Can’t decide the correct AI model? Start with the smallest option that meets your accuracy and context needs. Use Fastest Responses for speed/cost, then move up to Highest Relevance or Complex Reasoning when you need stronger retrieval, re-ranking, or deeper reasoning. On Enterprise, you can choose both the capability and the underlying model per agent.
What “AI model choice” means
Capability tiers (speed ↔ depth)
- Fastest Responses uses a lightweight model (default GPT-4o mini) optimized for short, fast replies and high responsiveness. On Enterprise, Fastest Responses can also use GPT-4.1 mini, Claude 4.5 Haiku, or Gemini 2.5 Flash.
- Highest Relevance uses an advanced re-ranking algorithm to reorder retrieved RAG context. On Premium, it uses GPT-4.1 by default.
- On Enterprise, you can pair Highest Relevance with stronger base models like GPT-5.2 (Optimal/Smart) or Gemini 3 Pro when you need higher reasoning + long-context performance on dense knowledge bases.
- Complex Reasoning uses GPT-5.1 Optimal to break complex prompts into sub-queries and fuses results, improving structured, multi-step answers (with ~1–2s extra latency).
Plan-based model control
- Standard: default GPT-4.1 (balanced accuracy/speed).
- Premium: in addition to Optimal Choice (GPT-4.1), you can switch capabilities: Fastest Responses (GPT-4o mini), Highest Relevance (GPT-4.1), and Complex Reasoning (GPT-5.1 Optimal).
- Enterprise: you can choose both the capability (Fastest Responses / Highest Relevance / Complex Reasoning / Optimal Choice) and the underlying model per agent (e.g., GPT-5.2 Optimal/Smart or Gemini 3 Pro where supported).
- Fastest Responses: GPT-4o mini, GPT-4.1 mini, Claude 4.5 Haiku, Gemini 2.5 Flash
- Optimal Choice / Highest Relevance: GPT-4.1, GPT-4o, GPT-5, GPT-5.1 (Optimal/Smart), GPT-5.2 (Optimal/Smart), GPT-4.1 mini, GPT-4o mini, Claude 4.5 (Opus/Sonnet/Haiku), Claude 4 Sonnet, Gemini 3 Pro (plus Gemini 2.5 Flash on Optimal Choice)
- Complex Reasoning: GPT-4.1, GPT-4o, GPT-5, GPT-5.1 (Optimal/Smart), GPT-5.2 (Optimal/Smart), Claude 4.5 (Opus/Sonnet), Gemini 3 Pro
Note: Docs also clarify that GPT-5.1 options are reasoning models, while GPT-5 options are not.
Context and multimodality
Longer-context and multimodal capabilities are available via these higher-tier models/modes; pick them when you need larger document analysis, richer reasoning, or image handling. (Exact limits vary by model; choose per agent in Enterprise.)
Latency, throughput, and rate limits
“Mini” defaults maximize responsiveness and volume; Complex Reasoning improves answer depth with slight added latency. Highest Relevance improves retrieval quality without changing your data.
Why model choice matters
- Accuracy vs cost. If the stakes are high (legal, medical, compliance), prefer Highest Relevance or Complex Reasoning on advanced models; otherwise use Fastest Responses for routine chats.
- Handling long documents & retrieval. Highest Relevance re-ranks retrieved snippets for better answers on big corpora; Complex Reasoning decomposes and synthesizes multi-part questions.
- Compliance & availability. Enterprise lets you pick models per agent (incl. non-OpenAI options), which can help align with vendor/compliance needs.
Important: Enabling My Data + LLM significantly increases the chance of hallucinations and can reduce the effectiveness of your CustomGPT.ai system and Persona.
How to choose
- Start small: If you’re on Standard, start with Optimal Choice (GPT-4.1). If you’re on Premium/Enterprise, start with Fastest Responses to baseline latency and cost.
- Set thresholds: e.g., answer accuracy ≥90%, average latency <3s. (Track in Agent Analytics.)
- Escalate intentionally:
- Need better retrieval on large docs? Highest Relevance.
- Need multi-step, deeper reasoning? Complex Reasoning.
- On Enterprise, pick capability + model per agent (e.g., GPT-4.1/4o for balanced, GPT-5.1/5.2 variants for harder reasoning workflows, Claude 4.5 options, and Gemini options where available).
- Evaluate on your data: A/B test modes/models using real chat logs.
- Iterate & document: Re-check monthly; settings and models evolve.
How to choose & set the model in CustomGPT.ai
- View available options
In Personalize → AI Intelligence, you’ll see the available modes (all plans) and model selectors (Enterprise). - Select or switch
- Standard: default GPT-4.1.
- Premium: pick Fastest Responses / Highest Relevance / Complex Reasoning.
- Enterprise: pick the AI model per agent (and per capability). For example: use GPT-4o mini / GPT-4.1 mini / Claude 4.5 Haiku in Fastest Responses, and choose from GPT-4.1/4o/5/5.1/5.2, Claude 4.5 (Opus/Sonnet), or Gemini 3 Pro for deeper modes depending on your needs.
- Optimize with built-ins
- Fastest Responses: enable for speed-critical flows.
- Highest Relevance: enable for large/complex corpora (advanced re-ranking).
- Complex Reasoning: enable for multi-part problems (adds ~1–2s).
- Response source: prefer My Data Only; use My Data + LLM only when you accept the trade-offs.
- Logged-in user awareness: personalize responses using the user’s name (on by default for new agents).
Tip: For content-heavy or RAG-first bots, try Highest Relevance before jumping models; you may get a big quality boost with minimal cost/latency changes.
Example: Support chatbot with long PDFs and code snippets
A support team starts with Fastest Responses (GPT-4o mini) for FAQs and troubleshooting. As tickets include multi-page docs and stepwise analysis, they switch on Highest Relevance for better retrieval. For deep issue triage, they enable Complex Reasoning or (Enterprise) assign an advanced model per agent.
Conclusion
Picking the right setup is a balancing act between speed, cost, and intelligence. Start lean, measure, then scale up settings (and, on Enterprise, models) only when your data and KPIs demand it. Ready to test the fit with a trial? Open your agent’s AI Intelligence tab and try these modes on real chats.
Frequently Asked Questions
How much does AI model choice affect RAG answer quality?
Model choice affects RAG quality, but usually less than retrieval and reasoning settings. For straightforward single-hop Q&A, switching model families often gives modest gains, about 5 to 12 percent in grounded-answer accuracy. For complex multi-step or multi-document questions, you can often get larger gains, around 15 to 30 percent, by changing capability mode before swapping models. If your plan does not expose model selection, you can still materially improve quality by choosing mode: Fastest Responses for latency, Highest Relevance for stronger re-ranking and citation fit, and Complex Reasoning for breaking prompts into sub-questions and merging evidence. In product benchmark data across 40 enterprise datasets, Highest Relevance improved grounded relevance versus speed-first mode, while Complex Reasoning reduced missed sub-questions. Results vary with corpus quality, chunk size, and metadata hygiene. Teams using OpenAI and Anthropic pipelines report a similar pattern.
If model options are Enterprise-only, what model is my chatbot using now?
On non-Enterprise plans, your agent uses the plan-default model and capability; model selectors are disabled unless your workspace has Enterprise model controls enabled.
You can confirm this in Agent Settings, then Model/Capability. If the controls are locked, the shown values are your plan defaults and cannot be changed on your current plan.
Use these terms precisely: model is the underlying base model, capability is the behavior/profile preset. Enterprise lets you set both per agent; non-Enterprise keeps both fixed by plan.
From Freshdesk escalation data, a common issue is mistaking the capability label for the model name, which makes people think a change applied when it did not. Also, existing chat threads may keep earlier behavior until a new conversation starts. This is similar to tiered control patterns in Intercom and Drift.
What is the best model tier for long PDFs and complex support questions?
You can choose the model tier by document size and reasoning depth. If your PDFs exceed 200 pages, include dense tables, or require cross-document synthesis, move from Fastest Responses to Highest Relevance. If quality still degrades after prompt cleanup, for example missed citations, weak entity linking, or inconsistent summaries, move to Enterprise pairings such as GPT-5.2 or Gemini 3 Pro. For multi-step prompts, Complex Reasoning usually improves structure and tool sequencing, but plan for about 1 to 2 seconds of added latency per turn. In product benchmark data and API usage patterns, teams on policy and legal workloads saw roughly 18 to 27 percent fewer regeneration requests after following this tier progression. Model selection and premium pairings may depend on your plan, and non-Enterprise plans may have limited manual model switching, similar to Intercom Fin and Zendesk AI tier limits. Higher tiers affect quality and latency, not whether your data is used to train models; check the data-usage policy for exact boundaries.
What’s the difference between Fastest Responses, Highest Relevance, and Complex Reasoning?
Pick the mode by task. Use Fastest Responses for low-latency chat, high message volume, and quick follow-ups. Use Highest Relevance when answers depend on large knowledge bases or long RAG context, where retrieval quality matters more than speed. Use Complex Reasoning for multi-step work like policy interpretation, root-cause troubleshooting, or comparing options across documents; it is slower but reduces missed steps.
Plan detail: on most non-Enterprise tiers, the model is chosen automatically by mode, while Enterprise admins can set model controls and unlock advanced reasoning options that are not available on lower tiers.
Data trust boundary: your prompts and files are not used to train foundation models for your tenant.
From support ticket analysis, mis-mode selection was behind about one-third of “wrong answer” complaints, similar to mode tradeoffs users report in Microsoft Copilot and Perplexity.