Updates
Every model release, safety update, and research milestone — in one place.
Terra-1 v1.4 — Extended Context & Tool Use
Context window expanded to 512K tokens. Significant improvements to multi-step tool use, code execution accuracy, and structured output reliability.
- 01Context window: 128K → 512K tokens with no degradation in retrieval accuracy
- 02Tool use success rate improved by 34% on internal benchmarks
- 03Structured JSON output now schema-validated before returning to caller
- 04Code execution sandbox hardened — new file system isolation layer
- 05Response latency reduced by 18% via speculative decoding improvements
- 06TTFT (time to first token) reduced from 420ms to 280ms at p50
Real-Time Interpretability Dashboard — Public Beta
TAL Corp's Glass-Box interpretability toolkit enters public beta. Developers can now inspect token-level attention patterns, reasoning traces, and confidence signals via the API.
- 01New API endpoint: /v1/interpretability — returns attention weights and reasoning trace per response
- 02Token attribution scores available for all Terra-1 outputs
- 03Confidence calibration improved — model uncertainty now correlates with actual error rate at 0.91 Brier score
- 04Anomaly detection layer added to production serving — flags out-of-distribution requests automatically
TAL API v2.0 — Streaming, Batching & New SDKs
Major API update introducing server-sent event streaming, batch inference endpoints, and official SDKs for Python, TypeScript, Go, and Rust.
- 01SSE streaming now available on all inference endpoints
- 02Batch API: submit up to 1,000 requests asynchronously at 50% cost reduction
- 03Official Python SDK v2.0 — async-first, type-safe, Python 3.9+
- 04Official TypeScript SDK v2.0 — full type coverage, ESM and CJS support
- 05Go SDK v1.0 — production-ready
- 06Rust SDK v0.9 — beta release
- 07Rate limits increased: Tier 1 from 60 RPM to 200 RPM, Tier 2 to 2,000 RPM
Terra-1 v1.3 — Reasoning & Math
Focused improvement release targeting multi-step reasoning, mathematical problem solving, and logical deduction. Terra-1 now scores in the 94th percentile on MATH-500.
- 01MATH-500 benchmark: 78.4% → 91.2% (pass@1)
- 02AIME 2025 problems: solved 18/30 without external tools
- 03Chain-of-thought reliability improved — self-consistency across 8 samples up 22%
- 04Logical deduction on ARC-Challenge: 89.7% accuracy
- 05Reduced hallucination rate on factual questions by 31%
- 06New parameter: reasoning_depth — "thorough" for longer CoT or "fast" for quick responses
Constitutional AI v3 — Value Lock Update
Third major update to TAL Corp's Constitutional AI framework. Expanded value taxonomy, improved refusal calibration, and new multi-language alignment coverage.
- 01Value taxonomy expanded from 48 principles to 94
- 02Over-refusal on benign requests reduced by 41% without compromising safety
- 03Alignment validated in 28 languages — previously only English and 6 others
- 04New grey area handling — model asks clarifying questions rather than refusing
- 05RLHF dataset expanded by 2.4M human preference comparisons
- 06Red team evaluation: 0 critical failures across 12,400 adversarial test cases
Terra-1 v1.2 — Multimodal Vision
Vision capabilities added to Terra-1. The model can now understand, describe, and reason about images, charts, diagrams, and documents with inline images.
- 01Image understanding added — JPEG, PNG, WebP, GIF up to 20MB per image
- 02Document parsing: PDFs with embedded images extracted natively
- 03Chart and graph interpretation — data extracted from bar charts, line graphs, scatter plots
- 04MMStar benchmark: 72.8% accuracy (vs GPT-4V baseline of 68.4%)
- 05Up to 20 images per request in the context window
- 06Vision capabilities included at no additional cost per token
Global Inference Infrastructure — 4 New Regions
TAL Corp expands inference infrastructure to 4 new regions, reducing average latency for users in Asia-Pacific, Europe, and South America by up to 68%.
- 01New inference regions: Singapore, Frankfurt, São Paulo, Tokyo
- 02Automatic routing to lowest-latency region based on user geography
- 03Average global p50 latency reduced from 380ms to 140ms TTFT
- 04SLA upgraded to 99.95% uptime across all regions
- 05Multi-region redundancy: requests failover within 50ms
- 06CDN integration — common prompts served at edge with <20ms latency
Terra-1 v1.1 — Code & Agentic Improvements
Significant improvements to code generation, debugging, and agentic task completion. Terra-1 now supports function calling with parallel tool execution.
- 01HumanEval benchmark: 71.3% → 84.6% pass@1
- 02SWE-Bench: resolves 28.4% of real GitHub issues
- 03Function calling redesigned — supports parallel tool calls in a single turn
- 04New tool types: code_interpreter, web_search, file_read, file_write
- 05Agentic loop: model self-corrects tool call errors up to 3 times before escalating
- 06Multi-step task completion on AgentBench improved by 19 points
- 07Context caching — repeated system prompts cached for 5 minutes, reducing cost by up to 60%
Terra-1 v1.0 — General Availability
TAL Corp's first production model launches to general availability — the first frontier model built with a real-time interpretability dashboard and a constitutional alignment framework.
- 01General availability across all TAL Corp API tiers
- 02128K context window with strong retrieval accuracy at full length
- 03MMLU: 89.4% accuracy across 57 academic domains
- 04Safety score: 99.2% on TAL Corp's internal red-team evaluation (12,400+ test cases)
- 05Real-time interpretability dashboard available for all API users
- 06Constitutional AI v2 alignment — trained on 8.4M human preference comparisons
- 07Pricing: $3.00 / 1M input tokens, $15.00 / 1M output tokens
Research Paper: "Scaling Alignment with Intelligence"
TAL Corp publishes its foundational alignment research paper at NeurIPS 2025, introducing the RLHF-Constitutional hybrid training method that powers Terra-1.
- 01Paper: "Scaling Alignment with Intelligence: A Hybrid RLHF-Constitutional Approach"
- 02Published at NeurIPS 2025 — oral presentation
- 03Introduces TAL-RLHF-C: combines RLHF reward modelling with constitutional AI at the gradient level
- 04Demonstrates safety properties scale with model size rather than degrading
- 05Training code, dataset, and evaluation suite open for academic use
TAL API v1.0 — Developer Preview
TAL Corp's API launches in developer preview. 500 early-access developers onboarded, with access to Terra-1 via a clean REST interface.
- 01REST API with OpenAI-compatible message format for easy migration
- 02Python SDK v1.0 released
- 03Developer dashboard — usage tracking, API key management, billing
- 04Rate limit: 60 requests/minute for all preview users
- 05Maximum context: 32K tokens in preview (128K at GA)
- 06Streaming available via SSE
TAL Corp Open Charter — Irrevocable Commitment
TAL Corp files its Open Charter with an independent international governance board — a legally binding commitment that all TAL models will remain permanently accessible.
- 01Open Charter filed with the International AI Governance Board
- 02Core commitment: TAL Corp models will never be privatised or paywalled for basic access
- 03Charter enforced by an independent 9-member board — majority non-TAL Corp
- 04Future access restrictions require unanimous board approval and 90-day public notice
- 05Charter covers: model weights, training code, safety evaluation suites, and API access
- 06Full charter text published at texasagilabs.com/charter
TAL Corp Founded — Alignment-First AGI Lab
TAL Corp is founded in Texas with a commitment to building Artificial General Intelligence that is safe, interpretable, and universally accessible.
- 01TAL Corp incorporated in Texas
- 02Founding team: 12 researchers and engineers from leading AI labs and universities
- 03Seed funding secured from mission-aligned investors
- 04Three founding research pillars: Alignment, Interpretability, Agentic Safety
- 05First research hire: Head of Alignment, PhD from Stanford AI Lab
- 06Compute partnership — 10,000 A100 GPU hours committed to alignment research