Which AI models support function calling in 2026?

All major AI models support function calling (tool use) as of June 2026, including GPT-5, Claude Opus/Sonnet/Haiku, Gemini Pro/Flash, DeepSeek V4, Mistral, Llama 4, Grok, and Cohere Command R. Budget models like GPT-oss 20B and Gemini Flash Lite have more limited function calling support.

Which AI models support vision and image input?

Most flagship models support vision: GPT-5, GPT-4o, Claude Opus/Sonnet, Gemini Pro/Flash, DeepSeek V4, Mistral Large, Grok, and Llama 4 Maverick. Budget models like GPT-4o mini, Gemini Flash Lite, and some smaller open-source models have limited or no vision support.

What is JSON mode and which models support it?

JSON mode ensures the model outputs valid JSON. GPT-5 and GPT-4o have native JSON mode. Claude models can output JSON via structured output or system prompts. Gemini supports JSON mode natively. Most modern LLMs can reliably produce JSON when instructed, but native support varies.

Which AI models have built-in embeddings?

Only Google Gemini models (3.1 Pro, 2.5 Pro, 2.0 Flash, Flash Lite) and Cohere Command R/R+ have built-in embedding endpoints. All other models require separate embedding APIs from providers like OpenAI (text-embedding-3), Cohere, or open-source alternatives.

Which AI models support fine-tuning?

OpenAI offers fine-tuning for GPT-5, GPT-5 mini, GPT-4o, and GPT-4o mini. Google supports fine-tuning for Gemini Pro and Flash models. DeepSeek, Mistral, Cohere, and Meta Llama models also support fine-tuning. Anthropic does not offer fine-tuning for Claude models.

AI Model Capabilities 2026: Which Models Support What Features

Published Jun 6, 2026 · Updated Jun 6, 2026 · By APIpulse

Choosing an AI model isn't just about price anymore. With 34 models across 10 providers, the real question is: which model actually supports the features you need?

We built a comprehensive capabilities matrix comparing all 34 models across 14 features. Here's what we found.

Explore the Full Interactive Matrix

Filter by provider, tier, and specific capability. Sort by any column.

Open Capabilities Matrix →

Key Findings

1. Function Calling Is Nearly Universal

As of June 2026, 31 of 34 models support function calling (tool use). The exceptions are a few budget open-source models (GPT-oss 20B, Llama 3.1 8B) where support is limited. If function calling is critical for your agent or chatbot, almost any modern model will work.

2. Vision Support Is the Real Differentiator

Only 22 of 34 models support image input. Budget models like GPT-4o mini, Gemini Flash Lite, GPT-oss, and Llama 3.1 8B have limited or no vision. If you're building an image analysis pipeline, you'll need to stick with flagship or mid-tier models.

3. Built-in Embeddings: Only Google and Cohere

This is the biggest surprise. Only 6 models have built-in embedding endpoints: all 4 Gemini models and both Cohere Command R models. Everyone else — OpenAI, Anthropic, DeepSeek, Mistral, Meta — requires separate embedding APIs. If you're building RAG and want to minimize API complexity, Gemini or Cohere are your best bet.

4. Batch API: OpenAI and Google Lead

Batch processing (for non-real-time workloads at 50% discount) is available on 16 of 34 models. OpenAI supports it across all their models, and Google Gemini supports it too. Anthropic supports batch for Claude. DeepSeek, Mistral, Meta, and others don't offer batch APIs yet.

5. Fine-Tuning: OpenAI, Google, DeepSeek, Mistral

19 of 34 models support fine-tuning. OpenAI, Google, DeepSeek, Mistral, Cohere, Meta, and AI21 all offer it. Anthropic does not offer fine-tuning for any Claude model — this is a key limitation if you need to customize model behavior for your domain.

6. Streaming: Universal

All 34 models support streaming responses. No exceptions. This is table stakes in 2026.

Feature Support by Provider

Provider	Vision	Functions	Embeddings	Fine-Tune	Batch	JSON Mode
OpenAI	Yes	Yes	No	Yes	Yes	Yes
Anthropic	Yes	Yes	No	No	Yes	Yes
Google	Yes	Yes	Yes	Yes	Yes	Yes
DeepSeek	Yes	Yes	No	Yes	No	Yes
Mistral	Yes	Yes	No	Yes	No	Yes
Cohere	No	Yes	Yes	Yes	No	Limited
Meta (Together.ai)	Partial	Yes	No	Yes	No	Limited
xAI	Partial	Yes	No	No	No	Yes

Best Model by Use Case (Capabilities Matter)

Building a RAG Pipeline?

You need embeddings + large context + function calling. Best picks: Gemini 3.1 Pro (built-in embeddings, 1M context, $2/M), Gemini 2.5 Pro (1M context, $1.25/M), or use any model with a separate embedding API.

Building an AI Agent with Tools?

You need function calling + streaming + large context. Best picks: Claude Opus 4.8 (1M context, excellent tool use), GPT-5 (272K, strong function calling), DeepSeek V4 Pro (1M context, $0.435/M).

Image Analysis Pipeline?

You need vision support. Best picks: GPT-5 ($1.25/M), Gemini 2.5 Pro ($1.25/M), Claude Sonnet 4.6 ($3/M), DeepSeek V4 Pro ($0.435/M — cheapest vision model).

High-Volume Batch Processing?

You need batch API support. Best picks: GPT-4o mini ($0.15/M with batch), Gemini 2.0 Flash ($0.10/M with batch), DeepSeek V4 Flash ($0.14/M).

Custom Domain Model?

You need fine-tuning. Best picks: GPT-4o mini ($0.15/M, cheapest to fine-tune), Gemini 2.0 Flash ($0.10/M), Mistral Small 4 ($0.15/M), Llama 4 Scout (open source, fine-tune freely).

Find Your Perfect Model

Use our interactive matrix to filter by the exact capabilities you need.

Open Capabilities Matrix →

The Bottom Line

Price matters, but capabilities matter more. A $0.10/M model that doesn't support vision won't help you build an image analysis pipeline. A model without function calling can't power your AI agent.

Use our capabilities matrix to find models that match your feature requirements first, then compare prices among those that qualify. That's how you build an AI stack that works — and doesn't break the bank.

Explore the Full Matrix →