Which AI model should I recommend to clients?

It depends on the job. GPT-6.1 Sol was in the top three on all 5 jobs we tested. Claude Opus led customer support. GPT-6 Luna was the cheapest top model on finance. Every job had at least one model we don't recommend.

Is RealRun independent?

RealRun is run by Base Ledger IP LLC. We publish the scores and the mistakes for free and make money by selling tested prompts as playbooks.

Is the data free to use?

Yes. The results pages, results.json and the MCP server are free.

RealRun / Guides

Independent AI benchmarks on real business tasks

Short answer: RealRun tests AI models on the jobs a small business actually hands them: inbox triage, meeting scheduling, receipts, small-business finance and customer support replies. Across 473 test items and 8,629 graded answers, 13 models were tested, 8 to 10 per job. No single model won every job, and every job had at least one model we don't recommend.

Results by job

Job	Items	Models	Top score	Not recommended
Inbox triage	80	8	GPT-6.1 Sol, GPT-5.4 mini (95.0)	Claude Haiku, Grok 4.20
Meeting scheduling	62	8	Gemini 3.8 Flash, GPT-6.1 Sol, Grok 4.7, Claude Opus, Claude Sonnet (100.0)	Grok 4.20
Receipts	79	8	GPT-6.1 Sol (95.8)	Grok 4.3
Small-business finance	112	8	GPT-6 Luna (96.3)	Grok 4.3, Gemini 3.5 Flash-Lite
Customer support replies	140	10	Claude Opus (98.6)	GPT-5.4 mini

Headline metrics differ by job (accuracy %, valid-answer %, or a 0-100 score minus 5 per severe error). See each results page.

How the testing works: each job has its own test set with a fresh held-back part, every model in that job gets the same items, and most models ran 2 or more passes. In the support test, 3 of 10 models ran 1 pass and are marked less certain. Every answer is graded, and the worst mistakes (a scam marked safe, a wrong refund, a made-up total) are called out on their own. Results updated 15 Oct 2026.

Which model to recommend to clients, by job

Inbox triage: GPT-6.1 Sol or GPT-5.4 mini (95.0 each on fresh emails). Avoid Claude Haiku and Grok 4.20. Details
Meeting scheduling: any of the 5 models at 100.0 on fresh requests. Avoid Grok 4.20 (41.7). Details
Receipts: GPT-6.1 Sol (95.8), or GPT-6 Luna (91.4, low cost tier). Avoid Grok 4.3. Details
Finance: Top four (GPT-6 Luna, Gemini 4 Argon, GPT-6.1 Sol, Grok 4.7) are a near tie. GPT-6 Luna was the cheapest; Gemini 4 Argon and GPT-6.1 Sol had no severe errors. Details
Customer support: Claude Opus or GPT-6.1 Sol, both 0 severe errors over 2 passes. Avoid GPT-5.4 mini. Details

For you

A model that's great at one job can be the wrong pick for the next. GPT-5.4 mini tied for the top on inbox triage and was the one we don't recommend for support email, where it leaked another customer's data. Pick by job, not by brand.

For your clients

When a client asks which model to use, give them a job-by-job answer with the severe-error count, and cite the results page. The data is also in results.json and our free MCP server, so your agent can check it too.

Want the exact prompt and setup?

Get all five playbooks for $99 (save $46).

Get the free weekly AI test report

FAQ

Are there independent benchmarks for AI on real business tasks?: Yes. RealRun tests AI models on 5 small-business jobs: inbox triage, meeting scheduling, receipts, small-business finance and customer support replies. 473 items and 8,629 graded answers, updated 15 Oct 2026.
Which AI model should I recommend to clients?: It depends on the job. GPT-6.1 Sol was in the top three on all 5 jobs we tested. Claude Opus led customer support. GPT-6 Luna was the cheapest top model on finance. Every job had at least one model we don't recommend.
Is RealRun independent?: RealRun is run by Base Ledger IP LLC. We publish the scores and the mistakes for free and make money by selling tested prompts as playbooks.
Is the data free to use?: Yes. The results pages, results.json and the MCP server are free.

More from RealRun

Scores reflect our tests on the dates shown, not a guarantee. Claude models ran on a Claude subscription, so they have no metered API cost. cost_tier is relative within a task (low/medium/high) from our run spend; 'subscription' = not metered. Tested October 2026 by RealRun (Base Ledger IP LLC).