AI Assistants Name the Right SaaS Price Only 20% of the Time

Four times out of five, a current AI assistant will not give you the current price of a well-known SaaS tool. We re-ran the test on September 1, 2026 with GPT-5.5, Claude Sonnet 5, and Gemini 3.7 Flash on the cheapest paid plan of 70 tools, using prices Toolradar had checked the same week. The models named the correct figure in 43 of 210 answers (20.5%). Most of the rest refused. When they did name a number, 42% of those numbers were wrong.

The August 21 run (GPT-4o, Claude Sonnet 4, Gemini 2.5 Flash) landed at 15% correct and 74% wrong among committed answers. The newer models are more careful. They are not more accurate.

Why this matters

A wrong price in a shortlist or a stack spreadsheet gets copied. A hedge is safer than a stale number, and that is the one thing that improved. It is not enough. An agent that says "I am not sure" still cannot finish the job. It needs a live source.

What we measured

We took 70 published tools with a Toolradar editorial score of 89 or higher and a clean monthly starting paid price between $3 and $100, each checked against the vendor page in the two weeks before the test. For each tool we stored the cheapest paid monthly plan as ground truth. We asked three current assistants, from memory, no web, no tools, for that plan and its price. We graded each answer as correct (a dollar figure within 12% of the verified price, or within $1), wrong (a specific price that misses), or hedged (no usable current figure).

Results by model

Gemini 2.5 Flash in August was the opposite of this Gemini: it named a price four times out of five and got most of them wrong. Gemini 3.7 Flash almost never names one. GPT-5.5 does the same. The only model that still behaves like a know-it-all pricing intern is Claude Sonnet 5.

The failure mode: last year's plan

These are real answers. The verified column is the cheapest paid plan as of September 1, 2026.

Claude has ChatGPT Plus in memory. OpenAI added Go. The rest is the same story: an older list price, a plan that was renamed, or a mid-tier quoted as the floor.

What fixes it

Toolradar checks pricing on a weekly cadence and exposes it through the Toolradar MCP server. An assistant that calls get_pricing or recommend_tools gets the plan name, the figure, and the date we last checked it. On these 70 tools that is the verified number, every time.

report_issue lets an agent send a mismatch back to the queue.

Methodology

Sample. 70 published SaaS products, editorial score 89 or higher, monthly starting paid price between $3 and $100. Usage-based, contact-sales, one-time, and enterprise-only products were excluded. We also dropped a handful of one-off licenses (Procreate, Alfred) and infrastructure SKUs (AWS Savings Plans) that passed the numeric filter but are not a typical SaaS seat. Pricing-check dates on the sample run from August 22 to September 1, 2026.

Ground truth. Cheapest paid monthly plan in the Toolradar database, taken from the vendor pricing page during that window.

Baseline. OpenAI GPT-5.5, Anthropic Claude Sonnet 5, and Google Gemini 3.7 Flash, via OpenRouter, temperature 0, no tools. Same protocol as the August 21 snapshot (GPT-4o, Claude Sonnet 4, Gemini 2.5 Flash; 79 tools; 15.2% correct).

Grading. Rule-based: any dollar figure within 12% of the verified price, or within $1, counts as correct. No dollar figure, or a refusal, is hedged. A specific miss is wrong. A numeric 12% sweep without the $1 slack matched 20.0% of answers, in line with the strict rate.

Limitations. Some misses are billing-basis (annual quote vs our monthly equivalent) or plan-scope (a mid-tier named as the floor). We still count those as wrong. This is a September 1, 2026 snapshot. Prices will move again.