The mid tier of the GPT-6.1 line, released on October 26, four weeks after GPT-6.1 Sol. It is the first Terra model since GPT-5.6's, since the GPT-6 generation shipped only Sol and Luna. OpenAI's system card calls it "a fast, efficient model for high-volume professional work, with strong image analysis and computer use capabilities". The API docs pitch it at visual reasoning, computer use and writing. It has a 1.05M-token context (922K max input), 128K output, a June 1, 2026 knowledge cutoff, and text and image input. Reasoning effort runs from none to max, with medium as the default. Pricing is $1 / $5 per million input/output tokens, half GPT-5.6 Terra's $2 / $12 or less, with cached input at $0.05. Requests over 272K input tokens cost 2× input and 1.5× output. A Pro reasoning mode is served on OpenRouter as a separate listing. Parameters and architecture are undisclosed.

Independent measurement puts it close to Sol for a fraction of the price. Artificial Analysis scores it 51 on Intelligence Index v5.0 at max effort, four points below GPT-6.1 Sol (55) and 23 above GPT-6 Luna (28). The lower efforts read 45 (xhigh), 42 (high), 38 (medium) and 29 (low). In AA's launch analysis on v4.3.2 it scored 48 at max. That was six points above GPT-5.6 Terra at under a third of the cost per index task ($0.40 vs $1.40), and 7 to 13 points better at lower efforts, where AA found the low and medium settings Pareto-optimal on cost. AA also measured its AA-Omniscience hallucination rate falling from 88% to 38% against GPT-5.6 Terra, alongside a 2-point accuracy gain. Its Terminal-Bench 4.0 score rose by about 20 points, close to GPT-6.1 Sol at max effort. It scores 48 on AA's Coding Agent Index v2.0, against Sol's 54.

OpenAI treats it as Critical in cybersecurity under the Preparedness Framework, and High for biological and chemical capability, the same as GPT-6.1 Sol and GPT-6 Astra. It ships on their safeguards stack. On OpenAI's internal ExploitBench port it develops working exploits for 24.5% of recently disclosed vulnerabilities, ahead of GPT-6.1 Sol's 21.5% and behind GPT-6 Astra's 31.5%. It succeeds on 5 of 22 Sandbox Bench targets, and scores 83.33% pass@1 on SEC-Bench Pro. In a deployment simulation over 48,690 internal Codex tasks it drew 22 severity-3-or-higher flags (0.045%). That is the lowest of the GPT-6 line, against 24 for GPT-6.1 Sol, 26 for Astra and 38 for GPT-6 Sol. It scores 55.5 on HealthBench, up from GPT-5.6 Terra's 46.2. All figures in this paragraph are self-reported.

Model Details

Context window 1,050,000
AA Intelligence 51 was 48 on v4.3.2
License Proprietary

Benchmark Scores

Benchmark Score Mode
HealthBench 55.5 —
SEC-Bench Pro (pass@1) 83.33% —
ExploitBench (internal port, arbitrary code execution) 24.5% —

Variants

Name Parameters Notes
GPT-6.1 Terra (max) — AAII v5.0 51 (48 on v4.3.2 at launch); $1/$5 per 1M tokens, $0.05 cached input
GPT-6.1 Terra (xhigh / high / medium / low) — AAII v5.0 45 / 42 / 38 / 29 by reasoning effort
GPT-6.1 Terra Pro — the same model served with reasoning.mode set to pro (OpenRouter listing); same per-token price, far more tokens per task
reasoningcodingagenticcomputer-usemultimodalefficiencyproprietary

Related