The business guide to AI models.
Search a dated reference of fourteen models, price a real workload, check whether an open model runs on your hardware, and leave with a plan to test. Confirm current prices with each provider before you buy.
Showing 6 of 14 models
Strengths
- Natural writing style
- Multimodal capabilities
- Largest ecosystem
- Reliable at scale
Best for
Teams needing a versatile, well-supported model for content generation and general tasks.
Strengths
- Best-in-class coding
- Strong reasoning
- Long context
- Safety-first design
Best for
Complex coding tasks, detailed analysis, and applications requiring instruction precision.
Strengths
- 1M token context
- Multimodal
- Good value
- Strong quality
Best for
Processing large documents, long transcripts, or extensive datasets in a single pass.
Strengths
- Great value
- OpenAI ecosystem
- Function calling
- Structured output
Best for
High-volume tasks where cost matters more than peak quality.
Strengths
- Best safety in class
- Precise instructions
- Fast responses
- Low cost
Best for
Customer support bots and brand-safe conversational AI.
Strengths
- Very low cost
- Strong benchmarks
- CoT reasoning
- Open-weight roots
Best for
Budget-conscious teams needing strong reasoning without premium pricing.
Reference snapshot, May 2026. Model names, prices and limits change often; treat this as a starting point.
Open-source or API?
Two ways to the same result. Budget, your team and how much control you need decide which one fits.
Open-source, self-hosted
You download the model and run it on hardware you control.
- Data stays with you: nothing has to leave your servers.
- Fixed costs: you pay for hardware and upkeep, not per token.
- Customizable: it can be tuned on your own data.
- No lock-in: you can swap models when a better one ships.
Best for: Teams with GPU capacity, strict data rules, or enough volume that per-token pricing gets expensive.
Hosted API
You send requests to a provider and pay for what you use.
- Fast to start: an API key and a few lines of code.
- Kept current: the provider handles upgrades.
- Scales on its own: traffic spikes are their problem.
- Lower barrier: no machine learning team needed.
Best for: Small teams that want results without running infrastructure, and most first projects.
Will it run on your machine?
Name a device, pick an open model, and see whether there is enough memory and speed headroom to run it locally.
Llama 3.1 8B
Llama 3.1 8B should run locally, but keep context and parallel users modest.
- Weights
- 4.7GB
- KV cache
- 0.9GB
- Runtime
- 1.6GB
- Speed margin
- +0.5
Good portable setup for smaller local models and single-user testing.
Use MLX or Ollama and close memory-heavy apps before running long context windows.
Better fits on this device
Side by side.
Price per million tokens, in and out. On a phone, the table scrolls sideways.
| Model | Provider | Type | Input per 1M | Output per 1M | Context | Quality | Speed |
|---|---|---|---|---|---|---|---|
| GPT-4o | OpenAI | API | $2.50 | $10.00 | 128K | Excellent | Fast |
| Claude 3.5 Sonnet | Anthropic | API | $3.00 | $15.00 | 200K | Excellent | Fast |
| Gemini 1.5 Pro | API | $1.25 | $5.00 | 1M | Very Good | Fast | |
| DeepSeek-V3 | DeepSeek | API | $0.27 | $1.10 | 128K | Very Good | Fast |
| GPT-4o mini | OpenAI | API | $0.15 | $0.60 | 128K | Good | Very Fast |
| Claude 3 Haiku | Anthropic | API | $0.25 | $1.25 | 200K | Good | Very Fast |
| Gemini 1.5 Flash | API | $0.075 | $0.30 | 1M | Good | Very Fast | |
| Qwen3-Coder | OpenRouter | Free | Free | Free | 128K | Good | Varies |
| Kimi-K2.5 | OpenRouter | Free | Free | Free | 128K | Good | Varies |
| Llama 3.1 405B | Meta | Open Source | Self-host | Self-host | 128K | Excellent | Varies |
| Mistral Large 2 | Mistral | Open Source | Self-host | Self-host | 128K | Very Good | Varies |
Published API rates as of May 2026. Quality and speed are our working read, not benchmark scores. Open-source costs depend on where you host.
What will it cost to run?
Pick a job and a model to see a monthly estimate from the reference rates. Check current provider terms before you decide.
- Per task
- Under $0.01
- Per month
- $75.00
- Per year
- $900.00
- Tasks a month
- 15,000
Reference rates as of May 2026. Real costs move with prompt length, caching, retries and review time. Free tiers carry rate limits and change often.
Plan the first test.
Four questions about the work, the review rule and where it can run. You get a short evaluation plan, not a model pick.
What is an LLM and why should my business care?
A Large Language Model (LLM) is an AI system trained on text and other data to understand and generate language. For businesses, LLMs can help summarize documents, qualify leads, draft replies, write code, and support customer operations. The useful question is not whether a model can generate text, but whether it improves a specific workflow with the right controls and review.
What is the difference between open-source and API models?
API models like GPT-4o or Claude are hosted by the provider: you send requests and pay under their current terms. Open-weight models like Llama or Qwen can be run on infrastructure you control, but hardware, hosting, operations, licensing, and model availability still shape the cost and flexibility. The trade-off is that self-hosting requires technical expertise and capacity planning.
Are free LLM models good enough for production use?
Free access can be useful for prototyping, but availability, rate limits, data terms, and model versions change often. Treat a free tier as a way to evaluate a narrow workflow, then confirm the provider terms and operational requirements before using it for customer-facing work.
How do I calculate the cost of running an LLM for my business?
LLM costs are usually measured in input and output tokens. Start with: tokens per task, times tasks per month, times the provider rate per million tokens, then add retrieval, hosting, monitoring, and human review where relevant. The cost calculator on this page gives a planning estimate, not a provider quote.
Which model should I use for customer support automation?
For customer support, prioritize instruction-following, escalation behavior, cost, and a clear approval boundary. Test candidate models against real but safely handled conversations, including edge cases, before routing a customer-facing workflow automatically.
What is OpenRouter and how does it offer free models?
OpenRouter is a unified API gateway that routes requests to multiple AI model providers through a single endpoint. Some models are offered free during promotional periods or because providers subsidize access to gain market share. The free tier typically has rate limits and may not offer the same SLAs as paid plans.
Can I switch between models without rewriting my application?
Yes, and you should design for it. Most LLM APIs follow the OpenAI chat completions format, so switching from GPT-4o to Claude to an open-source model requires minimal code changes. Tools like LiteLLM, OpenRouter, and the Vercel AI SDK abstract provider differences behind a single interface.
How does AnovaGrowth help businesses implement AI models?
We help teams evaluate the workflow, model options, source data, review boundary, and delivery approach before deciding what to build. The appropriate scope can include model evaluation, integration planning, automation design, or a controlled implementation, depending on the work and written engagement.
Tell us the job. We help you test a few models against your real work, then write down what it will cost to run and who reviews it.