Default pick
Claude Haiku 4.5. $1 input / $5 output per 1M tokens. Excellent tone, sub-second response. Covers 80% of SMB chatbot use cases.
AI Model Selection Guide.
How Fabi AI Labs picks between OpenAI, Anthropic, and Google models — based on volume, tone, cost, and honest tradeoffs. Updated April 2026.
How to read this.
This guide is auto-indexed into every Fabi AI Labs chatbot. When a user asks “which model should I pick?” the bot retrieves and answers from this document. You're reading the same source material.
TL;DR
Three bullets if you only read one thing.
High volume
Switch to Gemini 2.5 Flash above ~30,000 conversations/month. Half the cost at similar quality, 0.63s time-to-first-token.
Quality first
Switch to Claude Sonnet 4.6 when conversations are complex and volume stays under 20,000/ month. Near-flagship quality for the price.
Pricing snapshot
Every current model, every rate, April 2026.
Per 1 million tokens. Figures verified against official provider pages.
| Model | Input | Output | Context | Status |
|---|---|---|---|---|
| OpenAI GPT-5.4 | $2.50 | $15.00 | 1M | current flagship |
| OpenAI GPT-5.4 mini | $0.75 | $4.50 | 400k | current mid-tier |
| OpenAI GPT-5.4 nano | $0.20 | $1.25 | 128k | current budget |
| OpenAI GPT-4o mini | $0.15 | $0.60 | 128k | legacy |
| Anthropic Claude Opus 4.7 | $5.00 | $25.00 | 1M | current flagship |
| Anthropic Claude Sonnet 4.6 | $3.00 | $15.00 | 1M | current mid-tier |
| Anthropic Claude Haiku 4.5 | $1.00 | $5.00 | 200k | current budget/fast |
| Google Gemini 2.5 Pro | $1.25 | $10.00 | 1M | production stable |
| Google Gemini 2.5 Flash | $0.30 | $2.50 | 1M | current mid-tier |
| Google Gemini 2.5 Flash-Lite | $0.10 | $0.40 | 1M | budget |
⚠ Gemini 2.0 Flash shuts down June 1 2026. Do not build on it.
Quality snapshot
Arena ELO — human preference, 5.87M votes.
Pulled April 19 2026 from arena.ai/leaderboard. Lower rank = more preferred in head-to-head chat.
| Model | Arena ELO | Rank | Notes |
|---|---|---|---|
| Claude Opus 4.7 Thinking | 1504 | #1 | Best overall |
| Claude Opus 4.7 | 1497 | #3 | Top non-thinking |
| GPT-5.4 | 1467 | #18 | Highest benchmark intelligence |
| Claude Sonnet 4.6 | 1463 | #20 | Strong tone fidelity (verbose) |
| Gemini 2.5 Pro | 1448 | #35 | Solid production choice |
| Gemini 2.5 Flash | 1411 | #82 | Best quality-per-dollar at speed |
| Claude Haiku 4.5 | 1408 | #85 | Competitive with Flash |
| GPT-4o | 1345 | #159 | Aging rapidly |
Speed snapshot
How fast do they actually feel?
Under 1s feels instant · 1–2.5s feels normal · over 3s kills conversions.
| Model | Output speed | Time-to-first-token | UX verdict |
|---|---|---|---|
| Gemini 2.5 Flash-Lite | 393 t/s | 0.29s | Instant |
| Gemini 2.5 Flash | 208 t/s | 0.63s | Near-instant |
| Claude Haiku 4.5 | 80–120 t/s | ~0.9s | Fast |
| GPT-4o mini | 100–200 t/s | ~0.7s | Fast |
| Claude Sonnet 4.6 | 44 t/s | 1.37s | Acceptable |
| GPT-5.4 | 90 t/s | ~2.0s | Acceptable |
Reasoning models (o3, o4-mini) are NOT recommended for live chat — TTFT of 60+ seconds. Built for batch, not conversation.
Cost by scenario
Four common deployments, four picks.
Assumptions: 500 input tokens + 300 output tokens per chat turn — typical for RAG chatbot.
Cheapest possible, high volume
50,000 chats/month, willing to trade some quality for cost.
Winner: Gemini 2.5 Flash — $45/month.
GPT-4o mini is cheaper but declining in quality. Flash delivers nearly Haiku-level quality at half Haiku's cost with faster response. At this volume, it's the clear pick.
Balanced, quality matters, mid-volume
10,000 chats/month, tone matters, budget-aware.
Winner: Claude Haiku 4.5 — $20/month.
$11 more per month than Flash buys noticeably better tone fidelity — worth it for customer-facing brand work.
Premium experience, small volume
1,000 chats/month, willing to pay top dollar.
Winner: Claude Sonnet 4.6 — $6/month.
At this volume, cost differences are trivial. Sonnet 4.6 delivers near-flagship quality with strong tone fidelity. Opus is only $4/month more if you want the absolute ceiling.
“I don’t know, you pick”
Wide-range SMB chatbot, you want a confident default.
Winner: Claude Haiku 4.5.
It covers the widest range of SMB deployments without requiring a pricing conversation. $1/$5 per 1M tokens, sub-second response, excellent tone, 200k context. Most clients at 5k–15k chats/month will spend $10–30/month and get consistent results. Revisit at 40k+ chats/month — switch to Gemini 2.5 Flash to halve costs.
Provider strengths and weaknesses
Honest view of each of the three.
OpenAI
Wins
- Most mature API ecosystem, best docs, widest integration support
- GPT-5.4 mini is a compelling all-rounder at $0.75 input
Losses
- Model naming is fragmented (5.4, 5.4 mini, 5.4 nano, 5.2, 4.1, 4o…) — confusing to explain
- GPT-5.4 flagship at $15/1M output is expensive for chat
Anthropic
Wins
- Best conversational tone across all tiers — dominates Arena for human preference
- Haiku 4.5 punches far above its $1/$5 price
- Prompt caching can cut repeated-context costs by up to 90% — huge for RAG
Losses
- Haiku’s 200k context is fine for chatbots but smaller than rivals’ 1M
- Sonnet 4.6 is verbose — output token cost creeps up at scale
Wins
- Gemini 2.5 Flash at 208 t/s + 0.63s TTFT + $0.30/$2.50 is exceptional for high-volume
- 1M context at every tier
- Flash-Lite at $0.10/$0.40 is the cheapest non-deprecated option
Losses
- Gemini feels slightly more mechanical for nuanced tone vs Claude
- Google history of fast deprecations (see Gemini 2.0 Flash)
Final recommendation
Default to Haiku 4.5. Don't overthink it.
It solves 80% of SMB chatbot cases cleanly. If the client has a strong opinion or volume concern, use the scenario matrix above to pick a different model — but every model listed here is capable. The pick matters less than shipping.
Sources
Every number verified.
All figures verified against official provider pages as of April 22 2026. Pricing subject to change.
The right model is the one that gets your chatbot shipped. Message us with your volume, tone needs, and timeline — we'll reply with a scoped pick.
