A low pass rate is never the goal. What matters more is why the model failed.

Today, we are introducing SciCode++, a new dataset for assessing how well frontier models solve scientific problems through code.

Introducing SciCode++: Scientific coding data for frontier models

The original SciCode idea

SciCode-Verified shifted the challenge

SciCode++: Separating model failures from evaluation failures

Anatomy of a SciCode++ ‘task’

The test code for validation

What building SciCode++ at scale tells us about difficulty

Does the signal transfer? A post-training experiment

Building SciCode++ at scale

References

Introducing KernelQuest: Can agents write Triton kernels?

The challenge: speed is easy to claim and hard to verify

Why we built KernelQuest

What KernelQuest measures

Quality is part of the benchmark, not a post-processing step

One task shows why the gates matter

Leaderboard

Failure modes

Building the evidence needed to push frontier models forward

CEO Bench evaluates whether frontier agents can complete long-horizon financial and operational work inside a complete, expert-authored fictional company rather than merely crunch numbers in isolation.

Introducing CEO Bench: Evaluating frontier agents on real company work

The challenge: measuring judgment, not string matches

Introducing CEO Bench

From gold answers to graded judgment

Quality and methodology

Evaluation results

Failure-mode analysis

Next steps

Developed by Turing Frontier Research Lab, CyberStrike evaluates if agents can complete offensive, defensive, and incident response cybersecurity tasks.

Announcing CyberStrike: Measuring whether frontier agents can complete real cybersecurity work

Security work does not end at detection

The challenge: measuring completion, not partial progress

Introducing the CyberStrike benchmark

From plausible answers to verified outcomes

Task spotlight: when self-reported success does not survive replay

How CyberStrike advances the lab’s mission

Agents often solve the wrong problem, verify against their own misreading, and mark the task complete.

Terminal-Bench 3.0: Why capable agents fail confidently

Why self-verification fails

The data

Failure has moved upstream

The agent's own evidence is uninformative

Types of failure

Misreading the problem

Under-modeling the spec

Missing the numeric bar

Breaking under re-entry or concurrency

The traps persist across configurations

What independent verification requires

What we take from this

We identify systematic mistakes incurred when running agents on operational tasks in real enterprise settings.

CompanyBench: Evaluating Frontier AI Agents On Real Operational Knowledge Tasks

What makes CompanyBench different

Results: Frontier models still fall short

Systematic failure patterns

Conclusion

Explore CompanyBench

Terminal-Bench 3.0 evaluates agents on hard command-line work. Turing contributed to the benchmark, with some of the tasks authored and verified by our team. We share how we built these tasks and where the benchmark is headed.

Terminal-Bench 3.0 evaluates agents on hard command-line work. Turing contributed to the benchmark, with some of the tasks authored and verified by our team.

Terminal-Bench 3.0: Hard for the Right Reasons

Building to a higher standard

What a task consists of

Proving a task is fair

Looking beyond the leaderboard

Where this goes

The future of AI will be shaped by the data and deployment loop. Discover how real enterprise workflows, training data, and rapid iteration drive frontier model performance and economic impact.

Coding and Enterprise SuperIntelligence is Won on Data and Deployment

Data companies don’t see deployment. Deployment companies don’t see data. Turing sees both.

Every deployment sharpens the next model.

The next decade is won by closing the loop fastest.

Enterprise Superintelligence Will Be Won on the Data and Deployment Loop

OpenEnv is an open-source framework from Meta and Hugging Face for creating standardized, isolated, and reusable environments for training and deploying AI agents, especially for Reinforcement Learning (RL) and agentic workflows.

Evaluating Tool-Using Agents in Production-Oriented Environments with OpenEnv

What is OpenEnv - A Comprehensive Overview

Introduction to OpenEnv

Evolution from Classical RL Frameworks

From OpenAI Gym to OpenEnv

OpenEnv’s Modern Architecture

OpenEnv’s Domain Coverage

1. Browser Automation (BrowserGym)

2. Calendar Management (Turing’s Contribution)

3. Code Development (Coding Env)

4. Gaming Environments (OpenSpiel, Atari, Snake)

5. Financial Trading (FinRL)

6. Text-Based Games (TextArena)

Core Concepts and Architecture

Step-State-Reset Paradigm

Model Context Protocol (MCP) Integration

WebSocket Communication Protocol

Deployment and Operations

Containerized Environments

CLI Tooling

Monitoring and Observability

Comparison Summary

Why OpenEnv is Important and What It Achieves

Bridging Research and Production

Turing’s Technical Contribution: The Calendar Gym

Why We Chose the Calendar Environment

1. Real-World Complexity

2. Alignment with Turing’s Expertise

3. Measurable Outcomes

Architecture of the Calendar Gym

Key Technical Innovations

How to Use the Calendar Gym

Defining Reward Functions

Verifiers: Automated Success Criteria

Documentation Links

Insights Gained from RL Gyms

Key Findings from Calendar Gym Benchmarks

1. Multi-Step Reasoning is the Bottleneck

2. Ambiguity Resolution is Underrated

3. Tool Selection is Not Enough

Based on delivery patterns across dozens of enterprise programs, here are five projections that will define the year and how Turing is positioned to help you meet them with confidence.

AI in 2026: Five Projections Every Enterprise Must Prepare For

Projection 1: Governance becomes the universal requirement, not a differentiator

Projection 2: Industry specific intelligence replaces general-purpose models for core workflows

Projection 3: Continuous evaluation becomes the backbone of enterprise-scale AI

Projection 4: Efficiency and sustainability reshape model strategy at scale

Projection 5: Human expertise becomes the defining multiplier for enterprise AI maturity

The bottom line for 2026

See how 2025 shifted AI from general tools to proprietary intelligence. Explore key themes in agentic workflows, vertical specialization, and data governance.

2025: The Year of Proprietary Intelligence | Turing

Key Themes from 2025

Where This Is Already Working

Looking Ahead

2025: The Year Enterprise AI Became Proprietary Intelligence

RL environments are a foundational layer in how frontier labs evaluate, harden, and iterate on AI agents before deployment. This article explains why structured simulation environments are essential for reliable agent behavior and how they fit into a mature post-training workflow.

Learn why frontier AI labs rely on RL environments to evaluate agent reasoning, tool use, and workflow stability in simulation before real-world deployment.

RL Environments for Agent Evaluation: Why Simulation Comes Before Deployment

Ad Hoc Evals Happens in Production. Structured Evaluation Happens in Simulation

What RL Environments Provide That Human Oversight Cannot

A Feedback Loop That Reduces Downstream Instability

Structured Digital Twins for Agent Research

To Build Reliable Agents, Start With the Environment

Why Labs Use RL Environments Before Agents Touch Real Workflows

Explore expert insights on AI at Turing’s blog. Learn how AI-powered solutions are driving innovation, model performance, and business growth.

Turing Blog: Articles, Insights, Company News and Updates

Insights, analysis, and updates from Turing research labs  on the latest evaluations and benchmarks.

Blog

Blog posts

Request dataset samples or subscribe to our weekly newsletter to stay informed of what’s happening at the frontier of AI.

Blog

Introducing SciCode++: Scientific coding data for frontier models

Blog posts

Introducing SciCode++: Scientific coding data for frontier models

Introducing KernelQuest: Can agents write Triton kernels?

Introducing CEO Bench: Evaluating frontier agents on real company work

Announcing CyberStrike: Measuring whether frontier agents can complete real cybersecurity work

Terminal-Bench 3.0: Why capable agents fail confidently

CompanyBench: Evaluating Frontier AI Agents On Real Operational Knowledge Tasks

Terminal-Bench 3.0: Hard for the Right Reasons

Enterprise Superintelligence Will Be Won on the Data and Deployment Loop

Evaluating Tool-Using Agents in Production-Oriented Environments with OpenEnv

AI in 2026: Five Projections Every Enterprise Must Prepare For

2025: The Year Enterprise AI Became Proprietary Intelligence

Why Labs Use RL Environments Before Agents Touch Real Workflows

Contact us