Key Features

Read the Model Signal report →

GPT-5.5

Identity

GPT-5.5 is a flagship reasoning-class artificial intelligence model developed by OpenAI, released on April 23, 2026 (Source 1). It serves as the successor to the GPT-5.4 lineage, maintaining per-token latency parity with its predecessor while increasing intelligence and autonomous capabilities (Source 5, 6). The model is deployed in several variants, including GPT-5.5 Standard, GPT-5.5 Pro, and GPT-5.5 Instant, the latter of which replaced previous iterations as the default model for ChatGPT (Source 1, 17, 19). In June 2026, OpenAI acquired the cloud platform Ona to provide managed runtimes for the model's agentic workflows (Source 4). A specialized version, GPT-5.5-Cyber, is restricted to "critical cyber defenders" and state-level infrastructure protection (Source 13).

What it is

According to the provider, GPT-5.5 is designed for advanced reasoning and agentic workflows (Source 1).

• Provider: OpenAI
• Release Date: April 23, 2026
• Category: Reasoning
• Context Window: 1M tokens
• API Pricing: $5 per 1M input tokens / $30 per 1M output tokens (Standard variant)
• Key Features: Codex integration for engineering tasks; managed agentic execution via Ona; native multimodality (Source 1, 4, 12).

The model is optimized for "long-horizon" autonomous tasks, though its performance in this area is contested by specialized competitors (Source 8). The "Instant" variant is tuned for lower hallucinations and concise responses, specifically targeting high-stakes domains like medicine, law, and finance (Source 18, 21).

Capabilities & benchmarks

GPT-5.5 demonstrates high-tier performance across reasoning and agentic benchmarks, though it does not lead in every category.

• General Intelligence: The model scored 55 on the Artificial Analysis Intelligence Index, placing it third behind Fable 5 and Opus 4.8 (Source 7).
• Agentic Reasoning: GPT-5.5 won the "Agents’ Last Exam" (ALE), a benchmark designed by UC Berkeley RDI and 300 experts to measure real-world agentic competence, defeating Claude Fable 5 (Source 10).
• Factuality: OpenAI claims GPT-5.5 Instant produces 52.5% fewer hallucinated claims on high-stakes prompts compared to previous versions (Source 18).
• Cybersecurity: In testing by the AI Security Institute (AISI), GPT-5.5 solved one of two cyber ranges measuring offensive attack capabilities (Source 16).
• Precision: The model is reported to trail DeepSeek V4 Pro in specific precision-based tasks (Source 11).
• Coding: While integrated with Codex, GPT-5.5 is outperformed on long-horizon coding benchmarks by GLM-5.2 (Source 8).

How it compares

• Vs GPT-5.4: GPT-5.5 matches the per-token latency of GPT-5.4 while providing a significant step up in intelligence and autonomy (Source 6).
• Vs Claude Fable 5: GPT-5.5 (55) trails Fable 5 (60) on the Artificial Analysis Intelligence Index (Source 7), but GPT-5.5 outperformed Fable 5 on the agent-specific ALE benchmark (Source 10).
• Vs Opus 4.8: GPT-5.5 (55) sits slightly behind Opus 4.8 (56) on general intelligence indexing (Source 7).
• Vs GLM-5.2: Z.ai’s open-weights GLM-5.2 (51) trails GPT-5.5 on general intelligence but beats it on multiple long-horizon coding benchmarks at 1/6th the cost (Source 7, 8).
• Vs Mythos: Mythos Preview completed both AISI cyber ranges, whereas GPT-5.5 completed only one (Source 16).
• Vs MiniMax-M3: MiniMax-M3 is reported to eclipse GPT-5.5 on key benchmark performance while costing 5-10% as much (Source 12).
• Vs DeepSeek V4 Pro: DeepSeek V4 Pro reportedly outperforms GPT-5.5 Pro on precision-specific metrics (Source 11).
• Vs Gemini: Google's Gemini I/O 2026 model is positioned in the same class as GPT-5.5, though reportedly short of Mythos (Source 15).

Where it fits

GPT-5.5 is the primary engine for OpenAI's consumer and enterprise ecosystem. It serves as the default model for ChatGPT, replacing older versions to provide near-frontier capabilities to free-tier users (Source 19, 20). Its integration with the Ona platform positions it as a "managed execution" environment for autonomous agents, moving beyond simple chat interfaces (Source 4). In the public sector, GPT-5.5-Cyber is utilized by Japan’s megabanks (MUFG, SMBC, and Mizuho) as a defensive shield against AI-driven cyberattacks (Source 13).

Open Questions

• Cost Translation: Users have reported difficulty translating per-million-token pricing into predictable operational spend for complex workflows (Source 9).
• Autonomy vs. Control: While GPT-5.5 is marketed for agentic workflows, the degree of human-in-the-loop verification required for its "autonomous" engineering tasks remains a point of debate (Source 6, 8).
• Regional Competition: The emergence of Chinese models like MiniMax-M3 and GLM-5.2 offering similar or superior performance at a fraction of the cost raises questions about the long-term pricing power of the GPT-5.5 lineage (Source 8, 12).

Contradictions

• Default Model Replacement: Source 19 states GPT-5.5 Instant replaced GPT-5.3 Instant as the default ChatGPT model, while Source 20 claims it replaced GPT-3.5 Instant.

Sources

• source 1: sn_model_face:29628a5e-26af-463f-a351-e86c3f7825a7 (GPT-5.5 specs)
• source 2: model_provider_url:https://openai.com/index/introducing-gpt-5-5/
• source 3: sn_article:6454a6dc-6829-4fdf-b438-dc26d278dbc8 (Daily Signal June 17, 2026)
• source 4: sn_article:b04fabba-387a-4b3f-8270-ac35eb1da177 (Weekly Signal June 16, 2026 - Ona acquisition)
• source 5: sn_article:fc984399-ff1c-4ca1-8f46-0c8bd4bc35c7 (Daily Signal May 8, 2026)
• source 6: sn_article:0b910fd2-3813-4a94-9053-03a592785098 (Daily Signal April 24, 2026 - Latency parity)
• source 7: sn_wire_item:21e43baf-a1f7-4e00-bbce-34d607deeb3f (Techmeme - Intelligence Index)
• source 8: sn_wire_item:d8661cfc-e804-4b17-b585-621f23d64249 (VentureBeat - GLM-5.2 vs GPT-5.5)
• source 9: sn_wire_item:88edbb42-634f-47bf-8729-94754391505f (Gizmodo AI - Token cost complexity)
• source 10: sn_wire_item:a576c79e-8127-4a80-a386-0b8c7c2c7837 (VentureBeat - Agents' Last Exam)
• source 11: sn_wire_item:57bd771f-9b49-4d2e-86cb-2f4ecb251680 (Hacker News - DeepSeek precision)
• source 12: sn_wire_item:dc0410df-5c3b-484a-9e61-34d879ad5955 (VentureBeat - MiniMax-M3)
• source 13: sn_wire_item:17a79bbc-6792-4264-9de3-4e2a0a90f1a3 (The Next Web - Japan Cyber Defense)
• source 14: sn_wire_item:972271a7-fd9b-487b-8911-f1e2a674b10b (Business Insider - Cyber dash)
• source 15: sn_wire_item:14d8036b-48b7-454e-aa79-8b9398459972 (Techmeme - Gemini I/O)
• source 16: sn_wire_item:561be460-0df4-4b54-ae0f-370c06c3926c (Techmeme - AISI Cyber Ranges)
• source 17: sn_wire_item:efb53993-02e2-4076-a073-8d524c9fb293 (TechRadar Pro - GPT-5.5 Instant UX)
• source 18: sn_wire_item:23efd790-b2e5-4cd0-bafd-3c2ae8f8cb05 (Techmeme - Hallucination stats)
• source 19: sn_wire_item:568bce5b-23b1-4321-8fcf-b68cf8848f14 (Techmeme - GPT-5.3 replacement)
• source 20: sn_wire_item:d720a426-fa2d-45fd-9f3e-73c937fe1407 (TechCrunch AI - GPT-3.5 replacement)
• source 21: sn_wire_item:92d254ea-2548-44a9-9194-f04bedf4fe0a (The Verge AI - Factuality claims)