ModelPort

GPT-5.5

Released: April 23, 2026 (API: April 24, 2026) Model ID: gpt-5.5 Pricing: $5 / $30 per million input / output tokens Cached input: $0.50 per million tokens Long-context (>270k input): 2x input, 1.5x output Status: legacy. Superseded by gpt-6.md. Still the required layer when upgrading from GPT-4o or o-series. Last verified: legacy, written at the GPT-5.5 launch (April to May 2026)

Variants

Variant	Model ID	Use case	Pricing (in/out)
GPT-5.5	gpt-5.5	Complex agentic tasks, coding, reasoning	$5 / $30
GPT-5.5 Pro	gpt-5.5-pro	High-stakes deep reasoning, max compute	$30 / $180
GPT-5.5 Instant	gpt-5.5-instant	Fast everyday tasks, low latency	TBD

Reasoning effort levels

GPT-5.5 supports a reasoning.effort parameter that controls internal chain-of-thought depth. Reasoning tokens are billed as output tokens — higher effort = higher cost.

Level	When to use	Trade-off
none	Simple lookups, classification, formatting	Fastest, cheapest, no reasoning overhead
low	Straightforward Q&A, information retrieval	Fast with minimal reasoning
medium (default)	Most professional workflows	Balanced speed / quality / cost
high	Complex multi-step problems, coding, analysis	Higher accuracy, more reasoning tokens
xhigh	Hardest problems — novel research, deep debugging, ambiguous goals	Maximum reasoning depth, highest cost and latency

Effort tuning best practices

• Start with medium — only escalate if output quality is insufficient.
• For batch/eval workloads, run a sample at medium vs high to quantify the quality delta before committing to higher cost.
• xhigh should be reserved for genuinely hard problems — it can 3–5x your output token cost.
• none is useful for structured-output-only calls where the model just needs to follow a schema, not reason about it.

Key features

• Natively omnimodal: Unified architecture for text, images, audio, and video — no separate endpoints.
• 1M token context window: Full million-token input capacity.
• 128k output tokens: Large output generation in a single pass.
• Responses API: Recommended over Chat Completions for agentic workloads. Built-in support for tool calls, structured outputs, and reasoning control.
• Structured outputs: 100% schema adherence via constrained decoding. Define schemas with JSON Schema or Pydantic — no post-processing needed.
• Tool use: Native web search, computer use (browser/desktop navigation), hosted shell, and multi-step tool orchestration.
• Prompt caching: Cached input at $0.50/M tokens (10x cheaper than uncached).
• Tool search: Built-in tool discovery and selection for complex workflows.

Changelog highlights

Date	Change
Apr 23, 2026	GPT-5.5 and GPT-5.5 Pro released
May 5, 2026	GPT-5.5 Instant released (default for ChatGPT free tier)
May 12, 2026	Older DALL·E snapshots and Realtime API Beta deprecated
May 28, 2026	GPT-5.5 Instant updated — improved style/readability; canvas dropped from Instant/Thinking

Migration from GPT-4o

This is a significant migration — not just a model ID swap.

Step 1: API surface

Area	Change required
Endpoint	Migrate to Responses API for best results. Chat Completions still works but is not optimized for reasoning-heavy or agentic workloads.
Model ID	Change to gpt-5.5
Reasoning	New reasoning.effort parameter (default medium). GPT-4o had no equivalent — if you were using o3 reasoning, effort levels map similarly.
Structured outputs	Now 100% schema-adherent via constrained decoding. Remove post-processing hacks for malformed JSON.
Tool schemas	Verify tool definitions — GPT-5.5 has stricter schema validation and better tool selection.

Step 2: Prompt strategy

Area	Guidance
Over-prompting	GPT-5.5 follows instructions more literally and is more capable. Remove verbose scaffolding, hand-holding, and workarounds from GPT-4o prompts.
Outcome-first	Focus system messages on goals, success criteria, and output shape — not step-by-step process. The model plans its own execution.
Fresh baseline	Don’t carry over legacy prompt hacks. Start with a clean prompt and add constraints only as needed.
Few-shot examples	Still effective for GPT-5.5 — use them for format/style calibration.
Chain-of-thought	For gpt-5.5, explicit CoT prompting (“think step-by-step”) still helps. For o3 models, avoid it — they handle reasoning internally.

Step 3: Operational changes

Area	Notes
Latency	Higher than GPT-4o, especially at high/xhigh effort. Add loading indicators for user-facing apps.
Cost	Output tokens are more expensive ($30 vs GPT-4o’s $15). Reasoning tokens add to output cost. Monitor spend during first week.
Output quality	Cleaner, more structured output. Update downstream parsers that expected “messy” GPT-4o responses.
Tokenizer	Different tokenizer — re-evaluate token counts and cost estimates.
Snapshot pinning	Use gpt-5.5 for latest, or pin to a dated snapshot for stability.

Step 4: Parallel testing

• Run GPT-4o and GPT-5.5 side-by-side for critical workloads.
• Log response differences systematically.
• Compare on: correctness, format adherence, latency, cost, tool-call accuracy.
• Only cut over after evidence supports it.

Migration from o3 / o3-mini

If migrating from o-series reasoning models:

Area	Notes
Model ID	gpt-5.5 or gpt-5.5-pro for hardest problems
Reasoning effort	Maps directly: low, medium, high work the same. xhigh is the new maximum.
Prompt style	GPT-5.5 benefits from more structured prompts than o3. Add explicit format/output constraints.
Tool use	GPT-5.5 has broader native tool support. Verify tool schemas still match.
Cost	GPT-5.5 is generally cheaper than o3-pro for equivalent quality on most tasks.

Quirks and best practices

• Reasoning tokens are hidden by default: Unlike o3’s visible reasoning summaries, GPT-5.5’s reasoning tokens are internal. Use the API’s reasoning config to control depth, but don’t expect to see the chain-of-thought unless you opt in.
• Responses API is the future: Chat Completions works but is a legacy surface. New features (tool search, computer use, hosted shell) are Responses API only.
• Structured output strictness: GPT-5.5’s constrained decoding means your schema must be valid. Loose/partial schemas that “worked” on GPT-4o may cause errors.
• Canvas deprecation: Canvas is no longer available in GPT-5.5 Instant or Thinking variants. If your workflow depended on canvas, use legacy models temporarily.
• Prompt caching ROI: At $0.50/M for cached input (vs $5/M uncached), caching is a 10x cost reduction. Cache system prompts and static context aggressively.
• Long-context pricing jump: Above ~270k input tokens, pricing doubles for input. Chunk large contexts or use summarization if cost-sensitive.
• Hybrid architecture: Consider using GPT-5.5 for execution and o3 for planning/judgment in complex workflows. They complement each other well.

Model ID reference

Version	Model ID	Status
GPT-5.5	gpt-5.5	Current (GA)
GPT-5.5 Pro	gpt-5.5-pro	Current (GA)
GPT-5.5 Instant	gpt-5.5-instant	Current (GA)
GPT-4o	gpt-4o	Active
o3	o3	Active
o3-mini	o3-mini	Active

This site is open source. Improve this page.