Self-hosted open model, chat plan or API: costs for five analysts

For five analysts, published prices on 18 October 2026 give three very different shapes: seats at a fixed monthly fee, API usage that scales with tokens, and a rented GPU that costs the same whether busy or idle. The right choice depends on usage, data rules and staff time, so run your own numbers.

All prices below were fetched on 18 October 2026 from the provider pages listed under Sources. They change, so recheck before you budget. This post makes no claim about any provider's internal costs.

What do the published prices say?

Chat plan seats (Claude Team). Standard seat: $25 per seat per month billed monthly, or $20 billed annually. Premium seat: $125 monthly, or $100 annually. The page says all plans have usage limits on a rolling five hour window, paid plans also have weekly limits, and usage from web, desktop, mobile and Claude Code draws from one shared pool. Team Premium seats get 5x the usage of Standard seats.

API tokens. The pricing page lists Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens.

Rented GPU for a self-hosted model. Lambda's page lists an NVIDIA H100 PCIe 80 GB at $3.29 per GPU hour in its single GPU listing. Hugging Face Inference Endpoints lists one H100 at $4.50 per hour and one L40S at $1.80 per hour.

What does each option cost for five people?

Seats, five analysts, per month:

• Standard seats: $125 billed monthly, or $100 on annual billing.
• Premium seats: $625 billed monthly, or $500 on annual billing.

API, with an assumption you should replace. Say each analyst sends 1 million input tokens and 0.2 million output tokens per working day, over 22 days. That is 22 million input tokens ($44) and 4.4 million output tokens ($44) per person, $88 each, or $440 for the team per month. This is an illustration, not a measurement. Agent style work with long contexts can use far more.

Self-hosted GPU, one H100 PCIe at $3.29 per hour:

• Running around the clock for 730 hours: $2,401.70 per month.
• Running only 8 hours on 22 working days, 176 hours: $579.04 per month.

On Hugging Face at $4.50 per hour, those become $3,285 and $792.

What do the numbers hide?

• Capacity and quality. A GPU price says nothing about which open model fits in 80 GB, how fast it responds, or whether it does the task well. Test on your work.
• Staff time. Someone deploys, patches, monitors and updates the model server. That time is real money and is not in the table.
• Idle cost. A GPU left on overnight costs the same as one in use. Scheduling matters more than the hourly rate.
• Plan limits. Seats are predictable but capped. Heavy use may hit the five hour window or weekly limits.
• Token growth. API cost grows with context length and agent loops, not with headcount.

When does data control decide it?

If security data, code or customer records cannot leave your environment, the comparison changes: a hosted plan or API may be off the table, and self-hosting becomes a requirement, not a cost choice. Safeguard supports Full AI, Zero AI and bring your own model, including local models, so the platform can follow whichever choice your data rules force. The platform itself does not need a language model to run in Zero AI mode.

What to do

1. Measure one week of real use: tokens per analyst per day.
2. Price the three options with your measured numbers, not the illustration above.
3. Add staff time for the self-hosted option, at your loaded rate.
4. Check which data classes may go to a hosted service.
5. Pilot the cheapest option that meets the data rules, for a month.
6. Recheck published prices every quarter.

Quick check

Sources

Fetched on 18 October 2026.

• Claude pricing: plans
• Claude Developer Platform: model pricing
• Lambda: GPU instance pricing
• Hugging Face: pricing

Last reviewed: 18 October 2026.