GPT-6.1 Terra
GPT-6.1 Terra is OpenAI's fast, low-cost mid-tier GPT-6.1 model (released 26 Oct 2026) built for high-volume agentic, computer-use and document/visual work at $1/$5 per 1M tokens with a 1M-token context. Externally verified scores remain sparse for this exact variant: AA Intelligence Index 48.1 (Benchmark Heaven ground truth) and OpenAI's own Terminal-Bench 4.0 50.3% at max reasoning effort, measured at ~93 tps.
Empirical Evaluation Results
Architectural Profile & Capabilities
GPT-6.1 Terra is a mid-tier member of OpenAI's GPT-6 series, published 26 October 2026 alongside an addendum to the GPT-6 Astra system card rather than a standalone system card. OpenAI positions it between GPT-6.1 Sol and the smaller GPT-6 Luna line as a 'fast, efficient model' for high-volume professional work, emphasising three things: roughly 3x the throughput of the flagship tier, standard API pricing at half the rate of its GPT-5.6 Terra predecessor, and improved accuracy/coding over that predecessor (OpenAI reports its error rate on its own error-inducing prompt suite falling from 53.6% to 16.7% at low effort). Commercial shape confirmed across multiple independent listings: $1.00 per 1M input tokens and $5.00 per 1M output tokens, with cached input commonly quoted at $0.05 per 1M (cache-write tiers of $1.25/$2.50 per 1M at 30-minute/12-hour retention appear in third-party listings only and are not in an official price sheet). Context is 1,000,000 tokens per OpenRouter-model telemetry, with third-party listings quoting 1.05M-1.1M and a 128K max-output figure that has no primary-source confirmation. Capability evidence is genuinely thin for this exact variant. OpenAI's launch material reports Terminal-Bench 4.0 at 50.3% at max reasoning effort (more than double GPT-5.6 Terra's 21.5%), states that at only medium effort Terra scores 15 percentage points above GPT-5.6 Terra on Agents' Last Exam, and states that OSWorld 2.1 at medium effort exceeds GPT-5.6 Terra's best score - but the latter two are disclosed only as deltas/qualitative claims, with no absolute Terra number published. Benchmark Heaven's independently tracked AA Intelligence Index of 48.1 at max effort is the most defensible third-party anchor. SciCode 55.1% appears in an aggregated third-party listing for this slug. No exact-variant figures are published for GPQA Diamond, HLE (with or without tools), SWE-bench Verified, SWE-bench Pro, FrontierMath, Terminal-Bench-Science, BrowseComp, ARC-AGI-2, MRCR v2, BigLaw, CaseLaw, MedQA, FinQA or SecBench - and per ARMES governance those must not be inherited from GPT-6.1 Sol, GPT-6 Astra, GPT-6 Sol/Luna or GPT-5.6 Terra. Operationally Terra behaves like a throughput-optimised reasoning model rather than a frontier reasoning model: it exposes explicit low/medium/high/xhigh/max reasoning-effort tiers (AA tracks a 'non-reasoning' configuration separately), and its economics only make sense when work is batched at medium effort. The Astra system-card addendum assigns it a Critical cyber-risk rating and reports 22 severity-3-or-higher flags across 48,690 simulated internal Codex tasks - slightly fewer than GPT-6.1 Sol (24) and GPT-6 Astra (26), but still material for autonomous code execution.
OpenAI publishes no parameter count, active-parameter count or sparsity ratio for GPT-6.1 Terra, and no source describes MoE routing; with no evidence of sparse expert routing it is best classified as a dense autoregressive decoder-only Transformer in the GPT-6 lineage, trained as a throughput-optimised sibling of GPT-6.1 Sol with a documented evaluation addendum in the GPT-6 Astra system card. Attention/tokenizer internals (MLA vs GQA, layer/head counts) are not disclosed. Context window is 1,000,000 tokens (third-party listings quote 1.05M-1.1M; a 128K max-output figure is unverified). Modality per OpenRouter metadata is text output with image/file input available on vision-capable endpoints; the launch blog markets visual reasoning and precise computer use, so pixel-grounded work should be routed to the vision endpoint. Test-time compute is exposed as explicit reasoning-effort tiers (low, medium, high, xhigh, max) with hidden reasoning tokens billed at the output rate - AA also tracks a separate 'non-reasoning' (GPT-6.1 Terra Non-reasoning) configuration, which must not be conflated with the ::max variant. Native tool calling and computer-use post-training are core to the variant's design, and prompt caching plus 1M-token retrieval are the primary cost levers.
Recommended Workloads & Primary Use Cases
- •Exact-variant benchmark coverage is thin and lagging: only AA Intelligence Index 48.1 (Benchmark Heaven), OpenAI's vendor-reported Terminal-Bench 4.0 50.3% at max effort, and a third-party SciCode 55.1% are published for openai/gpt-6.1-terra. There are no exact-variant GPQA Diamond, HLE (tools/no-tools), SWE-bench Verified/Pro, BrowseComp or MRCR v2 numbers - any Sol/Astra/Luna/predcessor figure substituted for Terra is invalid.
- •Reasoning-effort tier drives both score and error rate: the 50.3% Terminal-Bench 4.0 figure is a max-effort result, while OpenAI's own low-effort error rate is 16.7% (down from 53.6% on GPT-5.6 Terra). Set effort explicitly per task class; do not benchmark or budget at default effort and expect max-effort accuracy.
- •Long-context behaviour beyond ~256K is undocumented for this variant: the 1M window is advertised but no MRCR v2 or equivalent retrieval score is published for Terra, and third-party context figures disagree (1.0M vs 1.05M vs 1.1M). Validate recall on your own needle tasks before relying on the tail of the window, and expect TTFT to scale sharply with prefill size.
- •Modality and endpoint ambiguity: OpenRouter metadata lists text as the modality while the launch blog markets detailed visual reasoning and computer use, so image/computer-use fidelity depends on the serving endpoint. Third-party vision and pixel-grounding tests report glyph/OCR-style weaknesses that would not appear in text-only evaluations.
- •Conflicting third-party aggregations and safety constraints: some leaderboards list AA Intelligence Index ~51.0 or Terminal-Bench 55.6 versus OpenAI's 50.3 and Benchmark Heaven's 48.1, reflecting unlabelled effort tiers and suite versions - treat version/effort-labelled numbers as canonical. Terra also carries a Critical cyber-risk rating in the Astra system-card addendum with 22 severity-3+ flags across 48,690 simulated Codex tasks, so autonomous shell/code execution needs guardrails and human checkpoints.
Calls to GPT-6.1 Terra are routed through strict Zero Data Retention inference channels. Prompts and outputs are never stored, indexed, or monitored by OpenAI.
More Models from OpenAI
Compare other engines available in this laboratory.
OpenAI's flagship September 2026 frontier model with a 1.05M context window, MoE architecture, and SOTA reasoning across coding, formal mathematics, and long-horizon agentic automation.
GPT-6.1 Sol (openai/gpt-6.1-sol) is OpenAI's September 29, 2026 mid-tier reasoning model, an upgrade to GPT-6 Sol positioned one step below the flagship GPT-6 Astra. It targets agentic coding, computer use, and document-heavy professional automation at roughly one-fifth of Astra's per-token cost, with a 1.05M-token context window, text+image input, and a measured output throughput of ~53 tokens/sec.
GPT-6 Sol is OpenAI's mid-tier, cost-efficient high-end GPT-6 model (launched 22 Sept 2026, ~three weeks after flagship GPT-6 Astra), exposing a 1,050,000-token context window, 128K max output, image input, selectable reasoning effort, and hidden test-time reasoning tokens at $2/$10 per 1M tokens — positioned between GPT-6 Astra (quality) and GPT-6 Luna (throughput/price).