ReasonCore AI @ NeurIPS 2026

Terminal-Bench-Science launched in August with a top score of 30%. Today Claude Fable 5.2 reports 74.6%. Three things worth noting: 1/ Resolution. It's 70 tasks, and the system card gives a task-clustered standard error of 4.3 points. Fable 5.2's 0.7-point gap to Mythos 5.2 is noise, and its 6.5-point gap to GPT-6 Astra is about 1.5 SE. 2/ Supply. 920 proposals became 70 merged tasks, a 7.6% yield after expert and technical review. Hard, verifiable science tasks are being used up faster than they're built. 3/ Signal. Near 75%, more tasks get solved on nearly every trial, and they stop separating models or giving RL a gradient.

A hard problem with an unreliable answer key is bad training data. Frontier models need challenging tasks and solutions that hold up under scrutiny. We build both: expert-authored scientific and coding reasoning tasks with verified solutions. ReasonCore has 5,000 CritPt-style tasks, 5,000 SciCode-style tasks 1,000 Terminal-Bench-Science-style tasks, and Terminal-Bench 4-style tasks available. Building your next post-training run? DM us for samples.

Hard benchmarks make headlines. For training, “hard” is the wrong target. A science task every model fails teaches RL nothing. So does one every model passes. The signal lives in the middle: a model solves it 1 to 6 times out of 8. Hitting that band is the expensive part. One terminal-bench-science task can take a frontier agent up to 3 hours. Certifying it takes 8 runs, and all 8 again every time a new model ships or a single file changes. That’s up to 24 agent-hours before a task is worth handing to a lab.

We're scientists and coders — geeks and nerds. We specialize in post-training AI models with SciCode, CritPt, DeepSWE, and Terminal-Bench-Science style training tasks. If you're a researcher and want to see samples, DM us here or checkout ReasonCore.ai.

SciCode, 2024 → 2026: what happened, and why linkedin.com/posts/anshumanl…

ReasonCore AI has a data partnership with @PlusAI_Inc where we transform their autonomous vehicle stills and videos for RL for frontier and world model training. Gain access to millions of training sets with expert reasoning with demonstrated headroom! benchmarks.reasoncore.ai/spa…

ReasonCore AI published a benchmark called "The SciCode 500." It's tested against Grok 4.5, Ox Alpha, DeepSeek V4 Pro, and Muse Spark 1.1. The benchmark has wide range of solvability by design. The 100% headroom problems are verified solved by an off benchmark SOTA model. SciCode is a specialty of ReasonCore and we have 5,000+ in inventory. We are building inventory in CritPt as well. Inquire within! reasoncore.ai/scicode500

Most AI benchmarks are saturating. CritPt is not. CritPt ("critical point") asks whether frontier LLMs can do real physics research. 50+ active physicists wrote 71 research-scale challenges from their own unpublished work. These are problems you'd hand an incoming PhD student, with nothing to memorize. Answers are guess-resistant and machine-verifiable: numerical arrays, symbolic expressions, working Python functions. No LLM-as-judge handwaving. At launch, the best model scored 5.7% on full challenges. Models handled isolated checkpoints far better than composed research. The gap is sustained, integrated reasoning. Under a year later, top models are near 30%. Fast progress, with 70% of the map still dark. This is what good evaluation looks like: novel, contamination-resistant, aimed where models actually break. It's also our specialty at ReasonCore, where we build frontier scientific benchmarks like CritPt and SciCode. Paper: arxiv.org/abs/2509.26574