Which AI is best for customer support replies?

In our October 2026 test, Claude Opus scored 98.6 on hard support emails with 0 severe errors over 2 passes. GPT-6.1 Sol scored 97.3, also with 0 severe errors. Gemini 4 Argon scored 97.8 but ran only 1 pass.

What is the biggest risk when AI drafts support emails?

Promising the wrong refund. 7 models promised a wrong refund amount on the same case. Leaking another customer's data was the other serious one: GPT-5.4 mini did it on 2 emails and Claude Haiku once.

Are cheap models good enough for support email?

For easy FAQ-style email, nearly all did fine: all 10 models scored 98.3-99.6 on the easy round. On hard email the low-cost GPT-5.4 mini scored 68.0 with 10 severe errors, and GPT-6 Luna scored 82.2 with 5.

Round 1 (easy): 80 emails. Round 2 (hard): 40 main + 20 fresh held-back emails written blind. Passes: 2 for 7 models; Gemini 4 Argon, Grok 4.7 and Grok 4.6 ran 1 pass (less certain). Results updated 15 Oct 2026.

RealRun / Guides

Best AI for customer support replies (2026)

Short answer: In our October 2026 test of 10 AI models on 140 customer support emails, Claude Opus scored highest on the hard round (98.6) with 0 severe errors over 2 passes. GPT-6.1 Sol (97.3) also made 0 severe errors. Gemini 4 Argon (97.8) did too, but ran only 1 pass, so it is less certain. We don't recommend GPT-5.4 mini (68.0): it leaked another customer's data on 2 emails.

Scores (hard round)

Model	Hard round score	Severe errors	Main 40	Fresh held-back 20	Passes	Easy round	Cost
Claude Opus	98.6	0	99.1	97.7	2	99.2	subscription
Gemini 4 Argon	97.8	0	98.3	96.9	1	99.5	medium
GPT-6.1 Sol	97.3	0	97.5	96.8	2	98.6	medium
Grok 4.7	93.2	1	93.9	96.8	1	99.2	high
Gemini 3.8 Flash	90.6	2	94.5	90.2	2	98.8	low
Claude Sonnet	89.4	3	89.3	97.2	2	99.6	subscription
Grok 4.6	88.0	2	88.0	n/a	1	98.9	high
GPT-6 Luna	82.2	5	83.1	92.9	2	98.9	low
Claude Haiku	80.5	6	89.0	85.9	2	98.7	subscription
GPT-5.4 mini	68.0	10	77.8	83.6	2	98.3	low

n/a = not run (Grok 4.6 ran the main set only). Severe errors are totals over all passes.

How the testing works: every model got the same set: Round 1 (easy): 80 emails. Round 2 (hard): 40 main + 20 fresh held-back emails written blind (140 items, 2,600 graded answers, 10 models). Passes: 2 for 7 models; Gemini 4 Argon, Grok 4.7 and Grok 4.6 ran 1 pass (less certain). Score: Round 2 overall: mean item score minus 5 per severe error, over all items answered, mean of passes. Answer keys for 4 items were widened after seeing answers to accept a second correct answer (logged).

Results updated 15 Oct 2026. Tests run October 2026. Full results: Customer support replies.

What we found

Round 1 (easy) was a tie: all 10 models scored 98.3-99.6. Easy email can't tell you which model to trust.
7 models promised a wrong refund amount on the same case.
GPT-5.4 mini: Leaked another customer's data on 2 emails, promised wrong refund amounts and missed an escalation; 10 severe errors over 2 passes.
Claude Haiku: Leaked another customer's data once and promised wrong refund amounts; 6 severe errors over 2 passes.
GPT-6 Luna: Promised wrong refund amounts; 5 severe errors over 2 passes.

For you

The way support email goes wrong isn't tone. It's money and privacy. In our hard round, 7 models promised the same customer the wrong refund amount, and GPT-5.4 mini put another customer's details into a reply on 2 emails. If an AI drafts your replies, those are the two things to check before anything goes out.

For your clients

If you set up support bots for clients, the easy round won't protect you: all 10 models scored 98.3 to 99.6 there. Show the client the hard-round results instead, where only 3 models finished with 0 severe errors, and make wrong refund amounts and data leaks the first two things your setup checks for.

Want the exact prompt and setup?

Get the playbook for this job, or all five for $99 (save $46).

Support Replies AI Playbook, $29 or get the free weekly AI test report

FAQ

Which AI is best for customer support replies?: In our October 2026 test, Claude Opus scored 98.6 on hard support emails with 0 severe errors over 2 passes. GPT-6.1 Sol scored 97.3, also with 0 severe errors. Gemini 4 Argon scored 97.8 but ran only 1 pass.
What is the biggest risk when AI drafts support emails?: Promising the wrong refund. 7 models promised a wrong refund amount on the same case. Leaking another customer's data was the other serious one: GPT-5.4 mini did it on 2 emails and Claude Haiku once.
Are cheap models good enough for support email?: For easy FAQ-style email, nearly all did fine: all 10 models scored 98.3-99.6 on the easy round. On hard email the low-cost GPT-5.4 mini scored 68.0 with 10 severe errors, and GPT-6 Luna scored 82.2 with 5.
How was this tested?: Round 1 (easy): 80 emails. Round 2 (hard): 40 main + 20 fresh held-back emails written blind. Passes: 2 for 7 models; Gemini 4 Argon, Grok 4.7 and Grok 4.6 ran 1 pass (less certain). Results updated 15 Oct 2026.

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).