Available Models
AI CTRL provides access to models from four providers: OpenAI, Anthropic, Google, and Perplexity. This page groups them by the four types described in Choosing a Model.
For what each model costs to run, see the AI CTRL Ratecard.
Your list will be shorter than this one. Your administrators choose which models appear in your dropdown, and most deployments enable a curated handful rather than everything available. If something here isn't in your list, ask them: it's a deliberate choice, not an oversight.
Within each type below, models are listed roughly from most economical to most expensive. GPT-5 mini is the standard baseline that credit allocations are measured against.
Everyday / lightweight
General questions, drafting, summarizing, rewriting, quick lookups. Fast and inexpensive: these handle more than people expect, so try one first and move up only when the answer isn't good enough.
| Model | Provider | Best for |
|---|---|---|
| GPT-5 nano | OpenAI | Summaries, tagging, and high-volume triage |
| Gemini 2.0 Flash-Lite | Simple input–output and minimum spend per request | |
| Gemini 2.0 Flash | Balanced multimodal assistant | |
| Gemini 2.5 Flash-Lite | Routing, translation, and extremely high request volume | |
| GPT-4.1 nano | OpenAI | High-volume classification, routing, and low-latency tasks |
| GPT-4o mini | OpenAI | Budget multimodal chat, vision, and simple extraction |
| GPT-5.4 Nano | OpenAI | Snippets, small fixes, classification, and routing |
| GPT-5 mini ★ | OpenAI | General chat, analysis, and light coding |
| Claude Haiku 3 | Anthropic | Maximum economy and minimum latency at high scale |
| Gemini 3.1 Flash-Lite | High-throughput tasks where cost and speed both matter | |
| Gemini 2.5 Flash | Chat, summarization, and extraction at large scale | |
| GPT-4.1 mini | OpenAI | Cost-effective coding, extraction, and everyday assistants |
| GPT-5.4 Mini | OpenAI | Quick iterations, lighter coding, and tool-heavy workflows |
| Claude Haiku 4.5 | Anthropic | High-volume assistants and chat with strong answer quality |
| Gemini 3.5 Flash | Fast reasoning, multimodal agents, and large-scale production workloads | |
| GPT-5.6 Luna | OpenAI | Classification, summarization, lightweight chat, and high-throughput automation |
| GPT-4o | OpenAI | Vision and general multimodal assistants |
| Gemini 3.6 Flash | Agentic and multimodal workloads: stronger coding and knowledge work than 3.5 Flash, at lower cost | |
| Gemini 3.7 Flash | Fast everyday work at large context | |
| GPT-5.6 Terra | OpenAI | Everyday professional work: the balanced default where Sol is overkill |
Reasoning
Multi-step problems, analysis, technical writing, and code. Slower and more verbose than a short question needs, and the most expensive per credit: reserve them for work that needs the thinking.
| Model | Provider | Best for |
|---|---|---|
| o4-mini | OpenAI | Cost-efficient math, logic, and structured reasoning |
| GPT-5 | OpenAI | Reasoning, coding, and general-purpose assistants |
| GPT-5.1 | OpenAI | Everyday coding and agent-style tasks |
| GPT-5.1 Codex | OpenAI | Economical coding, implementation, and code review |
| GPT-5.2 | OpenAI | Coding, tool use, and multimodal workflows |
| GPT-5.2 Codex | OpenAI | IDE-style edits, debugging, and repo-aware changes |
| GPT-5.3 Codex | OpenAI | Agentic coding, refactors, and multi-file software work |
| o3 | OpenAI | STEM, logic, and careful step-by-step analysis |
| Gemini 3 Pro | Complex analysis, coding, and multimodal assistants (preview) | |
| Gemini 3.1 Pro | Agents, coding, and multimodal reasoning (preview) | |
| GPT-5.4 | OpenAI | Coding, agents, math, and broad analysis |
| Claude Sonnet 4 | Anthropic | Implementation, documentation, and general software tasks |
| Claude Sonnet 5 | Anthropic | Agents, coding, and near-Opus quality at lower cost |
| Claude Opus 4.5 | Anthropic | Elite reasoning, research depth, and strategic analysis |
| Claude Opus 4.7 | Anthropic | Stronger coding, vision, and complex multi-step tasks |
| Claude Opus 4.8 | Anthropic | Most capable Opus — complex reasoning and long-horizon agentic work |
| Claude Opus 5 | Anthropic | Complex agentic coding, autonomous workflows, and advanced research where accuracy outweighs cost |
| GPT-5.5 | OpenAI | Flagship for coding, agents, computer use, and professional work |
| Claude Opus 4 | Anthropic | Broad flagship-quality coding, analysis, and writing |
| Claude Opus 4.1 | Anthropic | Premium coding, instruction following, and high-stakes work |
| GPT-5 Pro | OpenAI | Deep analysis and planning with extended thinking |
| GPT-5.2 Pro | OpenAI | Very demanding reasoning, long-form analysis, complex decisions |
| GPT-5.4 Pro | OpenAI | Hardest problems, maximum precision, long-running work |
| GPT-5.5 Pro | OpenAI | Highest-precision reasoning, audits, and high-stakes professional work |
| GPT-5.6 Sol | OpenAI | Agentic multi-step tasks, research synthesis, and high-stakes technical workflows |
Large context
Long documents, large data sets, and anything used with the Structured Data Tool. Use when the volume is genuinely large, not by default.
| Model | Provider | Best for |
|---|---|---|
| Gemini 3 Flash | Fast multimodal work and thinking-style tasks at large context | |
| Gemini 2.5 Pro | Complex code, deep analysis, and very long context | |
| GPT-4.1 | OpenAI | Strong coding and long-context instruction following |
| Claude Sonnet 4.5 | Anthropic | Product engineering and coding across long sessions, up to 1M with opt-in context |
| Claude Sonnet 4.6 | Anthropic | Coding, agent planning, and drafting with strong long-context reasoning |
| Claude Opus 4.6 | Anthropic | Hardest agents, complex coding, and very long documents |
Several newer models also handle very long context and are listed above under their primary use rather than repeated here: the GPT-5.6 family (Luna, Terra, and Sol) at 1.05M tokens, and Claude Opus 5 and Gemini 3.6 Flash at 1M. Any of them will take a large document.
Specialized
Single-purpose models. Availability depends on your organization's configuration.
| Model | Provider | Best for |
|---|---|---|
| Sonar | Perplexity | Web-grounded answers |
| Sonar Reasoning | Perplexity | Structured analysis grounded in live search results |
| Sonar Pro | Perplexity | Multi-step research questions, broad sources, nuanced synthesis |
| Sonar Deep Research | Perplexity | Long reports, due diligence, and literature-style synthesis |
| Sonar Reasoning Pro | Perplexity | Deep multi-step reasoning over retrieved web evidence |
| o4-mini Deep Research | OpenAI | Faster, more affordable deep research |
| o3 Deep Research | OpenAI | Long reports, due diligence, and literature-style synthesis |
| DALL-E 3 | OpenAI | Creative illustrations, concepts, and marketing-style imagery |
| GPT Image 1.5 | OpenAI | Strong image quality, variations, and multimodal workflows at scale |
| GPT Image 2 | OpenAI | Latest image generation with rich detail, edits, and complex scenes |
The Smart Model Router
Not a model: a dispatcher. It reads your question and assigns whichever model suits it, choosing only from those you already have access to. For most people doing most things, this is the right default.
Model retirements
Providers retire their own models on their own schedules, and a retired model is removed from AI CTRL when that happens. Expedient gives notice ahead of any retirement and works with your administrators through the change.
If you've built a custom model, check what it runs on. A custom model is locked to a single base model: when that base model retires, the custom model stops working until it's pointed at a replacement. Review anything your team depends on whenever a retirement is announced.
New models
New models are added roughly every two weeks, typically two to four weeks after a provider's public release, once they've been tested against our security requirements. They arrive restricted by default, so your administrators decide who gets access to each one.