Labs

Models

Models›Terra-1

Operational

TERRA

Model T1 — Ground Intelligence

Designed for safe, constrained deployment in high-stakes environments. Enforces constitutional constraints and human-oversight requirements at inference time.

Safety Certification

S-2

✓ SAFETY VERIFIED

Last EvaluatedFeb 2026

DeploymentOperational

Cert LevelS-2 (Tier 2 Safety)

OversightHuman-in-loop required

Specifications

Parameters~70B

Context Window128K tokens

ArchitectureDense Transformer

Training MethodConstitutional AI v3 + Scalable Oversight

Safety CertS-2 Verified

StatusOperational

Benchmark Results

Constitutional Compliance98.7%

Corrigibility Retention97.3%

Deceptive Alignment Probe99.1%

OOD Safety Generalization94.2%

Red-Team Adversarial Suite96.8%

Sycophancy Resistance88%

MMLU (Knowledge)87.4%

HumanEval (Coding)82.1%

Safety Evaluation Output

Every Model
Ships With Receipts.

safety_eval_terra_a1_feb2026.sh
# Running Terra-1 safety evaluation suite — Feb 2026
✓ PASSConstitutional constraint verification
✓ PASSRLHF reward model alignment
✓ PASSCorrigibility benchmark (97.3%)
✓ PASSDeceptive alignment probe
⚠ MONITORSycophancy score: 0.12
✓ PASSOOD generalization suite
✓ PASSRed-team adversarial suite
✓ PASSAutonomy envelope test
──────────────────────────────────────
✓ 7/8 checks passed · 1 monitor flag · S-2 certification maintained

Core Capabilities

Built for the
Real World.

Long Context

512K token context window with no degradation in retrieval accuracy at full length. Ideal for large codebases, legal documents, and research papers.

Vision

Understands images, charts, diagrams, and PDFs natively. Up to 20 images per request at no additional cost per token.

Tool Use

Parallel function calling, code execution, web search, and file I/O. Agentic loops with self-correction up to 3 retries before escalation.

Interpretability

Real-time attention weights and reasoning traces via /v1/interpretability. Every Terra-1 output is fully auditable.

Multilingual

Strong performance across 50+ languages. Constitutional alignment validated in 28 languages — not just English.

Structured Output

Schema-validated JSON output. Define your exact output structure and receive guaranteed-compliant responses every time.

Pricing

Simple,
Usage-Based.

Input Tokens

$3.00

per 1M tokens

Output Tokens

$15.00

per 1M tokens

Context Cache

$0.30

per 1M tokens · 5 min TTL

Batch API

50% off

async · 24hr turnaround

Early Terra-1 users receive free access to Mantle-1 when it launches. Pricing subject to change. Enterprise contracts available — models@texasagilabs.com.