OpenAI

GPT-5.6 Terra

Released Jul 9, 2026 · knowledge cutoff unpublished

Status: Superseded
Location: United States
Type: LLM
Modality: Multimodal
Context window: 1M
Max output: 128K
Speed (medium effort): 86 tok/s · 1.65s to first answer token
Price ($/MTok in / out): $2 / $12
Cost per task (medium effort): $0.12
Open weights: No

Benchmarks3 of 10 reported

SWE-bench Pro

63.4%

Terminal-Bench 2.1

87.4%

ARC-AGI-2

83.9%

MMLU-Proretired

—

GPQA Diamond

—

SWE-bench Verified

—

AIME (latest)retired

—

Humanity's Last Exam

—

LMArena Elo

—

GDPval-AA v2

—

Strengths & weaknesses

Strengths

Behaves like the flagship on ordinary work — the gap only opens on the hardest long-horizon agentic tasks
Long-context recall holds up close to Sol's, where the budget Luna tier falls away sharply
Full feature set rather than a cut-down one: image input, tool calling, extended reasoning, prompt caching and structured outputs

Weaknesses

OpenAI published no per-tier academic benchmarks, so where it really sits against rivals is hard to pin down
Trails Sol most on exactly the long-horizon agentic work that justifies reaching for a frontier model

Middle tier of the three-model GPT-5.6 family. OpenAI published only agentic suites per tier — Terminal-Bench 2.1 87.4%, Coding Agent Index 77.4, Agents' Last Exam 50.4, MRCR long-context recall 89.6% — and none of the site's tracked academic benchmarks (GPQA Diamond, HLE, MMLU-Pro, AIME stay null); SWE-Bench Pro and ARC-AGI-2 were filled later from OpenAI's launch table and ARC Prize respectively (below). Artificial Analysis Intelligence Index 55 (vs Sol 59, Luna 51); speed is AA's medium-effort measurement, matching the costPerTask effort (an earlier 140.4 tok/s reading in these notes was a different listing). Pricing is OpenAI's list rate, cut from $2.50/$15.00 to $2.00/$12.00 on 2026-07-30 (Luna was cut 80% in the same move). Context window per Artificial Analysis and OpenRouter; max output is tracker-reported, not from OpenAI's docs. Knowledge cutoff not published per tier. ARC-AGI-2 83.9% is ARC Prize-verified at max effort (xhigh 74.2%, high 67.1%, medium 37.5%, low 18.8%), published 2026-07-09 at arcprize.org/results/openai-gpt-5-6-terra; cost per task not stated. SWE-Bench Pro 63.4% is OpenAI's launch-table figure (Sol 64.6 / Terra 63.4 / Luna 62.7), as reproduced by LayerLens, DataCamp and llm-stats.

For developers

API model strings

openaigpt-5.6-terra

Licence

Proprietary — weights not released

Retirement

No retirement announced

Lineage

No recorded predecessor