Models reference

Summary table

Model	Purpose	Cost coefficient	Context
gpt-5.4	Base GPT-5, balanced default	1.0×	256k
gpt-5.5	Extended, better for long-context tasks	1.4×	400k
gpt-5.4-mini	Cheap and fast, for classification	0.2×	128k
codex-5.3	Code-specialized for Codex CLI	1.1×	256k

The coefficient is applied to the base OpenAI token price and multiplied by the 1.09 margin.

GPT-5.4

The default general-purpose model. Good for chat, content generation, simple reasoning. Pick it when unsure — it offers the best price/quality trade-off.

{"model": "gpt-5.4", "messages": [...]}

GPT-5.5

Extended version with larger context and deeper reasoning. Use it for:

• long documents (>100k tokens),
• complex code analysis,
• multi-step reasoning,
• math and logic problems.

It costs 40% more, but pays off when gpt-5.4 underperforms.

GPT-5.4-mini

The cheapest model. Suitable for:

• classification, tagging, labeling,
• structured-data extraction from text,
• simple fallback flows,
• embedding-alternative use cases.

Do not use it for long-form generation or code — quality drops.

Codex-5.3

Specialized for Codex CLI and code generation. Supports the Responses API (POST /v1/responses) for interactive agent sessions. Pick it for:

• IDE code completion,
• patch/diff generation,
• agent loops with tool use.

codex --model codex-5.3 "Generate a REST API in FastAPI"

Tiers: Fast and Priority

You can choose a processing tier with the service_tier request parameter:

• default — standard queue.
• priority — higher priority, lower latency, +30% price.
• flex — batched processing, −40% price, latency up to several minutes.

{"model": "gpt-5.4", "service_tier": "priority", "messages": [...]}

Reasoning effort

GPT-5.x models accept a reasoning_effort parameter that controls internal reasoning depth:

Value	Output-token multiplier	When to use
minimal	1.0×	Plain answers, chat, classification
low	1.5×	Basic reasoning, typical tasks
medium	2.5×	Complex analysis, multi-step logic
high	4.0×	Math, proof checking, debugging

{
  "model": "gpt-5.4",
  "reasoning_effort": "medium",
  "messages": [{"role": "user", "content": "Solve this equation..."}]
}

Higher effort means more internal reasoning tokens spent — and a more expensive request. Start with minimal and increase only if the response quality is insufficient.

How to choose

1. General-purpose chat / agent → gpt-5.4 + minimal.
2. Long PDF / large-repo analysis → gpt-5.5 + medium.
3. Codex CLI / IDE assistant → codex-5.3.
4. Bulk classification → gpt-5.4-mini + minimal.
5. Production agent with tool use → codex-5.3 + low/medium.

See Billing for token-cost details.