Nobody can tell you what your team will spend, because it depends on how many agent sessions each person runs and how long they go. What you can do is build a range from published list prices and replace the guesses with your own logs. This post is an estimate, not a quote and not anyone's measured bill. Prices were read from Anthropic's pricing page on 18 October 2026.
The assumptions
- Ten developers, 21 working days a month.
- One session is the invented one used in our caching guide: 40 turns, a 20,000 token stable prefix, 3,000 new input tokens and 600 output tokens per turn.
- Sessions per developer per day: 3 (light), 6 (typical guess), 12 (heavy).
- API list prices, no discounts, no seat plans. Cached means the prefix is read from the cache after the first turn.
Per session, from the list prices: Claude Sonnet 5.5 costs $2.08 uncached and $0.61 cached. Opus 5.5 costs $4.16 and $1.22. Fable 5.1 costs $10.40 and $2.85. Haiku 5.5 costs $0.10 and $0.03.
Monthly estimate for ten developers, caching on
| Model | 3 sessions a day | 6 sessions a day | 12 sessions a day |
|---|---|---|---|
| Haiku 5.5 | $22 | $43 | $86 |
| Sonnet 5.5 | $383 | $766 | $1,532 |
| Opus 5.5 | $766 | $1,532 | $3,064 |
| Fable 5.1 | $1,792 | $3,585 | $7,169 |
With no caching at all, the Sonnet 5.5 row becomes $1,310, $2,621 and $5,242. That gap is the value of keeping the prompt prefix stable.
What this leaves out
- Sessions that run longer than 40 turns. Cost grows faster than the turn count because context grows too.
- Reasoning or thinking tokens, which are billed as output.
- Tool surcharges. Anthropic lists web search at $10 per 1,000 searches.
- A mix of models. Most teams use a smaller model for routine steps.
- Seat based plans. They are priced and limited differently, and the vendors do not publish their limits in tokens, so we do not convert them here.
- Regional uplifts: Anthropic lists a 1.1x multiplier for US only inference on Claude 4.6 and later, and OpenAI lists a 10 percent uplift for regional processing endpoints.
Run it with your numbers
devs, days = 10, 21
sessions_per_dev_day = 6
turns, prefix, new, out = 40, 20000, 3000, 600
inp, write, read, outp = 2.00, 2.50, 0.10, 10.00 # Sonnet 5.5 per MTok
cached = (prefix*write + (turns-1)*prefix*read
+ turns*new*inp + turns*out*outp) / 1e6
uncached = turns*((prefix+new)*inp + out*outp) / 1e6
n = devs * days * sessions_per_dev_day
print("sessions", n)
print("cached", round(n*cached), "uncached", round(n*uncached))
Take the usage data from a real week, fill in actual turns and token counts, and the range collapses to a number you can defend.
Turning the estimate into a decision
If your logged spend lands well below the monthly price of a team plan, the API is cheaper for you. If it lands above, a plan with fixed pricing may be. See our guide to choosing between subscription, API and enterprise. Safeguard can be connected to Claude, ChatGPT or your own agent over MCP, tool calls are unlimited, and you can bring your own model, so you pick the model bill separately.
Quick check
Which input moves this estimate the most?
Sessions per developer per day and the model tier. The model tier alone changes the monthly figure by more than 80 times between Haiku 5.5 and Fable 5.1 in the table.
Sources
- Claude API pricing, fetched 18 October 2026
- OpenAI API pricing, fetched 18 October 2026
