AI Models: A Comparative Overview
Choosing a model is a trade-off between performance, cost, and control. The landscape splits into two families, and most real deployments end up using both:
- Cloud-based API models — managed by providers such as OpenAI, Google, and Anthropic. State-of-the-art capability with minimal setup, billed per token.
- Local and open-source models — run on your own hardware for privacy, control, and no per-token cost.
This overview is a decision aid, not a live price sheet. Model names, IDs, and rates change frequently — treat every number below as approximate and confirm against the provider’s current documentation before committing to production.
Cloud-based API models
Best when you need top-tier capability, scale, and near-zero infrastructure. Pricing is quoted per 1M tokens; input tokens are typically cheaper than output tokens, and features like prompt/context caching can shift the effective cost.
| Provider | Tier | Model (API ID) | Approx. price (per 1M tokens) | Best for |
|---|---|---|---|---|
| OpenAI | Premium | gpt-4.1 (gpt-4.1) |
~$6–12 (output) | High-stakes content, agents, complex reasoning |
| OpenAI | Mid-tier | gpt-4.1-mini (gpt-4.1-mini) |
~$1.60–3.20 (output) | Coding, document QA, enterprise tasks |
| OpenAI | Budget | gpt-4.1-nano (gpt-4.1-nano) |
~$0.40–0.80 (output) | Chatbots, short inference, high-volume tasks |
| Anthropic | Premium | Claude 4 Opus (claude-4-opus) |
~$15 in / $75 out | Agentic tasks, legal, safety, deep reasoning |
| Anthropic | Mid-tier | Claude 4 Sonnet (claude-4-sonnet) |
~$3 in / $15 out | Summarization, chat, enterprise workloads |
| Anthropic | Budget | Claude 4 Haiku (claude-4-haiku) |
~$0.25 in / $1.25 out | FAQ bots, retrieval, customer service |
| Premium | Gemini 2.5 Pro (gemini-2.5-pro) |
~$10 (output, ≤200k ctx) | Reasoning, coding, retrieval, long context | |
| Value | Gemini 2.5 Flash (gemini-2.5-flash) |
~$0.30 in / $2.50 out | Large-scale, low-latency summarization | |
| Budget | Gemini 2.5 Flash-Lite (gemini-2.5-flash-lite) |
~$0.10 in / $0.40 out | Inference at scale, high-frequency chat | |
| Cohere | Value | Command R+ (command-r-plus) |
~$3.50 (output, est.) | Document QA, enterprise retrieval |
| Perplexity | Specialized | pplx-7b-online, pplx-70b-online | ~$1.00 (combined) | Live web data, research, verifiable answers |
| DeepSeek | Budget | DeepSeek-V2 (deepseek-v2) |
~$0.10–0.50 (output) | Retrieval, bulk tasks, cost-sensitive processing |
Reading the table: match the tier to the hardest task in your pipeline, not the average one. A cheap model that fails 10% of the time on complex reasoning often costs more in rework than a premium model that succeeds on the first pass. Where a workflow mixes easy and hard steps, route each step to the cheapest model that can handle it (see the routing pattern in Oh My Claude Code).
Local and open-source models
These run on your own hardware — total privacy, full control of the pipeline, and no per-token fee beyond electricity. Ideal for fine-tuning on sensitive data and for high-volume workloads where API costs dominate.
| Category | Model family (developer) | Primary strength |
|---|---|---|
| Top-tier generalist | Llama 3 (Meta) | Strong all-around open performer for reasoning, writing, and instruction-following |
| Multilingual generalist | Qwen2 (Alibaba) | A Llama 3 competitor with exceptional multilingual strength |
| Coding specialist | DeepSeek-Coder-V2 (DeepSeek) | A leader for code generation, completion, and explanation |
| Efficient powerhouse | Mistral / Mixtral (Mistral AI) | Strong balance of performance and fast inference for interactive apps |
| Small and capable | Phi-3 (Microsoft) | Strong reasoning for its size; suited to on-device and modest hardware |
| Polished chat | Community fine-tunes | Often the most refined out-of-the-box chat experience for general Q&A |
How to choose
- Start with the task, not the model. Define the hardest thing your system must do reliably, then pick the smallest model that clears that bar.
- Cloud for capability, local for control. Reach for an API when you need frontier reasoning or long context; reach for local when data can’t leave your environment or volume makes per-token pricing untenable.
- Verify before you commit. Every price and ID here drifts. Confirm current figures in the provider’s own documentation before building a cost model.
