GPT-5.6 (OpenAI Reasoning Model Family)

Overview

OpenAI’s current-generation reasoning model family, documented in the official Reasoning Models guide. Reasoning models use internal reasoning tokens to “think” before producing a visible response — planning, using tools effectively, inspecting alternatives, and recovering from ambiguity. They work especially well for complex problem solving, coding, scientific reasoning, and multi-step agentic workflows, and are the recommended models for Codex CLI. OpenAI recommends starting with gpt-5.6 for most reasoning workloads.

Model Family (confidence: high)

Model	Role
gpt-5.6	Default choice for most reasoning workloads
gpt-5.6-sol	Highest-intelligence API option; for harder problems that tolerate more latency, used with reasoning.mode: pro
gpt-5.6-terra	Lower-cost tier
gpt-5.6-luna	Lowest cost and latency

Pricing and Capacity (confidence: high)

Model	Input $/MTok	Output $/MTok	Max output	Context	Knowledge cutoff	Reasoning
gpt-5.6-sol (alias gpt-5.6)	$5	$30	128K	1.05M	Feb 16, 2026	none→max
gpt-5.6-terra	$2	$12	128K	1.05M	Feb 16, 2026	none→max
gpt-5.6-luna	$0.20	$1.20	128K	1.05M	Feb 16, 2026	none→max

• All three: text+image input, text output, multilingual, vision; via Responses API + SDKs; tools = Functions, Web search, File search, Computer use.
• Specialized catalog (purpose-built, not reasoning-lineup): image (GPT Image 2), realtime speech (GPT-Realtime-2.1 / 2.1-mini / 2 / Translate / 1.5), TTS, transcription (GPT Transcribe, Live Transcribe, Realtime-Whisper, 4o / 4o-mini Transcribe).
• Selection framing: “Start with Sol for complex reasoning/coding; Terra to balance intelligence and cost; Luna for cost-sensitive, high-volume workloads.”

Reasoning models work better with the Responses API than Chat Completions; OpenAI reports improved model intelligence and performance through Responses. The family line is analyzed in detail in reasoning-effort.

Reasoning Effort and Mode (confidence: high)

• reasoning.effort — none / minimal / low / medium / high / xhigh / max, with model-dependent subsets and defaults (e.g. gpt-5.5 defaults to medium). Lower effort favors speed and token savings; higher effort produces more complete reasoning.
• reasoning.mode — standard (default) or pro on GPT-5.6 models. Mode and effort are independent: mode selects the execution tier, effort controls how much reasoning happens within it. Omitting effort defaults to medium in both modes.

Reasoning Token Mechanics (confidence: high)

• Reasoning tokens are not visible via the API but occupy context window space and bill as output tokens; the count is exposed as usage.output_tokens_details.reasoning_tokens.
• max_output_tokens caps reasoning + visible output + non-visible formatting tokens. Hitting it returns status: incomplete with incomplete_details.reason = "max_output_tokens" — possibly before any visible output exists, so you can pay for input + reasoning tokens without a visible answer.
• OpenAI recommends reserving at least 25,000 tokens for reasoning and outputs when experimenting; adjust as you learn your prompts’ reasoning-token appetite.
• Models before GPT-5.6 (e.g. gpt-5.5, gpt-5.4) support interleaved thinking: visible output before/between thinking, and thinking between tool calls.

Preserving Reasoning Across Calls (confidence: high)

• reasoning.context — auto (model default), current_turn, or all_turns. GPT-5.6 defaults to all_turns (renders compatible earlier reasoning into the next sample); earlier models default to current_turn.
• previous_response_id — the shortest stateful integration; the service ignores irrelevant reasoning items and keeps only what matters.
• Stateless mode (store: false or Zero Data Retention): reasoning items carry an encrypted_content property by default, passable to future calls.
• Persisted reasoning provides continuity, not exposure — reasoning items remain opaque; the API never returns their raw text.

Reasoning Summaries (confidence: medium)

The summary parameter exposes a summary of the model’s reasoning without revealing raw tokens. auto selects the most detailed summarizer available for the model (currently ~detailed for most; e.g. concise on the computer-use model). Summaries are opt-in and returned in the reasoning item’s summary array; organization verification may be required before using summarizers.

phase Parameter (confidence: high)

For long-running or tool-heavy GPT-5.5/5.4 flows, assistant messages should set phase: "commentary" (intermediate preambles, e.g. before tool calls) vs phase: "final_answer" (completed answer). Missing or dropped phase values can cause preambles to be treated as final answers; preserve each original value when replaying history.

Prompting Guidance (confidence: high)

• Give the model a clear goal, strong constraints, and an explicit output contract — without prescribing every intermediate step.
• Treat reasoning.effort as a tuning knob, not the primary way to recover quality.
• For agentic/research workflows, define what counts as done and how the model should verify its work.

Structured Outputs Support (confidence: high)

gpt-5.6 is the reference model for OpenAI’s Structured Outputs feature: text.format: { "type": "json_schema" } on the Responses API guarantees the response adheres to a supplied JSON Schema — no missing keys, no invalid enums — and refusals surface as an explicit, machine-detectable refusal/incomplete. For new type-safe integrations, use Structured Outputs rather than JSON mode. See structured-outputs.

Implications

Structured Outputs turns output-shaping into a schema contract the API enforces — for gpt-5.6-based integrations, type-safe parsing is a configuration choice, not a prompt skill. Refusals must still be handled, but as explicit first-class responses instead of parse failures.

Prompt Caching (confidence: high)

GPT-5.6 family models support automatic and explicit prompt caching to reduce cost and latency. Unlike earlier models, the default implicit breakpoint is placed at the latest message without fallback to shorter matching prefixes; developers must use explicit prompt_cache_breakpoint to reuse stable prefixes. Cache writes cost 1.25× the base input rate. See prompt-caching.

Model Selection (confidence: high)

The official Prompt Engineering guide frames model choice as the first prompt decision: reasoning models (gpt-5-6 family) think before answering and suit complex logic, coding, and multi-step tasks; non-reasoning GPT models answer directly and suit simple, low-latency work. Choose the smallest model that reliably does the job — reasoning tokens cost more and take longer. The guide’s GPT-5-series best practices add task-specific guidance for coding (role + workflow, tests, tool-use examples, Markdown standards), front-end (Tailwind/shadcn/Radix, Lucide/Material Symbols/Heroicons icons, Motion), and agentic workflows (planning, persistence, preambles, TODO rubrics).

Implications

Model choice and reasoning effort are now the two highest-leverage dials before any prompt text is written. Because reasoning models are costlier and slower, the marginal cost of verbose prompting rises — the guide’s best-practice sections exist precisely because “just add instructions” is no longer free.

Implications

The GPT-5.6 family shifts reasoning from a prompt technique (explicit chain-of-thought instructions) to inference-level control: the developer budgets and steers internal thinking via effort/mode/context parameters instead of coaxing it out of the model with instructions. That inverts classic CoT guidance — prescribing every step is now counterproductive. It also makes reasoning-token budgeting a first-class API concern (the incomplete/max_output_tokens failure mode) and introduces opaque-but-continuous reasoning state (all_turns, encrypted_content) that keeps reasoning useful across calls while never exposing it. Compared with the paired-thinking-response patterns this vault documents, OpenAI’s design hides reasoning both from the user and from the developer, exposing only summaries and token counts.

Open Questions

• Which subset of reasoning.effort values each model supports, and how minimal vs low differ in practice — the guide defers to per-model pages.
• Whether max effort materially beats xhigh on real evals; the guide only says to evaluate.
• How reasoning summaries evolve beyond the current auto ≈ detailed mapping.
• Pricing, context-window, and knowledge-cutoff values are snapshot values from the Models reference capture; verify against live model pages before quoting.

Related

• reasoning-effort — the effort/mode/context parameter taxonomy this family implements
• structured-outputs — JSON-Schema-guaranteed response format supported by gpt-5.6
• chatgpt — OpenAI’s earlier conversational model (GPT-3.5-turbo era)
• gpt-4 — OpenAI’s prior multimodal flagship
• prompt-engineering-guides — practical guide collection, including the reasoning-LLMs entry
• llm-models-guide — model-specific guidance catalog
• paired-thinking-response — user-facing hidden-reasoning patterns (contrast: OpenAI’s reasoning is developer-opaque too)

Sources

^[raw/prompts/articles/openai-reasoning-models-guide.md] ^[raw/prompts/articles/openai-structured-outputs-guide.md] ^[raw/prompts/articles/openai-prompt-engineering-guide.md]