Ubby provides access to a carefully curated selection of AI models that have been tested and optimized for agentic workflows. Not every AI model works well in autonomous agent scenarios—some excel at conversational interactions but struggle with tool use, complex multi-step reasoning, or reliable execution of structured tasks. The models available in Ubby represent those that perform best for building and running autonomous agents.

This curated approach means you can trust that any model you select has been validated for agentic use cases. You are choosing among models that all work well with agents; the question is which one best fits your specific task requirements and budget.

The available models span several major families:

Each model family brings different strengths to the table. Your job is not to find the single "best" model but to match models to tasks based on your specific requirements.

When evaluating AI models, several key characteristics determine their suitability for different tasks.

Some models demonstrate superior reasoning ability, handling complex multi-step problems, nuanced analysis, and sophisticated decision-making better than others. These models typically cost more credits per token but deliver higher-quality outputs for demanding tasks.

Claude Sonnet-4, for example, excels at tasks requiring deep understanding, careful reasoning, and nuanced responses. If your agent needs to analyze complex business situations, provide strategic advice, or handle ambiguous instructions, a high-intelligence model justifies its higher cost.

Lighter models may struggle with truly complex reasoning but handle straightforward tasks perfectly well. Using an expensive high-intelligence model for simple tasks wastes credits without adding value.

Different models process requests at different speeds. Some return responses in seconds, while others take longer. For agents that need to respond quickly—like customer service bots or real-time assistants—speed matters significantly.

Generally, smaller models process faster than larger ones. A model like GPT-5-nano delivers responses much quicker than GPT-5, though with reduced capability. Consider whether your use case prioritizes speed over maximum intelligence.

In automated workflows where agents run asynchronously without human waiting, speed matters less. A monthly report that takes 30 seconds versus 10 seconds to generate makes little practical difference. But an interactive agent where users expect immediate responses needs faster models.

The context window determines how much information a model can consider at once. This includes your prompt, any documents or data provided, the agent's instructions, and conversation history.

Models with larger context windows can handle longer documents, maintain more conversation history, or work with extensive background information. If your agents regularly process large documents or need to reference substantial context, context window size becomes a critical selection factor.

However, larger context windows consume more tokens and therefore more credits when fully utilized. Don't default to maximum-context models if your tasks rarely need extensive context.

Some models offer specialized capabilities that make them particularly suited to certain tasks:

Match specialized models to tasks that benefit from their particular strengths.

Every model has different credit costs per million tokens, and these costs vary between input (processing) and output (generation) tokens. Understanding these costs helps you optimize your credit usage.

At the high end, flagship models like Claude Sonnet-4 cost 4,500 credits per 1M input tokens and 22,500 per 1M output tokens. These premium prices reflect exceptional capability and should be reserved for tasks that genuinely benefit from top-tier performance.

Mid-range models like GPT-4.1 (3,000 input / 12,000 output) or DeepSeek-Chat (750 input / 1,500 output) offer strong performance at more moderate costs. These models handle most business tasks well and represent good value for routine agent operations.

At the efficient end, models like GPT-5-nano (75 input / 600 output) or GPT-OSS-20b (60 input / 225 output) provide basic capability at minimal credit consumption. While not suitable for complex reasoning, they excel at straightforward tasks where intelligence requirements are modest.

More expensive models deliver better performance, but the relationship is not linear. A model costing 10x more does not necessarily perform 10x better—the improvement might be 2x or 3x for most tasks.

This creates optimization opportunities. For many routine tasks, a mid-range model delivers 80-90% of a flagship model's quality at 30-50% of the cost. The performance gap matters more for some tasks than others.

Consider what "good enough" means for each task. A monthly executive report might justify premium model costs for maximum quality. A routine data extraction task that produces structured output might work perfectly well with an efficient mid-range model.

To understand true costs, calculate expected credit consumption per task execution:

Example 1 - Complex analysis task:

Example 2 - Same task with DeepSeek-Chat:

The cheaper model saves ~138 credits per execution (87% savings). Whether this tradeoff makes sense depends on whether the quality difference matters for your use case.

Effective model selection involves matching model characteristics to task requirements systematically.

Classify your tasks into categories based on their complexity and requirements:

This classification guides your default model choices, though you should validate with testing.

Create a simple matrix to guide model selection:

Do not assume model selection. Test your assumptions with real tasks:

Run the same task through multiple models and compare results. Does the expensive model deliver meaningfully better output? Or does a mid-range model produce virtually identical results at lower cost?

Pay attention to failure modes. Sometimes cheaper models fail occasionally where expensive models succeed consistently. If these failures create significant problems, the reliability of expensive models justifies their cost. If failures are easily caught and corrected, occasional failures might be acceptable given the cost savings.

Validate across multiple examples, not just one. A model might handle one example well but struggle with others that have different characteristics.

Certain patterns emerge across successful Ubby deployments for different types of agents.

For complex document analysis (contracts, technical documents, nuanced content):

For straightforward extraction (pulling data fields, categorizing, summarizing):

For simple document classification:

For client-facing content (reports, proposals, communications):

For internal documentation:

For template-based content (filling forms, standard letters):

For complex analytical reasoning:

For routine metrics and reporting:

For data extraction and transformation:

For sophisticated advisory conversations:

For customer service and support:

For simple FAQ and routing:

Track how different models perform in your specific use cases to refine selection over time.

Your Ubby usage logs show which models consumed how many credits over time. Review this data monthly to understand your model cost distribution:

For critical agents, track quality metrics alongside costs:

If a cheaper model requires frequent human correction, its nominal cost savings might be illusory once you account for the human time spent fixing errors.

Model selection should not be static. As new models become available, as your tasks evolve, and as you gain experience with different models, continuously refine your choices:

The AI model landscape evolves rapidly. Design your agent architecture to adapt easily as new models emerge.

Rather than hard-coding specific models into your agents, use configuration that can be easily updated. When a new, better model becomes available, you can switch with minimal effort.

Some organizations maintain model "tiers" (premium, standard, efficient) and assign agents to tiers rather than specific models. When a better model joins a tier, all agents in that tier benefit automatically.

AI providers regularly release improved models with better performance, lower costs, or new capabilities. Stay informed about these releases and evaluate whether they offer advantages for your use cases.

Ubby adds new models to the platform as they become available. When you see a new model appear in your model pricing page, investigate whether it might outperform your current selections.

Document which models work well for which tasks in your organization. This knowledge helps new team members make good model selections and prevents repeatedly testing the same model/task combinations.

Share learnings across your team about model performance. Someone might discover that a particular model excels at a specific task type, knowledge that benefits everyone.

You now understand the different AI models available in Ubby, their characteristics, and how to select the right model for each task. This knowledge enables you to optimize both the quality of your agents' work and the efficiency of your credit usage.

In the next article, we will explore Ubby's pricing plans and billing system in detail, helping you choose the right plan for your needs and understand how billing works.