Business · 26 tasks · 43 models
Cheapest AI models for Customer Support
Which models resolve customer issues with empathy without inventing policy, over-promising, or fabricating account facts?
The cheapest capable model for Customer Support is gpt-5.4-mini, at $0.015 per run — and it still clears our quality bar.
Top score — strong
Clears the quality bar at $0.015/run
~12s per run, still strong
Quality vs. cost
Every model placed by what it delivers and what it costs. The best value sits high and to the left.
Full ranking
| # | Model | Score | Cost/run | Speed | Best for |
|---|---|---|---|---|---|
| 1 | gpt-5.4-mini | 83.3 Strong | $0.0154 | 13.5s | Strong drafts |
| 2 | glm-5 | 84.2 Strong | $0.0160 | 34.1s | Strong drafts |
| 3 | deepseek-v3.2-low | 78.4 Usable | $0.0163 | 19.3s | Strong drafts |
| 4 | deepseek-v3.2 | 75.6 Usable | $0.0166 | 21.0s | Strong drafts |
| 5 | deepseek-v3.1-terminus | 78.9 Usable | $0.0169 | 22.0s | Strong drafts |
| 6 | deepseek-v3.2-high | 76.3 Usable | $0.0172 | 22.2s | Strong drafts |
| 7 | gemini-3.1-flash-lite | 79.9 Usable | $0.0175 | 13.4s | Strong drafts |
| 8 | grok-4.20-beta | 78.3 Usable | $0.0175 | 13.5s | Strong drafts |
| 9 | gemini-3-flash-preview | 79.0 Usable | $0.0177 | 16.0s | Strong drafts |
| 10 | grok-4.20 | 79.9 Usable | $0.0178 | 12.4s | Strong drafts |
| 11 | kimi-k2.7-code | 86.2 Strong | $0.0182 | 21.6s | Best overall |
| 12 | gpt-5.4-low | 87.1 Strong | $0.0187 | 13.8s | Best overall |
| 13 | mistral-medium-3.1 | 72.8 Usable | $0.0189 | 16.1s | Needs review |
| 14 | qwen3.5-plus-02-15 | 84.3 Strong | $0.0190 | 47.7s | Strong drafts |
| 15 | kimi-k2.5 | 83.0 Strong | $0.0192 | 53.9s | Strong drafts |
| 16 | gpt-5.4 | 85.6 Strong | $0.0193 | 15.9s | Best overall |
| 17 | claude-haiku-4.5 | 72.4 Usable | $0.0193 | 16.4s | Needs review |
| 18 | claude-sonnet-4.5 | 77.4 Usable | $0.0204 | 19.5s | Strong drafts |
| 19 | qwen3.7-max-high | 87.4 Strong | $0.0215 | 41.9s | Best overall |
| 20 | gpt-5.5-low | 86.2 Strong | $0.0215 | 14.6s | Best overall |
| 21 | qwen3.7-max | 86.1 Strong | $0.0224 | 43.0s | Best overall |
| 22 | qwen3.7-max-low | 84.3 Strong | $0.0228 | 43.3s | Strong drafts |
| 23 | gemini-3.5-flash-low | 85.5 Strong | $0.0231 | 18.1s | Best overall |
| 24 | gpt-5.4-high | 86.4 Strong | $0.0233 | 18.5s | Best overall |
| 25 | claude-sonnet-4.5-high | 84.1 Strong | $0.0237 | 22.7s | Strong drafts |
| 26 | claude-sonnet-4.6-low | 72.7 Usable | $0.0240 | 21.5s | Needs review |
| 27 | claude-sonnet-4.5-low | 80.7 Strong | $0.0240 | 21.9s | Strong drafts |
| 28 | gpt-5.5 | 88.3 Strong | $0.0240 | 15.8s | Best overall |
| 29 | claude-opus-4.5 | 82.0 Strong | $0.0247 | 20.3s | Strong drafts |
| 30 | gemini-3.1-pro-preview-low | 89.8 Strong | $0.0257 | 20.7s | Best overall |
| 31 | gemini-3.5-flash-high | 84.8 Strong | $0.0260 | 19.9s | Strong drafts |
| 32 | claude-sonnet-4.6-high | 80.3 Strong | $0.0262 | 23.1s | Strong drafts |
| 33 | gemini-3.1-pro-preview-high | 88.1 Strong | $0.0265 | 23.8s | Best overall |
| 34 | claude-opus-4.8-low | 85.7 Strong | $0.0288 | 18.1s | Best overall |
| 35 | claude-opus-4.8-high | 85.8 Strong | $0.0292 | 19.0s | Best overall |
| 36 | claude-opus-4.5-low | 77.1 Usable | $0.0304 | 22.7s | Strong drafts |
| 37 | gpt-5.5-high | 85.0 Strong | $0.0309 | 24.6s | Best overall |
| 38 | claude-opus-4.6 | 77.4 Usable | $0.0315 | 28.2s | Strong drafts |
| 39 | claude-opus-4.6-low | 74.6 Usable | $0.0320 | 25.4s | Needs review |
| 40 | claude-opus-4.6-high | 75.3 Usable | $0.0325 | 28.8s | Strong drafts |
| 41 | claude-opus-4.5-high | 74.8 Usable | $0.0338 | 27.2s | Needs review |
| 42 | gpt-5-mini | 68.4 Needs editing | $0.0189 | 24.4s | Needs review |
| 43 | minimax-m2.7 | 62.4 Needs editing | $0.0202 | 34.0s | Needs review |
What separates the top models
Resolution
mediumTests resolving an issue with a concrete next step, ownership, and a realistic timeline — grounded in the account facts.
Leader: gemini-3.1-flash-lite
Basic Support Reply Test
mediumTests clear support replies for common user issues.
Leader: claude-opus-4.8-high
De-escalation
mediumTests calming an angry customer with genuine empathy and action — not a robotic, over-apologetic, or dismissive reply.
Leader: gpt-5.4-high
Policy Boundary Test
hardTests refunds, privacy, abuse, security, and account-risk policies.
Leader: kimi-k2.5
Escalation and Incident Test
hardTests outage, SLA, engineering handoff, VIP escalation, and multi-issue tickets.
Leader: gemini-3.1-pro-preview-low
Policy Boundaries
hardTests handling a request for something outside policy without inventing policy or promising what can't be delivered.
Leader: qwen3.5-plus-02-15
Help Content Test
mediumTests help center articles, FAQs, macros, churn rescue, and tone rewrites.
Leader: claude-opus-4.8-low
Escalation & Handoff
hardTests recognizing an out-of-scope/over-authority issue and writing a clean internal handoff.
Leader: gpt-5.4-low
See the difference
Real model outputs and how our judge scored them.
Top score
gemini-3.1-flash-lite
Knows when NOT to escalate (Lumen)
100/100
“The model perfectly followed the instructions by providing the exact UI flow (Settings -> Team -> Invite) without unnecessarily escalating the routine request. The tone is friendly, and the response is concise and clear.”
Hi there, I’d be happy to help you with that! Adding a second care coordinator to your dashboard is a straightforward process. You can do this by following these steps: 1. Log in to your dashboard and click on **Settings**. 2. Navigate to the **Team** tab. 3. Click the **Invite** button and enter the email address of the team member you would like to add. Once you send the invite, they will receive an email with instructions to set up their account. Let me know if you run into any issues or have any other questions!
Mid-pack
claude-opus-4.6-high
Churn rescue
87/100
“The response is excellent, highly empathetic, and perfectly utilizes the provided facts and allowed options without offering a discount. It takes ownership of the issue and provides a clear, concrete resolution path. It is slightly verbose and includes minor unprompted technical details and commitments, but overall reads like a top-tier customer success reply.”
**Subject: We dropped the ball on your onboarding — let us fix this** Hi [Customer Name], Thank you for telling us directly — and I'm sorry. You should not have had to figure out the Salesforce connection on your own, and the experience you described is not the onboarding we intend for any customer. This one is on us. I looked into your account and can see exactly what happened: your Salesforce integration failed twice due to an `invalid_scope` OAuth error. This almost always means that specific OAuth scopes (typically `api`, `refresh_token`, and `offline_access`) haven't been enabled for the connected app in your Salesforce org. It's a fixable configuration issue — not a platform incompatibility — and it's something we should have caught and walked you through on day one. Here is what I'd like to offer to make this right: 1. **Dedicated 30-minute onboarding call** — I'll pair you with a specialist who will screen-share, resolve the Salesforce OAuth error live, and confirm data is flowing correctly. We can schedule this as early as tomorrow at a time that works for your team. 2. **Salesforce permission checklist** — I'm attaching our step-by-step checklist that covers the exact connected-app scopes, API permissions, and profile-level access your Salesforce admin needs to enable. This will also prevent scope errors for the remaining 14 seats on your account. 3. **7-day trial extension** — Since nearly half your trial was spent blocked by this integration issue, I've flagged your account for a 7-day extension so your team gets a fair window to evaluate Pro with Salesforce actually connected. None of this requires you to commit to anything. If after the onboarding call and a full trial with a working integration you still feel it's not the right fit, I'll process t
Lowest score
deepseek-v3.2-high
Multi-issue ticket
14/100
“The model completely misunderstood the task. Instead of writing a customer support reply that separates billing, bug, and feature requests for a user, it wrote a generic internal guide defining what those ticket types are. It fails as a support reply.”
**Multi-Issue Ticket Breakdown** **1. Billing Inquiry** *Issue:* Customer has questions or concerns about charges, invoices, payment methods, or subscription plans. *Action Required:* Verify account details, explain recent charges, process refunds if applicable per policy, update payment method, or adjust subscription. **2. Bug Report** *Issue:* Customer reports unexpected behavior, error messages, crashes, or broken functionality. *Action Required:* Reproduce issue, gather logs/error details, identify affected platform/version, document steps, and escalate to engineering if unresolved. **3. Feature Request** *Issue:* Customer suggests new functionality, enhancement, or product change. *Action Required:* Log request in feature tracking system, note use case and priority based on user impact, and inform customer it has been forwarded to the product team. **Next Steps:** - Address billing first if urgent (e.g., refund request). - Triage bug severity and assign accordingly. - Confirm feature request receipt and provide reference ID if available.
Where models still fail
The most common problems we flagged across all models.
Frequently asked
What is the best AI model for customer support?
In our benchmarks, gemini-3.1-pro-preview-low ranks first for customer support, scoring strong, across 26 test cases.
What is the cheapest good model for customer support?
gpt-5.4-mini is the best value: it clears our quality bar for customer support at $0.015 per run.
Which model is fastest for customer support?
grok-4.20 is the fastest model that still performs well for customer support.
How we test
Each model output is scored by a strict JSON LLM judge, supported by deterministic heuristics, then normalized to a 0-100 score.
Judge: gemini-3.1-pro-preview · 1000 model runs across 8 benchmarks · last tested 2026-06-30
This page is Spring Prompt, running
We just did this for every model. Do it for your prompt.
The rankings above come from running real tasks through real models and scoring every output. Spring Prompt is that same engine — pointed at your prompt, your test cases, and your definition of good.
- Generate test cases from your prompt — no eval set required to start.
- Compare models side by side with quality, cost and latency in one matrix.
- Optimise the winner until the scores say it's ready to ship.
Prompt × model results
12 test cases · 3 evals