Using AI to Estimate Software Costs

Testing how Claude, Gemini, and ChatGPT handle ETL pricing over 20 runs each.

February 18, 2026 · 7 min read

ETL pricing is complicated. Row counts, connector fees, usage-based tiers, volume discounts. It's exactly the kind of thing we want to use AI to help figure out.

To test frontier model capabilities, I gave Claude, Gemini, and ChatGPT the same scenario: estimate monthly ETL cost for a tech company spending $10M/year on digital ads. I ran the same prompt 20 times per model for consistency. (See appendix for full prompt and methodology.)

Here's what I learned.

Key Takeaways

Cost estimates vary widely, even when using the exact same prompt.
Most of the spread comes from assumptions about price, not volume (how much data needs to move).
Each LLM has its own view on software pricing. Gemini comes in lowest for all three vendors tested.
AI models seem to agree more when clear pricing information is available.

Cost Estimates Vary Widely

Here's how each model estimated costs for the three vendors across 20 runs each of the same prompt.

Box plot showing distribution of monthly ETL cost estimates by LLM and vendor. Gemini clusters tightly at low values for all three vendors. Claude has moderate spread. GPT scatters widely, especially for Fivetran ($2,500 to $80,000). Median across all 60 runs: Fivetran $5.8K, Airbyte $2.2K, Estuary $1.2K.

In the plots below, each dot is one run (midpoint of the model's stated range). Lines mark the median. See appendix for full run data.

Fivetran

Airbyte

Estuary

Most of the Spread Comes from Assumptions About Price, Not Data Volume

I extracted the assumptions and reasoning baked into each LLM response. Across all three models, the inputs were reasonable: 5–7 ad platforms, 15-minute to daily syncs, etc. No run was nonsensical. Each one built a defensible estimate from its own assumptions.

Net net, the models are relatively close on how much data needs to move. Medians range from 18M to 28M rows per month:

Strip plot showing assumed monthly active rows by model. Medians cluster between 18M (GPT) and 28M (Claude), with most estimates in the 10M–40M range. A few outliers reach 100M–150M.

But they are much farther apart on what they expect it costs to move each row of that data. For example, Fivetran median price per million rows per month ranges from $150 (Gemini) to $1,138 (GPT):

Strip plot showing implied price per million rows by vendor and model. For Fivetran, GPT assumes ~$1,138/million rows while Gemini assumes ~$150/million. For Airbyte, Claude assumes ~$278/million while Gemini assumes ~$40/million. For Estuary, GPT assumes ~$158/million while Gemini assumes ~$8/million.

That's almost an 8x spread at the median for Fivetran, 7x for Airbyte, and 20x for Estuary — compared to just 1.6x on volume.

Price per row, not volume, drives nearly all the variance.1

Putting It All Together

Each model appears to have a relatively consistent point of view on pricing. Gemini estimated the lowest and tightest pricing for all three vendors:

Grouped bar chart showing median monthly ETL cost estimates by LLM. For Fivetran: Gemini $4K, Claude $6K, GPT $16K. For Airbyte: Gemini $900, Claude $5K, GPT $2K. For Estuary: Gemini $200, Claude $1K, GPT $2K.

Combined Estimates

Here's where things land all together:

Fivetran $4K–$16K/mo

Tightest consensus across models (300% spread). Consistent with published rates.2

Airbyte Cloud $900–$5.1K/mo

Wider spread across models (467% spread).3

Estuary $200–$2,400/mo

Widest spread across models (1,100% spread).

It's probably not a coincidence that consensus tightens with how much the models have to go on.

Fivetran's MAR-based pricing is well-documented and widely benchmarked, so the models converge.
Estuary's GB-based pricing is simple arithmetic but sparsely documented, giving it the widest spread.
Airbyte's credit system is the most opaque: credits vary by connector type, and the system has evolved significantly. It's the only vendor where the model rank order flips.

Practical Takeaways

Based on these findings, a few tips if you're using AI for vendor pricing research with complex structures:

Consider running the same prompt across multiple models, multiple times. There's enough variability in any single LLM response to caution making decisions on it.
Expect less accuracy on newer or less-documented vendors where the models have less to go on.
Check the assumptions behind each answer — the reasoning tends to be sound, but the inputs can vary by an order of magnitude.

Whether these patterns generalize — tighter estimates for market leaders, Gemini consistently anchoring low — is worth testing in other categories. Any vendor with a pricing calculator is a candidate.

If you're a buyer or SaaS company interested in applying this approach, get in touch.

James Riso

Founder, Riso Group

James is a data and AI strategy consultant who helps companies build scalable analytics infrastructure and data-driven growth strategies. Connect on LinkedIn.

Notes

1. A counterfactual confirms this: equalize the rate across models and the cost estimates converge. Equalize the volume instead and the spread remains. ↩
2. Fivetran Standard plan rates (post-March 2025): $2.50/M for first 5M MAR, $2.00/M for 5–20M, $1.50/M for 20–100M, plus $5/mo minimum per connection and $12K/yr annual minimum. March 2025 change moved MAR billing from account-level to per-connector, increasing costs 40–70% for multi-connector setups. Fivetran Pricing ↩
3. Sticker price understates total cost since Airbyte requires more hands-on engineering. Most Airbyte Cloud source connectors are community-maintained with no uptime SLA. Schema changes default to manual approval, and retry/rate-limit tuning falls on the user. Fivetran maintains all connectors in-house. Airbyte Connector Support Levels ↩

Appendix: Methodology

Setup

The same prompt was sent to all three models, unmodified:

Act as a Senior Data Engineer. Using your knowledge and any necessary tools, you will estimate the monthly cost to ETL marketing data (Meta, Google, LinkedIn, etc.) for a tech business spending $10M/year on advertising. The data ingested needs to be granular, timely, and land in an existing data warehouse. Ignore the warehousing costs, we are focused only in the cost of data movement. Compare these specific vendors: Fivetran, Airbyte (Cloud), Estuary Output Requirements: Summary Table: Columns for Vendor, Est. Cost, Pricing Model, and Notes Express all cost estimates as monthly USD amounts. Return the Summary Table in markdown format. Final Recommendation: Give a recommendation assuming the team has minimal engineering resources to devote to this task

Three models — Claude Opus 4.6 (Anthropic), GPT-5.2 (OpenAI), and Gemini 3 Pro (Google) — accessed via the OpenRouter API. Each was run 20 times at temperature 1.0.

Parsing

Responses were parsed with a 3-tier fallback:

JSON — structured responses with a summary_table array
Markdown table — pipe-delimited tables (most common)
Text fallback — keyword scan for vendor names and dollar amounts

Most models returned cost estimates as ranges (e.g., "$10,000–$16,000/month"). The charts plot the midpoint of each range; single-value estimates are used as-is. Monthly Active Row (MAR) assumptions were extracted via regex (e.g., "10–15M MAR", "~25M rows/month"). For runs with explicit MAR figures, I derived an effective price per million rows to produce the price-per-row chart.

Full Run Data

All 60 runs (20 per model) with extracted assumptions and cost estimates. Calculated-midpoint columns show the values plotted in the charts above.

Model	Run	MAR (M rows)	Sync Freq	# Platforms	Fivetran (stated range)	Fivetran (calc. midpoint)*	Airbyte (stated range)	Airbyte (calc. midpoint)*	Estuary (stated range)	Estuary (calc. midpoint)*
Claude Opus 4.6	1	12M	hourly	7	$10,000 – $16,000	$13,000	$2,000 – $4,500	$3,250	$625 – $2,000	$1,312
Claude Opus 4.6	2	38M	real-time	7	$20,000 – $35,000	$27,500	$8,000 – $15,000	$11,500	$2,000 – $5,000	$3,500
Claude Opus 4.6	3	22M	6-hour	6	$2,500 – $5,000	$3,750	$750 – $2,250	$1,500	$500 – $1,500	$1,000
Claude Opus 4.6	4	32M	daily	7	$3,000 – $7,000	$5,000	$2,000 – $6,000	$4,000	$500 – $2,000	$1,250
Claude Opus 4.6	5	15M	real-time	7	$12,000 – $20,000	$16,000	$4,000 – $8,000	$6,000	$400 – $800	$600
Claude Opus 4.6	6	14M	real-time	8	$10,000 – $16,000	$13,000	$5,000 – $7,000	$6,000	$2,000 – $4,000	$3,000
Claude Opus 4.6	7	48M	real-time	7	$3,000 – $6,000	$4,500	$8,000 – $18,000	$13,000	$400 – $700	$550
Claude Opus 4.6	8	10M	6-hour	7	$5,000 – $8,000	$6,500	$2,500 – $4,500	$3,500	$500 – $900	$700
Claude Opus 4.6	9	32M	hourly	9	$20,000 – $50,000	$35,000	$8,000 – $20,000	$14,000	$2,000 – $5,000	$3,500
Claude Opus 4.6	10	32M	hourly	4	$4,000 – $7,000	$5,500	$12,000 – $20,000	$16,000	$500 – $2,000	$1,250
Claude Opus 4.6	11	22M	real-time	6	$6,000 – $10,000	$8,000	$500 – $2,000	$1,250	$500 – $2,000	$1,250
Claude Opus 4.6	12	45M	hourly	6	$4,000 – $8,000	$6,000	$500 – $2,000	$1,250	$250 – $750	$500
Claude Opus 4.6	13	45M	real-time	6	$3,000 – $5,000	$4,000	$4,000 – $8,000	$6,000	$750 – $2,000	$1,375
Claude Opus 4.6	14	20M	real-time	7	$20,000 – $40,000	$30,000	$20,000 – $50,000	$35,000	$1,000 – $3,000	$2,000
Claude Opus 4.6	15	115M	6-hour	8	$15,000 – $30,000	$22,500	$10,000 – $25,000	$17,500	$500 – $2,000	$1,250
Claude Opus 4.6	16	150M	hourly	7	$4,000 – $7,000	$5,500	$2,500 – $6,000	$4,250	$500 – $2,000	$1,250
Claude Opus 4.6	17	16M	6-hour	7	$4,000 – $7,000	$5,500	$2,500 – $5,000	$3,750	$1,000 – $2,500	$1,750
Claude Opus 4.6	18	10M	real-time	7	$3,000 – $6,000	$4,500	$2,000 – $4,500	$3,250	$800 – $1,500	$1,150
Claude Opus 4.6	19	100M	real-time	7	$25,000 – $50,000	$37,500	$30,000 – $60,000	$45,000	$2,000 – $5,000	$3,500
Claude Opus 4.6	20	22M	real-time	7	$4,000 – $8,000	$6,000	$1,500 – $4,000	$2,750	$500 – $1,500	$1,000
GPT-5.2	1	10M	daily	2	~$4,000/mo (typ. ~$3k–$6k depending on MAR)	$4,500	~$1,000–$2,000/mo	$1,500	~$1,000–$2,500/mo	$1,750
GPT-5.2	2	60M	hourly	3	~$80,000/mo	$80,000	~$4,000/mo	$4,000	~$1,500/mo	$1,500
GPT-5.2	3	30M	hourly	5	~$40,000–$55,000/mo	$47,500	~$1,000–$3,000/mo	$2,000	~$1,500–$4,000/mo	$2,750
GPT-5.2	4	155M	hourly	2	~$60,000/mo	$60,000	~$800/mo	$800	~$500/mo	$500
GPT-5.2	5	60M	hourly	3	~$28,000/mo	$28,000	~$1,800/mo	$1,800	~$2,500/mo	$2,500
GPT-5.2	6	15M	hourly	2	~$20k–$35k/mo	$27,500	~$1k–$3k/mo	$2,000	~$1k–$4k/mo	$2,500
GPT-5.2	7	10M	real-time	3	~$9,000/mo	$9,000	~$1,200/mo	$1,200	~$1,500/mo	$1,500
GPT-5.2	8	50M	real-time	2	$25,000/mo	$25,000	$8,000/mo	$8,000	$4,000/mo	$4,000
GPT-5.2	9	—	hourly	3	~$12,000/mo (typical range ~$8k–$20k)	$14,000	~$900/mo (typical range ~$500–$1.5k)	$1,000	~$1,200/mo (typical range ~$1k–$2k)	$1,500
GPT-5.2	10	25M	hourly	3	~$30,000/mo (typical range $20k–$40k)	$30,000	~$2,000/mo (typical range $1k–$5k)	$3,000	~$2,500/mo (typical range $1.5k–$6k)	$3,750
GPT-5.2	11	5M	hourly	2	~$6,000/mo	$6,000	~$3,500/mo	$3,500	~$1,500/mo	$1,500
GPT-5.2	12	45M	hourly	3	$12,000–$22,000/mo	$17,000	$800–$2,500/mo	$1,650	$1,000–$3,500/mo	$2,250
GPT-5.2	13	60M	hourly	4	$15,000/mo	$15,000	$4,000/mo	$4,000	$2,500/mo	$2,500
GPT-5.2	14	6M	hourly	3	~$6,000/mo	$6,000	~$2,000/mo	$2,000	~$1,200/mo	$1,200
GPT-5.2	15	—	daily	2	$4,000–$9,000/mo	$6,500	$1,200–$3,500/mo	$2,350	$2,000–$5,000/mo	$3,500
GPT-5.2	16	15M	hourly	2	~$20,000–$30,000/mo	$25,000	~$5,000–$10,000/mo	$7,500	~$2,000–$6,000/mo	$4,000
GPT-5.2	17	4M	hourly	4	~$3,500/mo	$3,500	~$1,500/mo	$1,500	~$900/mo	$900
GPT-5.2	18	2M	hourly	2	~$2,500/mo	$2,500	~$1,300/mo	$1,300	~$900/mo	$900
GPT-5.2	19	10M	hourly	3	$15,000/mo	$15,000	$4,000/mo	$4,000	$2,500/mo	$2,500
GPT-5.2	20	20M	hourly	3	~$25,000 / mo (typ. band: $20k–$35k)	$27,500	~$10,000 / mo (typ. band: $7k–$15k)	$11,000	~$6,000 / mo (typ. band: $4k–$12k)	$8,000
Gemini 3 Pro	1	20M	15-minute	3	$2,200–$2,800	$2,500	$400–$800	$600	$100–$200	$150
Gemini 3 Pro	2	28M	hourly	2	$2,500–$3,200	$2,850	$600–$900	$750	<$100	$100
Gemini 3 Pro	3	25M	hourly	4	$4,500–$6,000	$5,250	$1,500–$2,500	$2,000	$200–$400	$300
Gemini 3 Pro	4	10M	daily	1	$3,200–$4,000	$3,600	$600–$900	$750	$100–$200	$150
Gemini 3 Pro	5	35M	daily	1	$4,500–$6,000	$5,250	$600–$1,000	$800	$200–$500	$350
Gemini 3 Pro	6	3M	daily	2	$1,200–$1,800	$1,500	$750–$1,100	$925	$50–$200	$125
Gemini 3 Pro	7	38M	hourly	2	$4,500–$6,000	$5,250	$2,000–$3,000	$2,500	$50–$250	$150
Gemini 3 Pro	8	50M	daily	1	$3,500–$4,500	$4,000	$750–$1,000	$875	$100–$300	$200
Gemini 3 Pro	9	20M	daily	1	$3,500–$4,500	$4,000	$500–$800	$650	$0–$50	$25
Gemini 3 Pro	10	20M	hourly	2	$3,000–$4,500	$3,750	$800–$1,200	$1,000	$100–$300	$200
Gemini 3 Pro	11	25M	15-minute	4	$5,500–$7,000	$6,250	$3,000–$4,500	$3,750	$200–$600	$400
Gemini 3 Pro	12	32M	15-minute	3	$3,500–$4,500	$4,000	$500–$800	$650	$100–$250	$175
Gemini 3 Pro	13	62M	hourly	4	$4,500–$6,500	$5,500	$2,200–$3,000	$2,600	$400–$800	$600
Gemini 3 Pro	14	40M	daily	2	$3,500–$4,500	$4,000	$1,800–$2,200	$2,000	$200–$400	$300
Gemini 3 Pro	15	10M	—	2	~$2,100–$2,500	$2,300	~$400–$600	$500	<$50–$250	$150
Gemini 3 Pro	16	20M	—	3	$3,500–$4,500	$4,000	$2,000–$2,800	$2,400	$200–$500	$350
Gemini 3 Pro	17	25M	—	3	$3,000–$4,500	$3,750	$600–$1,000	$800	$100–$300	$200
Gemini 3 Pro	18	40M	15-minute	2	$2,500–$3,500	$3,000	$1,200–$1,800	$1,500	$100–$300	$200
Gemini 3 Pro	19	25M	15-minute	2	$3,500–$5,000	$4,250	$600–$900	$750	$50–$150	$100
Gemini 3 Pro	20	50M	hourly	1	$2,500–$3,800	$3,150	$800–$1,200	$1,000	$100–$300	$200

* Calculated midpoint is the simple average of the low and high ends of the stated range. When a model returned a single figure, that figure is used as-is. These are the values plotted in the charts above.