OpenAI

GPT-5.5

Released Apr 23, 2026 · knowledge cutoff 2025-12

Status: Superseded
Location: United States
Type: LLM
Modality: Multimodal
Context window: 1M
Max output: 128K
Speed (medium effort): 64 tok/s · 10s to first answer token
Price ($/MTok in / out): $5 / $30
Cost per task (medium effort): $0.41
Open weights: No

Benchmarks7 of 10 reported

GPQA Diamond

93.5%

SWE-bench Verified

88.7%

SWE-bench Pro

58.6%

Terminal-Bench 2.1

83.4%

Humanity's Last Exam

41.4%

LMArena Elo

1481

ARC-AGI-2

85%

MMLU-Proretired

—

AIME (latest)retired

—

GDPval-AA v2

—

Strengths & weaknesses

Strengths

Tenacious on long agentic runs — keeps pushing rather than declaring premature victory
Reliably pulls the one relevant detail out of a very long prompt
Deliberately built for agent workloads rather than for chat

Weaknesses

Chat quality regressed — shorter, flatter answers and more refusals than the 5.1 line
Struggles to recover once a long thread already contains a wrong assumption
Training artefacts surfaced in production, including a tic for inserting creature words that OpenAI later patched out

GPT-5.5 Pro variant ($30/$180) sits #2 on LMArena. Context reported as 1M API / 1.1M by some trackers. Knowledge cutoff Dec 2025. SWE-bench Verified 88.7 lab-reported; LMArena Elo 1481 is the GPT-5.5 (high) listing. Terminal-Bench 2.1 83.4 is the vendor-submitted Codex CLI + GPT-5.5 entry on tbench.ai (also on CodingFleet's 2.1 board); the same model scores 76.4 through the Terminus 2 harness, and OpenAI's launch post reported Terminal-Bench 2.0 82.7. Speed is Artificial Analysis' GPT-5.5 (medium) model-page figure on OpenAI's API (2026-09-03); AA's providers page for the same variant showed a different window (OpenAI 75.0 tok/s / 6.41 s). SWE-Bench Pro 58.6% is OpenAI's launch figure (per llm-stats' OpenAI-sourced entry and Anthropic's Claude Fable 5 comparison table, which lists GPT-5.5 at 58.6); one DataCamp piece quotes 59.4%, untraced. OpenAI later audited SWE-Bench Pro (~30% broken tasks) and stopped recommending it.

For developers

API model strings

Not researched

Licence

Proprietary — weights not released

Retirement

No retirement announced

Lineage

No recorded predecessor