GPT-5.6 Terra

Mid-tier

GPT-5.6 Terra is the balanced everyday model in the GPT-5.6 family, previewed June 26, 2026. OpenAI positions it as GPT-5.5-class performance at a fraction of the price. It launched at $2.50 per million input tokens and $15 per million output, and OpenAI cut it 20 percent on July 30 to $2 input, $0.20 cached, and $12 output, with a 1 million token context window and higher rates above 272K input tokens. Access at launch was limited to roughly 20 pre-approved organizations under a US Government preview arrangement; general availability followed on July 9, 2026. Detailed SWE-bench, MMLU-Pro, and GDPval results are not yet public; OpenAI disclosed coding, biology, and cybersecurity numbers in the preview post and said the rest land at GA.

Input Price

$2.00

per 1M tokens

Output Price

$12.00

per 1M tokens

Context Window

tokens

Released

2026-07

API access

Capabilities

textvisiontool-usecodereasoning

Key Strengths

✓$2/$12 per 1M since the July 30 cut, well under GPT-5.5 at $5/$30
✓1M token context window
✓Same lineage as Sol with cheaper economics
✓Reasoning, vision, tool use, and code included

Best For

▸Production agentic workflows where Sol is overkill
▸High-volume coding assistants
▸RAG over very long contexts
▸Migration target from GPT-5.5 once Terra reaches GA

Benchmark Scores

Benchmark	Score	Description
SWE-bench	95.4	Real-world software engineering tasks from GitHub issues (SWE-bench Verified)
MMLU-Pro	86.7	General knowledge and reasoning across 57 subjects
GPQA Diamond	92.9	Graduate-level science questions verified by domain experts
OSWorld 2.0	50.2	Computer use across real desktop applications and multi-step GUI tasks
BrowseComp	87.5	Agentic web search and browsing over hard-to-find facts
FrontierCode v1.1	41.3	Agentic coding on frontier software engineering tasks (Main split)
Terminal-Bench 4.0	26.3	Long-horizon agentic work in a terminal across software, science, ML, operations, hardware, security, and media (66 tasks, all-or-nothing verifiers)