Blog

We trained a 4B model on 1,000 expert-written rubrics from ComplexConstraints, our frontier instruction-following benchmark. It reached parity with a 60x larger model, and the gains transferred to external benchmarks it never saw.

A benchmark for long-context enterprise agents: MCP-native RL environments, expert-written handbooks up to 124 pages, deterministic grading. No frontier model exceeds 25%. Instead, they fire employees without authorization, approve self-submitted expenses, and send expired medical records to insurers.

Anthropic cited two Surge AI benchmarks, GDP.pdf and Riemann-bench, in their Fable 5 and Mythos 5 release. A look at why expert-built evaluations matter at the frontier.

A benchmark for entangled instruction following, where constraints depend on each other, fire conditionally, and must be inferred from context.

When Microsoft wanted to understand how MAI-Thinking-1 measured up, they used Surge human evaluations to prove it.

Post-training on Surge AI's agentic RL environments and why it generalizes to external tool-use benchmarks like Toolathlon, τ²-Bench, and BFCL-V4.

LMArena measures which answer you prefer in two seconds. Antidote measures which one you'd still be glad you got a month later, graded by doctors, lawyers, and engineers.

Can frontier models master the documents that run the world? GDP.pdf is a professional multimodal reasoning benchmark that takes real-world prompts and PDFs pulled directly from expert enterprise workflows.

Riemann-bench is a verifiable benchmark of extreme-tier mathematical problems where even frontier models score <10%.

Stop testing models in tiny, self-contained environments. We built CoreCraft, a large-scale startup world, and deployed AI agents to solve real tasks. Our goal: to move agents beyond the cleanliness of the lab and into the chaos of enterprise reality.

Stop rewarding slop. Hemingway-bench is an AI writing leaderboard that takes real-world writing tasks and puts them in front of master wordsmiths. Our goal: to push AI writing from two-second vibes to genuine nuance and impact.

Meta Superintelligence Labs partnered with Surge to build AdvancedIF, an instruction-following benchmark where every prompt and rubric was written by human experts – not synthetically generated by an LLM. In instruction-following domains, where frontier models still fail 22-30%, using these human-crafted rubrics as reward signals for RL yields a 13% gain.

Would you trust a medical system whose only metric was “which doctor wins the Internet?” No, you'd call that malpractice. Yet that's LMArena.

Our RL environment run on 9 models revealed the core capabilities all agents need to master: tool use, planning, adaptability, groundedness, and common sense.

We stress-tested GPT-5, Gemini 2.5 Pro, and Claude Sonnet 4.5 on 200+ expert finance tasks. Here's where even the best models break when they move from benchmarks to Wall Street.

After 100+ hours with Opus 4.1 and 20+ hours in the first week of Sonnet 4.5's launch, Nick Heiner, our VP of Product gives first impressions.

On Surge AI’s agentic coding benchmark, Claude Sonnet 4.5 outperformed GPT-5-Codex in accuracy, while GPT-5-Codex was more cost-efficient. Despite similar scores, the models were distinct in which tasks they failed in. In a refactoring case study, Claude succeeded after persistent debugging, while GPT-5-Codex failed due to an unexplained decision to end the task early. Both stayed focused and avoided hallucinations even when encountering difficulties.

At Surge AI, we work with the world’s sharpest minds to push the limits of AI. Professor Bogdan Grechuk – an IMO gold medalist and Associate Professor at the University of Leicester – is one of them. We interviewed him about the work he does to train SOTA models to perform frontier research.

When coding models spiral into self-reinforcing hallucinations, small mistakes compound into catastrophic failure. In SWE-bench, we saw SOTA models invent whole classes, methods, and terminal outputs – never realizing they had lost touch with the real codebase. In this case study, we’ll look at how three frontier coding agents tried to solve one particular SWE-bench problem: one spiraled into hallucinations and failed entirely, one spiraled but recovered, and one avoided hallucinations altogether. Our goal: to illustrate how dissecting real-world problems can steer models towards human-ready AGI.

Academic benchmarks make great headlines, and terrible AI.

The AI world loves climbing leaderboards. Companies race to hit #1 on LMSYS, chase perfect scores on academic benchmarks, and demo SVGs of pelicans on bicycles. These achievements make for great headlines and impressive presentations – even when these metrics are easily hacked.

GPT-5 was released on Aug 7, 2025. The swift removal of all legacy models from the ChatGPT UI was met with an even swifter backlash: some people online felt that GPT-4o was more personable, human, and engaging, whereas GPT-5 was stiff and robotic. This viral meme encapsulated the faction’s thesis:

An update on Astral Codex Ten's Image Generation Bet: close, but no dice. DALL·E 3 and Midjourney fail.

Learn how Anthropic partnered with Surge AI to gather high-quality human feedback at scale using the RLHF platform, resulting in one of the safest and most advanced large language models on the planet.

We measured ChatGPT vs. Google on 500 search queries, and found that ChatGPT crushes Google on coding and ties it on general information — despite not being optimized for a search experience at all. Dive into this post to learn more about OpenAI’s existential threat to Google.

How do you make large language models safer and adversarially robust to counterattacks? Learn about AI red teams of creative data labelers who try to interactively penetrate AI defenses in order to teach them.

We analyzed HellaSwag, a popular LLM benchmark, and found errors in 36% of its rows.

TikTok has been taking over the world — and now, your Google Search results too. But when are they actually helpful? We ran a large-scale personalized human evaluation, asking Surgers to rate hundreds of <query, TikTok> pairs to find out.

Has Astral Codex Ten's bet on AI progress really been won? We asked Surgers to evaluate DALL·E and Imagen on Scott's 5 compositionality prompts!

Why can't Meta A/B test its way back to greatness? To move Instagram beyond short-term engagement metrics, we ran a personalized human evaluation asking 100 users to compare TikTok vs. Instagram Reels. Learn why Gen Z considers Reels the place where TikToks go to die, and what Instagram should do about it.

Surge AI is partnering with NYU and the Fund for Alignment Research on the Inverse Scaling Prize. If you've found a task with LLM inverse scaling properties, and need help creating a dataset of 300-500+ examples, reach out. We’re a human alignment platform with deep expertise in training large language models on human feedback, and we’re here to help – including $500 of free data labeling credits to kickstart your submission.

Search quality measurement is one of the trickiest, but most important parts of building Search. Read how Neeva uses human evaluation of search quality to build a state-of-the-art search engine challenging Google.

Hugging Face's BLOOM is a new 176B parameter multilingual large language model. How does it compare to other state-of-the-art LLMs? We ran a human evaluation across 7 real-world categories to evaluate its performance.

Last year, Google released their “GoEmotions” dataset: a human-labeled dataset of 58K Reddit comments categorized according to 27 emotions. The problem? A whopping 30% of the dataset is mislabeled! Check out some of the egregious errors, and learn how to build better datasets.30% of Google's Emotions Dataset is Mislabeled

Our mission at Surge AI is to inject human values and intelligence into AI. We want to build a world where AI

Gary Marcus has several examples of AI mistakes. But are they really failures, or a sign of creativity? We gave them to 15 Surgers to complete GPT-3's "mistakes" to see how they would perform instead.

We built a dataset of 8,500 Grade School Math Problems for OpenAI. The goal of the dataset: to train language models like GPT-3 to solve natural language math problems and measure their reasoning ability. Learn about our process in this blog post!

Where do human artists fit in a world of rich, creative AI? We asked 100 Surgers to draw the DALL-E prompts.

Google Search is falling behind. We analyzed three areas – programming queries, sports queries, and cooking queries – to understand where Google Search lags behind its competitors.

Social media platforms optimize for clicks and engagement — but those same short-term optimizations drive clickbait, toxic content, and misinformation. How can we align their ML systems to human values instead? This post describes a data-driven approach with Facebook.

Are popular toxicity models simply profanity detectors? We show how toxicity models overweight profanity, and make mistakes when profanity is used in a positive way.

Has Google's Search Quality deteriorated in recent years? This post measures Google Search using human evaluation.

Data-centric AI requires radically rethinking the data that goes into your models. Surge AI provides data labelers with the skills you need to get context-sensitive labels.