GPT-5.5¶

Development line: project:gpt-5-5 · thread gpt-5-5-announcement
Last event: 2026-04-23 · 1 dated since 2026-04-23 · Researched: 2026-09-06 · confidence: high

What it is¶

GPT-5.5 — an OpenAI API model for teams building coding, tool-using, and professional-work agents.

• Supports Responses and Chat Completions, function calling, structured outputs, hosted tools, computer use, and MCP.
• Offers the pinned gpt-5.5-2026-04-23 snapshot.

1,050,000-token context; 128,000-token maximum output; API Free is unsupported. Retain it for a pinned or evaluated workload, but use GPT-6 Astra as the new-workload baseline.

Development line¶

• 2026-04-23 — GPT-5.5 announcement. On 2026-04-23, an OpenAI announcement marked a public milestone for GPT-5.5. The announcement confirms a dated public step for the project, but does not establish capabilities, availability, pricing, or migration details.

What changed¶

• 2026-04-23: GPT-5.5 began rolling out in ChatGPT and Codex for eligible plans, aimed at coding, computer use, knowledge work, and research.
• 2026-04-24: GPT-5.5 and GPT-5.5 Pro became available in the API; API-deployment safeguards were added.
• 2026-05-05: GPT-5.5 Instant replaced GPT-5.3 Instant as ChatGPT’s default and appeared in the API as chat-latest.
• 2026-05-07: GPT-5.5-Cyber entered limited preview for defenders securing critical infrastructure.
• 2026-05-28: GPT-5.5 Instant received a ChatGPT/API style update; Canvas was removed for Instant and Thinking.
• 2026-06-09: Instant personalization expanded to ChatGPT Go and Free, with Free using a reduced past-chat set.
• 2026-06-24: ChatGPT updated Instant for multi-turn context, complex constraints, and decision-oriented conversations.
• 2026-07-06: GPT-5.5 Instant Mini replaced GPT-5.3 Instant Mini as the ChatGPT fallback after Instant or Auto rate limits.

How to use this¶

1. Choose gpt-5.5 for the maintained alias, or gpt-5.5-2026-04-23 when behavior must stay pinned; confirm the API usage tier first. — https://developers.openai.com/api/docs/models/gpt-5.5
2. Use the Responses API for reasoning, tool-calling, or multi-turn work. — https://developers.openai.com/api/docs/guides/latest-model?model=gpt-5.5
3. Start with reasoning.effort: medium. Evaluate low for latency-sensitive work and raise effort only after representative evaluation. — https://developers.openai.com/api/docs/guides/latest-model?model=gpt-5.5
4. Use Structured Outputs, previous_response_id for response state, and a stable prompt prefix for cached repeated traffic. — https://developers.openai.com/api/docs/guides/latest-model?model=gpt-5.5
5. For a new system, establish the accuracy target with GPT-6 Astra, then compare cost and latency before retaining GPT-5.5. — https://developers.openai.com/api/docs/guides/model-selection

Best practices¶

• Treat GPT-5.5 as a new model family. Begin with a fresh, minimal prompt baseline instead of carrying forward an older prompt stack. — https://developers.openai.com/api/docs/guides/latest-model?model=gpt-5.5
• State outcome, success criteria, constraints, evidence rules, output shape, and stop conditions. Avoid procedural instructions unless the path itself is required. — https://developers.openai.com/api/docs/guides/latest-model?model=gpt-5.5
• Use Structured Outputs rather than placing a JSON schema in prompt prose. — https://developers.openai.com/api/docs/guides/latest-model?model=gpt-5.5
• Put stable prompt content first and dynamic context last. Use prompt_cache_key consistently and inspect cached-token usage. — https://developers.openai.com/api/docs/guides/latest-model?model=gpt-5.5
• Put tool inputs, side effects, retry safety, and failure modes in tool descriptions. Preserve phase exactly if assistant items are replayed manually. — https://developers.openai.com/api/docs/guides/latest-model?model=gpt-5.5

Superseded by this¶

• :{

Still unknown¶

• Availability, rate limits, and the ChatGPT model picker depend on the account, plan, region, and API tier. We checked no target account.
• GPT-5.5, GPT-5.5 Pro, Thinking, Instant, Cyber, and Instant Mini are distinct variants or access modes. Do not transfer access, safety, or capability claims between them.
• The reviewed Simplified-Chinese launch page established no separate China-specific rollout or regional API entitlement.
• Checked sources contain no official retirement date for the gpt-5.5 API alias.

Sources¶

Agent brief¶

• Subject: project:gpt-5-5, thread gpt-5-5-announcement, 1 dated events 2026-04-23 → 2026-04-23.
• Practical note: See the sourced usage and practice sections above, including their limits.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.