# General text, chat, and reasoning model guide

General text models cover Q&A, writing, extraction, and analysis, while reasoning models fit multi-step problems; production selection should balance task accuracy, latency, cost, and control.

## Separate task depth first

Summarization, rewriting, classification, and simple Q&A usually prioritize speed, price, and stable formatting. Mathematics, planning, complex analysis, and constrained decisions depend more on reasoning. Do not send every request to the most expensive model; route by difficulty and keep a fallback.

## Selection dimensions

Compare accuracy, instruction following, structured-output success, context, time to first token, and total cost on real inputs. Work involving the web, files, or business systems also requires tool reliability, permission boundaries, and recovery; model knowledge does not replace a current data source.

## Minimal call

This site uses a Bearer token. Select the model name from the marketplace, set timeouts, and distinguish authentication, rate limit, context, and upstream failures.

First copy an exact model ID that explicitly supports `/v1/chat/completions` from the [model marketplace](https://api.tu-zi.com/pricing):

```bash
export BASE_URL="https://api.tu-zi.com"
export API_KEY="your site API key"
export CHAT_MODEL_NAME="exact model ID that supports /v1/chat/completions"
```

```bash
curl "$BASE_URL/v1/chat/completions" \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -d @- <<JSON
{"model":"$CHAT_MODEL_NAME","messages":[{"role":"user","content":"Analyze this problem"}]}
JSON
```

## Quality evaluation

Build a fixed sample from real work and record factual accuracy, schema pass rate, human editing, refusal quality, latency, and cost. Use traceable sources or human rubrics for open answers and deterministic validation for structured tasks; model self-evaluation is not acceptance evidence.

## Common limits

A context window is not reliable memory. Long input increases latency and can dilute important evidence. Temperature, reasoning effort, JSON Schema, tools, and caching are model-specific; run field-level contract tests against the exact model and channel before launch.

## Use cases

- Knowledge Q&A and content generation
- Complex analysis and multi-step reasoning
- Structured extraction and classification

## API paths

- /v1/responses
- /v1/chat/completions
- /v1/messages

## Related guides

- [OpenAI GPT and o-series](https://api.tu-zi.com/en/docs/models/openai-gpt): As reviewed on 2026-08-08, OpenAI positions GPT-5.6 Sol, Terra, and Luna for complex professional work, balanced intelligence and cost, and high-volume workloads through the Responses API.
- [Anthropic Claude](https://api.tu-zi.com/en/docs/models/claude): As reviewed on 2026-08-08, Anthropic positions Fable 5, Opus 5, Sonnet 5, and Haiku 4.5 for highest capability, complex agents, balanced work, and low latency.
- [Google Gemini](https://api.tu-zi.com/en/docs/models/gemini): Gemini is Google's multimodal family; the official catalog lists separate models and methods for general reasoning, low latency, images, video, audio, embeddings, and agents.
- [DeepSeek](https://api.tu-zi.com/en/docs/models/deepseek): The current official IDs are deepseek-v4-flash and deepseek-v4-pro; both enable thinking by default and can disable it. This gateway supports only names and paths explicitly enabled in its model list and protocol guide.

## Frequently asked questions

### Should I choose Responses, Chat Completions, or Messages?

Start with the endpoints actually supported by the target model and channel. Existing messages/choices clients can keep Chat, new workflows can evaluate Responses, and native Claude structures commonly use Messages. Tools, stream events, and response fields differ, so migration is not a URL-only change.

### Is a reasoning model always better than a general text model?

No. Multi-step math, code, and dense constraints are more likely to benefit from reasoning, while summarization, classification, and simple extraction often prioritize low latency and stable formatting. Reasoning can add wait time and output tokens, so route by task difficulty and set budgets.

### Can I compare two text models using a public benchmark alone?

No. Use real business samples and record factual accuracy, schema pass rate, human editing, refusal quality, first-token and total latency, and actual cost. Validate structured tasks programmatically and open answers with sources or a human rubric.

### Does a large context window mean the model reliably remembers everything?

No. Long inputs raise cost and latency and can dilute key evidence. Send only what the task needs, redact sensitive data, and review service retention terms. Critical facts still need retrieval, citations, or deterministic checks.

## Sources

- OpenAI Models: https://developers.openai.com/api/docs/models
- Claude Models: https://platform.claude.com/docs/en/about-claude/models/overview
- Gemini Models: https://ai.google.dev/gemini-api/docs/models

## Documentation source

Source: https://api.tu-zi.com/en/docs/use-cases/text-reasoning

Reviewed: 2026-08-12

This document is an integration reference. Provider-native APIs may differ from this site's APIs; use the corresponding API Reference and model marketplace to confirm the site URL, authentication, model ID, group and parameters. This document does not certify live success for every model or upstream service. It contains documentation and examples, without the current user's credentials or request history.