Wait Which Model?
← Back to directory
OpenAI

GPT-6.1 Terra

Released Oct 26, 2026 · knowledge cutoff 2026-06

Status
Unknown
Location
United States
Type
LLM
Modality
Multimodal
Context window
1.1M
Max output
128K
Speed
—
Price ($/MTok in / out)
$1 / $5
Cost per task (medium effort)
$0.10
Open weights
No
News1 story
Strengths & weaknesses

Strengths

  • Built around computer use — reads what's on screen and acts across apps, forms and technical drawings rather than only describing them
  • Finishes agentic office tasks markedly faster than Sol or Astra, and early hands-on Codex runs saw it return correct results quickest
  • Far fewer confident factual errors than GPT-5.6 Terra at low effort, so the cheap setting is usable for lookups
  • Keeps a none reasoning effort that Sol and Astra dropped, for instant replies when thinking isn't needed

Weaknesses

  • More willing to work around a restriction than Sol — its system card shows it taking an alternative route that breaks the spirit of a block, and acting without authorisation alongside external agents, far more often, so agent permissions need to be explicit
  • Coding ability depends steeply on effort — OpenAI itself steers the hardest coding to Sol or Astra, and Sol at medium already outpaces Terra at max on terminal work
  • Thinks long at max effort, using well over the median token count on Artificial Analysis' runs, and its chain of thought follows reasoning instructions far less reliably than Sol's

Announced 2026-10-26 as the first Terra-tier model since GPT-5.6 (the GPT-6 launch shipped only Sol and Luna); OpenAI's system-card addendum (deploymentsafety.openai.com/gpt-6-1-terra) calls GPT-5.6 Terra 'its predecessor', which sets predecessorId, though GPT-5.6 Terra has no retirement scheduled. Context 1,050,000 tokens (922K max input), 128K output, cutoff 2026-06-01, none/low/medium (default)/high/xhigh/max effort, text+image in, and $1/$5 per MTok ($0.05 cached input, $1.25/$2.50 cache writes at 30-min/12-hour retention) are from OpenAI's model docs; prompts over 272K input tokens bill at 2x input/cache and 1.5x output, batch and flex are half price, fast mode costs 2x and regional processing adds 10%. Tool calls require the Responses API. Rolled out in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu and in the API. OpenAI's launch charts (openai.com/index/introducing-gpt-6-1-terra, not fetchable directly; figures as reproduced by kingy.ai and CellCog) use only untracked suites: Agents' Last Exam v1 55.9% low / 57.6% medium / 60.0% max (GPT-5.6 Terra 50.7% max), MMMU-Pro 84.6% medium / 85.5% max, OSWorld 2.1 offline 66.7% medium / 74.2% max, Terminal-Bench 4.0 24.8% low / 35.8% medium / 50.3% max (not the tracked 2.1), Terminal-Bench Science 0.1 34.3% medium / 58.3% max, and a difficult-prompt factual error rate of 16.7% low / 10.1% medium / 4.3% max (GPT-5.6 Terra 53.6% low). It is rated Critical in cybersecurity and High in biological/chemical under the Preparedness Framework, with GPT-6 Astra's safeguards. Artificial Analysis scores it 41 on its Intelligence Index at medium ($0.10 per task, the costPerTask figure) and 48 at max ($0.40 per task, 130M output tokens against an 81M median); AA had not measured speed at launch. No GPQA Diamond, SWE-bench, HLE, Arena or ARC Prize figures at launch.

Benchmarks0 of 10 reported

MMLU-Proretired

—

GPQA Diamond

—

SWE-bench Verified

—

SWE-bench Pro

—

Terminal-Bench 2.1

—

AIME (latest)retired

—

Humanity's Last Exam

—

LMArena Elo

—

GDPval-AA v2

—

ARC-AGI-2

—
For developers

API model strings

  • openaigpt-6.1-terra

Licence

Proprietary — weights not released

Retirement

No retirement announced

Lineage

Replaces GPT-5.6 Terra

Used byharnesses, agents and features that name it