Skip to main content
BenchLM
Data
Specialist vertical · BenchAlign v6.0 · 78 ranked (27 Supported, 51 Estimated)
Scored from the dataset of

Science and research leaderboard

As of November 21, 2026, Claude Fable 5.2 leads our science and research leaderboard at 94.1 (Supported, $8/$40 per 1M tokens), ahead of GPT-6.2 Astra at 90.1. Gemini 4 Neon (#3, 86.3) is the lowest-cost model near the top at $3/$15 per 1M tokens; Kimi K3.1 Pro leads open-weight models at #18 (65.4). We rank 78 models (27 Supported, 51 Estimated) across 18 domain protocols, and this vertical carries zero weight in the overall score.

Shortlist by workload and budget

Each pick comes directly from the live science & research rows below, with listed API rates and the case where another option wins.

Highest point score#1 · Supported

Anthropic · $8/$40 per 1M tokens

Costs $8/$40 per 1M tokens; Gemini 4 Neon (#3, 86.3) cuts token spend to $3/$15 per 1M tokens.

Closest contender#2 · Supported

OpenAI · $8/$40 per 1M tokens

Sits 4.0 points behind Claude Fable 5.2; test both on your own prompts when provider lock-in or latency matters.

Best API value#3 · Supported

Google · $3/$15 per 1M tokens

Keeps 92% of the top score (86.3 vs 94.1), but gives up 7.8 points on the hardest tail cases.

Top open-weight model#18 · Supported

Moonshot AI · Open weights · $2/$10 per 1M tokens

Trails Claude Fable 5.2 by 28.7 points (#18 overall here), so pick it when weight control or on-prem residency outweighs peak score.

Ranked models

Ordered by BenchAlign v6.0 point estimate as of November 21, 2026. An Estimated row backed by fewer than two source families never ranks above a row scoring within 2 points below it. Conditional ranges do not establish rank confidence.

Science & Research LLM Leaderboard
RankModelScoreEvidence
#1Anthropic94.1Supported
#2OpenAI90.1Supported
#3Google86.3Supported
#4OpenAI84.0Supported
#5Anthropic83.3Supported
#6OpenAI83.0Supported
#7Anthropic80.8Supported
#8Google80.3Supported
#9OpenAI79.4Supported
#10OpenAI77.1Estimated
#11Google73.9Supported
#12Google73.6Estimated
#13Anthropic73.3Supported
#14OpenAI71.9Supported
#15Anthropic69.8Supported
#16Anthropic67.0Estimated
#17OpenAI66.2Supported
#18
Kimi K3.1 ProOpen weights
Moonshot AI
65.4Supported
#19Meta66.4Estimated
#20Anthropic63.5Supported
#21Anthropic63.2Estimated
#22DeepSeek62.9Supported
#23Google61.6Supported
#24Google61.4Supported
#25OpenAI61.2Estimated
#26Google61.2Supported
#27Google60.9Estimated
#28Anthropic60.9Estimated
#29Anthropic60.9Estimated
#30Anthropic60.9Estimated
#31Anthropic60.8Supported
#32
GLM-5.3Open weights
Z.AI
60.7Supported
#33OpenAI60.6Supported
#34Thinking Machines Lab60.5Estimated
#35xAI60.0Supported
#36Google61.9Estimated
#37xAI59.7Estimated
#38
Kimi K3Open weights
Moonshot AI
59.5Supported
#39
GLM-5.2Open weights
Z.AI
59.4Estimated
#40Alibaba59.2Estimated
#41Google58.8Supported
#42OpenAI58.1Estimated
#43OpenAI57.6Supported
#44
Kimi K2.6Open weights
Moonshot AI
55.7Estimated
#45
MiMo-V2.6-ProOpen weights
Xiaomi
57.3Estimated
#46OpenAI55.0Estimated
#47
Muse Spark OpenOpen weights
Meta
56.4Estimated
#48
Mistral Large 4Open weights
Mistral
55.1Estimated
#49
MiniMax M3Open weights
MiniMax
54.9Estimated
#50xAI54.1Estimated
#51OpenAI51.9Estimated
#52
GLM-5.1Open weights
Z.AI
51.0Estimated
#53
Gemma 4 31BOpen weights
Google
50.9Estimated
#54
DeepSeek V4 ProOpen weights
DeepSeek
50.9Estimated
#55Alibaba50.6Estimated
#56OpenAI47.3Estimated
#57Google47.0Estimated
#58
DeepSeek-R1Open weights
DeepSeek
46.9Estimated
#59Xiaomi48.8Estimated
#60
Kimi K2.5Open weights
Moonshot AI
46.4Estimated
#61OpenAI45.0Estimated
#62
Qwen3.5 397BOpen weights
Alibaba
44.5Estimated
#63
Gemma 4 26B A4BOpen weights
Google
42.7Estimated
#64xAI41.2Estimated
#65Xiaomi42.3Estimated
#66Google42.3Estimated
#67
GPT-OSS 120BOpen weights
OpenAI
39.5Estimated
#68
DeepSeek V3.1Open weights
DeepSeek
39.5Estimated
#69
DeepSeek V3Open weights
DeepSeek
39.3Estimated
#70OpenAI39.3Estimated
#71Google37.4Estimated
#72Anthropic35.1Estimated
#73xAI37.0Estimated
#74Mistral32.8Estimated
#75OpenAI29.6Estimated
#76
GPT-OSS 20BOpen weights
OpenAI
27.9Estimated
#77OpenAI26.2Estimated
#78
Gemma 3 27BOpen weights
Google
23.6Estimated

Scored domain benchmarks (18)

Each percentage is the protocol's relative reference weight inside this vertical's calibrated BenchAlign v6.0 model, not a raw score average. Independent evaluator shares are capped at 40% whenever three or more independent evaluators contribute.

A separate Kaggle SciCode track for main research problems without background notes.

9% ref. weight

FrontierPhysics, from BenchFlow, gives an AI agent a research project written by a working physicist and checks whether it gets the physics right and writes it up well enough to publish. It has 56 tasks across nine fields of physics, and each one is run three times.

8% ref. weight

Accuracy on the 1,000-question HLE-Diamond set without tools at each model's highest available reasoning effort.

7% ref. weight

Accuracy on the 1,000-question HLE-Diamond set without tools, using the benchmark owner's high-reasoning evaluation.

7% ref. weight

Kaggle's few-shot run of the 448-question GPQA main set.

7% ref. weight

Kaggle's zero-shot run of the 448-question GPQA main set, kept apart from the harder Diamond subset.

7% ref. weight

A benchmark of AI agents completing expert-curated research workflows across the life, physical, Earth, mathematical, and engineering sciences.

7% ref. weight

An expert-authored benchmark designed to probe frontier knowledge and reasoning. BenchLM keeps protocol differences visible because tool-assisted and closed-book HLE runs answer different questions.

6% ref. weight

Tool-free variant of Humanity's Last Exam that isolates a model's raw frontier reasoning.

6% ref. weight

SciCode evaluates language models on generating code for realistic scientific research problems across 16 subfields of physics, math, chemistry, biology, and material science. Problems decompose into 338 subproblems requiring domain knowledge recall, scientific reasoning, and precise code synthesis. Based on real scripts from published research.

6% ref. weight

Kaggle's zero-shot run of the 198-question GPQA Diamond set, taken from the GPQA suite board. It stays separate from provider, Vals, Artificial Analysis, and Mercor Diamond results.

5% ref. weight

Kaggle's few-shot run of the GPQA Diamond set at task version 5, with the October 20 run kept below it.

5% ref. weight

Graduate-level science questions from provider comparison charts and open evaluations.

4% ref. weight

An end-to-end autonomous scientific research benchmark with 40 tasks across 10 scientific domains, where agents receive related literature and raw data, then attempt to rediscover the hidden target paper.

4% ref. weight

A benchmark of realistic biology-research tasks involving literature, figures, tables, databases, and bioinformatics files.

3% ref. weight

Computational biology challenges with objective answers that remained unsolved by independent human experts.

3% ref. weight

SciUniverse Level 1 tests whether models carry out scientific work across chemistry, biology, and materials science. Its 92 tasks include physical experiments, simulated lab work, and measurement analysis, with Pass@1 averaged equally across 17 task families.

3% ref. weight

An expanded version of GPQA that evaluates graduate-level knowledge and reasoning capabilities across 285 disciplines, providing comprehensive coverage of academic domains.

1% ref. weight

Executable research code, wet-lab protocol reasoning, and graduate STEM exams

Eighteen protocols anchor the science vertical, led by SciCode Main Standard (Kaggle) (9%), FrontierPhysics (8%), HLE-Diamond (highest effort) (7%) alongside Terminal-Bench-Science, HLE-Diamond, ResearchClawBench, LABBench2, BioMysteryBench, SciUniverse, and SuperGPQA. Rather than relying on textbook recall alone, the basket pairs PhD-level physics, chemistry, and biology derivations with executable scientific computing in SciCode and containerized research workflows in Terminal-Bench-Science.

Twenty-three of the 78 ranked models hold Supported labels, while recently launched frontier models with fewer independent reruns remain visible with Estimated labels until Kaggle and academic boards complete full sweeps.

Structured STEM benchmarks do not validate wet-lab safety or novel hypotheses

Solving closed-form quantum mechanics problems on FrontierPhysics or debugging numerical simulations in SciCode measures analytical execution on problems with known ground truth. It does not establish physical laboratory safety, reagent compatibility under real bench conditions, or statistical validity when interpreting noisy experimental assays.

Reasoning-heavy configurations also spend far more thinking tokens on HLE-Diamond and GPQA than standard chat defaults, so check effort settings and per-query latency before embedding a top-ranked model into an interactive literature assistant.

Matching frontier reasoning models to simulation, literature, and budget constraints

Claude Fable 5.2 (94.1, Estimated, $8/$40 per 1M tokens) and GPT-6.2 Astra (90.1, Estimated, $8/$40 per 1M tokens) are separated by just 4.0 points at the frontier, while Claude Fable 5.2 (#1, 94.1) leads Supported rows with broader multi-evaluator confirmation. Run both top models on your own domain code or derivation set before standardizing on one provider.

For batch literature screening or parameter sweeps where $8/$40 pricing is prohibitive, Gemini 4 Neon (#3, 86.3 at $3/$15 per 1M tokens) retains 94% of the leader's score at one-quarter the token rate, and Kimi K3.1 Pro (#18, 65.4, Supported) leads self-hosted weights.

Questions

Which LLM is best for scientific research right now?

Claude Fable 5.2 holds #1 on our scientific research leaderboard at 94.1 (Supported) as of November 21, 2026, ahead of GPT-6.2 Astra at 90.1 and Gemini 4 Neon at 86.3. Both top models hold Supported status across independent public evaluations.

What is the best open-weight LLM for scientific research?

Kimi K3.1 Pro from Moonshot AI leads open-weight models on this scientific research board at #18 with a calibrated score of 65.4 (Supported). That trails Claude Fable 5.2 by 28.7 points, so choose Kimi K3.1 Pro when self-hosting, VPC data residency, or weight customization matters more than peak frontier accuracy.

Which API model offers the best value for scientific research?

Gemini 4 Neon (#3, 86.3) is our top API value pick for scientific research, priced at $3/$15 per 1M tokens compared with Claude Fable 5.2 at $8/$40 per 1M tokens. You give up 7.8 calibrated points, making Gemini 4 Neon the better buy whenever those extra points do not cut downstream retries.

Which benchmarks make up the scientific research leaderboard score?

BenchAlign v6.0 calibrates 18 open and primary domain protocols for this board, led by SciCode Main Standard (Kaggle) (9%), FrontierPhysics (8%), HLE-Diamond (highest effort) (7%), HLE-Diamond (7%). Each percentage reflects share of the vertical's reference weight inside our item-response model rather than a raw score average, and no single independent evaluator can exceed 40% when three or more evaluators contribute.

How does the science leaderboard differ from the math and reasoning boards?

Math focuses on symbolic proof and competition problem solving, while general reasoning emphasizes domain-neutral logic and long-context deduction. This science board weights executable research software in SciCode and Terminal-Bench-Science alongside graduate physics, chemistry, and biology suites such as FrontierPhysics, GPQA Diamond, and LABBench2.

Why do some top science models show Estimated while lower rows show Supported?

Newly released models such as Claude Fable 5.2 enter with Estimated labels when fewer independent evaluators have finished running all 18 science protocols, whereas Claude Fable 5.2 (#1) has broader multi-evaluator confirmation. Single-source estimates cannot rank above models scoring within two points below them.

Science & research leaderboard updates

Weekly updates when new models or domain evaluations move the science & research rankings.

Read a sample issue

Join 7,000+ subscribers.

One email each week. Unsubscribe anytime.