | $1176.00 |
| gpt-5.6-sol est | 64.5M | — | $516.00 |
| claude-haiku-4-5 | 11.9M | — | $60.00 |
| gpt-5.6-luna est | 4.50M | — | $14.00 |
Harness · spend
Harness | Tokens | Responses | Cost |
|---|---|---|---|
| Claude Code | 76.6M | — | $4687.00 |
| Cursor est | 75.0M | — | $1010.00 |
| Codex est | 18.0M | — | $144.00 |
| Copilot est | 3.00M | — | $24.00 |
User · spend
User | Tokens | Responses | Cost |
|---|---|---|---|
| eval-runner | 17.2M | — | $1651.00 |
| Hana | 8.40M | — | $535.00 |
| Jordan est | 16.2M | — | $526.00 |
| Noor | 6.50M | — | $442.00 |
| Ravi est | 14.6M | — | $422.00 |
| Mark | 6.40M | — | $388.00 |
| Kofi | 6.50M | — | $353.00 |
| Dev | 5.80M | — | $268.00 |
| Lena est | 19.0M | — | $252.00 |
| Priya est | 16.0M | — | $240.00 |
| Mateo est | 16.5M | — | $220.00 |
| Alex est | 16.5M | — | $194.00 |
| User B est | 9.00M | — | $140.00 |
| User A | 7.00M | — | $130.00 |
| Riley est | 7.00M | — | $104.00 |
Same work, cheaper model
$ per million tokensgpt-5.6-sol does the most work (64.5M tokens, #1) for the 3rd-largest bill ($516). claude-opus-5 is the reverse: #1 by spend, #3 by tokens. The difference is the price of a million tokens.
| Model | Tokens | Spend | $ per M | By spend | By tokens |
|---|---|---|---|---|---|
| claude-opus-5 Anthropic | 42.7M | $4,099 | $96 | #1 | #3 ▼ |
| claude-sonnet-5 Anthropic | 49.0M | $1,176 | $24 | #2 | #2 |
| gpt-5.6-sol OpenAI | 64.5M | $516 | $8 | #3 | #1 ▲ |
| claude-haiku-4-5 Anthropic | 11.9M | $60 | $5 | #4 | #4 |
| gpt-5.6-luna OpenAI | 4.5M | $14 | $3 | #5 | #5 |
What was paid per million input and output tokens, cache reads included in the $. Sorted by spend.
- Who
- Jordan, Noor, Hana, Kofi, Ravi
- Routine
- 12.0M of their 20.0M claude-opus-5 tokens: sessions under 20 tool calls: edits, tests, reviews
- Price
- $96 → $24 per million tokens
Longer sessions stay on claude-opus-5. The estimate assumes the same token count on claude-sonnet-5.
Seats you're paying twice for
9 seats · $679 a month4 people use a personal account beside the enterprise seat for the same provider ($440 a month), and 5 enterprise seats had no sessions in 30 days ($239 a month).
| Person | Seat | What we saw | A month |
|---|---|---|---|
| Personal account beside the enterprise seat · 4 · $440 | |||
| User A Team A | Claude Max (personal) Anthropic | Claude Code signed in as usera@gmail.com on mac-a1, beside an Anthropic Enterprise seat | $200 |
| Dev Finance | ChatGPT Pro (personal) OpenAI | chatgpt.com from a personal account on win-fin-02, beside a ChatGPT Enterprise seat | $200 |
| Mateo Engineering | Claude Pro (personal) Anthropic | claude.ai from a personal account on mac-eng-03, beside an Anthropic Enterprise seat | $20 |
| Jordan Engineering | ChatGPT Plus (personal) OpenAI | Codex signed in with a personal ChatGPT account on mac-eng-11, beside a ChatGPT Enterprise seat | $20 |
| Enterprise seats with no sessions in 30 days · 5 · $239 | |||
| Priya Engineering | Claude Enterprise Anthropic | Last session Sep 3. Her Claude use goes through Cursor | $60 |
| Sam GRC | ChatGPT Enterprise OpenAI | Last session Aug 28 | $60 |
| Riley GRC | ChatGPT Enterprise OpenAI | Last session Sep 14 | $60 |
| Hana Engineering | Cursor Business Cursor | Last session Sep 11. She works in Claude Code | $40 |
| Mark GRC | Copilot Business GitHub | No session since the seat was assigned (Jul 15) | $19 |
Cache rate and runaway sessions
A typical weekday is $188; weekends are quiet. One day stands out in Spend over time, and one session explains it.
$1,363 of it is one session: S-8812, eval-runner on claude-opus-5.
- Agent
- eval-runner · Claude Code, headless on ci-runner-02
- Team
- Engineering · scheduled by Kofi
- Ran
- 9 h 40 min · 1,912 model calls · 14.2M tokens
- Cache
- 4% read from cache (the fleet: 96%)
A retry loop re-sent the full context after every failed tool call, and a timestamp in its system prompt changed on every call, so almost nothing was read from cache.
Token counts on this page exclude cache reads. A cached token costs a tenth of a fresh one, so a session that stops hitting the cache costs more for the same work.