Use the least expensive model that reliably passes your real test. Escalate to a frontier model only when the task is consequential, unusually difficult, or repeatedly fails on the cheaper tier.
OpenAI: Sol, Terra, and Luna
| Model | Best fit | Watch for |
|---|---|---|
| GPT-5.6 Sol | The hardest reasoning, coding, research, and high-value professional work. | Higher cost and latency are wasted on routine rewriting or classification. |
| GPT-5.6 Terra | Balanced everyday work: analysis, writing, planning, documents, and general business use. | Still verify facts and calculations; balanced does not mean infallible. |
| GPT-5.6 Luna | Fast, lower-cost work such as extraction, classification, simple drafting, and high-volume workflows. | Escalate when nuance, ambiguity, or multi-step reasoning matters. |
Inside ChatGPT or an API tool, model availability, context, usage limits, and reasoning controls can differ. Read the exact product label rather than assuming every GPT-5.6 experience is the same.
Claude: Fable, Opus, Sonnet, and Haiku
| Model | Best fit | Practical positioning |
|---|---|---|
| Claude Fable 5 | Frontier work where quality matters more than price: difficult analysis, complex coding, long-running agents, and high-stakes professional drafts. | Anthropic's flagship public tier. Do not spend Fable tokens on simple summaries. |
| Claude Opus 4.8 | Deep reasoning, coding, agentic work, and tasks that benefit from a powerful mature model. | A high-capability option that may remain available beside newer frontier releases. |
| Claude Sonnet 5 | The balanced workhorse for coding, documents, research, and business workflows. | Often the sensible default when you need strong quality at lower cost than the flagship. |
| Claude Haiku 4.5 | Fast, economical classification, extraction, routing, customer-support drafts, and repetitive sub-tasks. | Useful inside agents when many inexpensive calls matter more than maximum depth. |
It means Sonnet is being used with a medium effort or reasoning setting. The model name and the effort setting answer different questions: which model? and how much compute should it spend?
How effort settings change usage
Low or medium effort can be appropriate for routine coding, drafting, and analysis. Higher effort gives the model more room to reason, but can increase latency and token consumption. Test the same representative tasks at two settings before deciding that “high” should be your default.
Gemini: speed, general work, and frontier reasoning
Google's model names and previews change quickly, but the durable pattern is similar: a fast economical tier for volume, a balanced general-purpose tier, and a Pro or frontier tier for difficult multimodal reasoning, coding, and very large contexts. Gemini is especially useful when the work combines text with images, audio, video, code, or large document collections.
Do not confuse a large context window with perfect comprehension. Organize long inputs, identify the relevant evidence, and request citations, page references, or timestamps.
Model name versus product mode
A model is the underlying system. A product may add search, files, memory, connectors, code execution, agents, or a selectable reasoning level. Two people can say they used “Claude” or “ChatGPT” while actually using different models, tools, effort settings, and context.
When documenting a workflow, record all four:
- Model: Sol, Terra, Luna, Fable, Sonnet, Haiku, or the exact Gemini model.
- Effort or reasoning setting: low, medium, high, extended, or automatic where offered.
- Tools: web, files, code, email, calendar, database, or computer use.
- Evaluation: the test cases and standard the output must pass.
A simple selection ladder
- Start fast and inexpensive. Use Luna, Haiku, or the equivalent Gemini fast tier for bounded routine work.
- Move to the balanced model. Use Terra, Sonnet, or the equivalent general tier when the task needs stronger judgment or richer output.
- Escalate only on evidence. Use Sol, Fable, Opus, or a Gemini frontier model when cheaper tiers fail or the value of a better answer justifies the cost.
- Measure the whole workflow. Include retries, human review, tool calls, and agent handoffs—not merely the first prompt.
Frequently asked questions
Should I always use the newest flagship?
No. Flagships are appropriate when additional capability creates additional value. For routine extraction, rewriting, tagging, and routing, a smaller model is usually faster and cheaper.
Is medium effort the same as a medium-sized model?
No. A model tier and an effort setting are separate controls. Sonnet can be the selected model while “medium” controls how intensively it reasons.
Which model is best for a small business?
There is no universal winner. Begin with a balanced model for important work and a fast model for repetitive tasks. Keep the model that produces the best net result after cost, errors, and review time.
Which model should run an agent team?
Do not automatically give every role the flagship. A strong coordinator can receive concise work from cheaper specialist models. Evaluate each role separately.