Instructions and chat
We track 76 benchmarks in instructions and chat, or 82 if you count each version. They run from everyday level to expert level.
OpenAI held the top instructions and chat score 41% of the contested time
That's about 19 of the 46 months since January 2023, averaged over the benchmarks below. The easiest benchmarks are at the top. Versions of one test share a row, with a dark tick at each new version. We stop the clock, and fade the rest of the row, when the best score passes 85%, or at the newest score if 45 days go by without another.
OpenAI DeepSeek Z.ai Qwen Meta Anthropic Ai2 Google DeepMind Amazon ByteDance Seed Baidu Liquid AI Other labs MiniMax NVIDIA Shanghai AI Lab Skywork AI (Kunlun Tech) INF Technology (infly) Mistral AI Meituan
Everyday Most adults, without training
University Undergraduates and entry-level professionals
Share of contested time at number one
OpenAI 41% 
Qwen 15% 
Amazon 14% 
NVIDIA 8% Z.ai 5% 
ByteDance Seed 4% 
DeepSeek 3% Anthropic 2% Other labs 2% 
Google DeepMind 2% 
Mistral AI 1% Liquid AI 1% 
MiniMax 1% 
Baidu 1%
Qwen 15% Amazon 14% NVIDIA 8% Z.ai 5% ByteDance Seed 4% DeepSeek 3% Anthropic 2% Other labs 2% Google DeepMind 2% Mistral AI 1% Liquid AI 1% MiniMax 1% Baidu 1% Closed out: 9 benchmarks passed 85%
OpenAI 4 DeepSeek 2 Baidu 1 
Meta 1 Google DeepMind 1
DeepSeek 2 Baidu 1 Meta 1 Google DeepMind 1 - Feb 2026 Multi-IF ERNIE 5.0
- Aug 2025 COLLIE GPT-5
- Apr 2025 Arena-Hard v2.0 o3
- Jan 2025 AlpacaEval 2.0 DeepSeek-R1
- Dec 2024 Arena-Hard-Auto DeepSeek-V3
- Jul 2024 IFEval Llama 3.1-405B
- Jun 2024 MixEval GPT-4 Turbo Apr 2024
- May 2024 RewardBench Gemini 1.5 Pro May 2024
- May 2023 AlpacaEval GPT-4 Mar 2023
Scores come from each benchmark’s own leaderboard, from Epoch AI’s runs and from labs’ reports on their own models. We date Epoch runs by the model’s release and lab reports by when they came out. Labs pick their own settings, so a lab’s claim can beat an independent run of the same model. Turn off lab claims to see independent runs only.
OpenAI Anthropic Google DeepMind Meta DeepSeek Qwen Shanghai AI Lab Skywork AI (Kunlun Tech) INF Technology (infly)
Everyday Most adults, without training
School Students, from grade school to high school
University Undergraduates and entry-level professionals
Share of contested time at number one, independent runs only
OpenAI 52% Google DeepMind 20% Meta 18% Anthropic 7% DeepSeek 3%
Google DeepMind 20% Meta 18% Anthropic 7% DeepSeek 3% Only each benchmark’s own leaderboard and Epoch AI’s runs, dated by the model’s release. We leave out benchmarks that only labs report.
From everyday to expert level
Grouped by who could solve the tasks, from things most adults can do to open research problems. The bar under each name is its best score.
Everyday
Most adults, without training
- IFEval
- IFBench
- MultiChallenge
- EQ-Bench v2
- EQ-Bench 3
- AlpacaEval 2.0
- RewardBench
- Multi-IF
- MT-Bench
- MixEval
- EQ-Bench 4
- AgentIF (2026-10 setting)
- COLLIE
- AlignBench
- AlpacaEval
- SEAL Precise Instruction Following
- AgentIF
- AdvancedIF
- SysBench
- Internal API instruction following (OpenAI)
- Inverse IFEval
- MM-IFEval
- CFBench
- IFStruct
- MARS-Bench
- Meeseeks
- Apple side-by-side human evaluation
- Complex Workflow (ByteDance Seed in-house)
- IFEval (Mistral internal version)
- NV Chat Benchmark
- Brevity Benchmark
- CELLO
- Doubao Multi-Turn Bench
- EIFBench
- Infinite Generations (Mistral internal)
- InsCtrl (Tencent)
- MT-Bench-101
- Qwen Chinese human evaluation
- Robust IF Bench
- RoleBench
- RoleMRC
- SIFo
- xDailyBench
- Chat All-in-One Bench
- Chinese Complex Instruction Following (ByteDance Seed in-house)
- Consulting and Recommendations Benchmark (StepFun)
- DeepSeek complex instruction following and multi-turn writing evaluation
- General Assistant (MiniMax in-house)
- IF-Badcase
- IFEval-OOD
- InFoBench
- Instruction Following (MiniMax in-house)
- LengthCtrl (Qwen)
- MAI human side-by-side evaluation
- Meta Underspecified Intent evaluation
- Preference on ChatGPT and API prompts
- ThinkFollow (Qwen)
- CompassArena
- ComplexBench
- FollowBench
- Human feedback win rates (Anthropic)
- Kimi Chinese in-house open-ended benchmark
- LLMBoxing
- Real-scenario prompts human evaluation (Z.ai)
- RealCompanion
- SEAL Showdown
- Yupp VIBE Score
University
Undergraduates and entry-level professionals
Another 4 have no level. Levels don’t apply to arenas, safety evaluations or composite indices.
By kind
Following precise instructions
| Benchmark | Level | Scored | Best score | Status |
|---|---|---|---|---|
| IFEval Models answer about 500 prompts containing verifiable instructions such as word limits or required keywords. Code checks each instruction for strict and loose accuracy. | Everyday | Checked automatically | 98.9% | In use |
| IFBench 58 new out-of-domain verifiable output constraints. Scored by programmatic constraint checks (strict or loose). | Everyday | Checked automatically | 82.9% | In use |
| AgentIF (2026-10 setting) Models follow the same AgentIF instructions in thinking and non-thinking modes with vendor decoding settings. A gpt-5.6-luna judge scores 3 runs per model. | Everyday | Graded by a model | 67.4% | Not in cards |
| COLLIE Models generate text that meets compositional constraints at word, sentence, paragraph, and passage level. Scored by the share passing automatic constraint checks. | Everyday | Checked automatically | 99% | Fading |
| SEAL Precise Instruction Following Models answer 1,054 private prompts with detailed instructions in nine categories. Human annotators check 12 yes or no criteria and compare responses side by side on a 7-point scale. | Everyday | Graded by people | 92.0 | Not in cards |
| AgentIF Models follow long system prompts from 50 real agentic applications, each with many constraints. Scored by constraint and instruction success rates. | Everyday | Graded by a model | 59.8% | In use |
| MathIF Reasoning models solve 420 math problems from GSM8K, MATH-500, Minerva, Olympiad and AIME while obeying added verifiable instructions. Code checks the instructions and the answer. | School | Checked automatically | 50.7% | Not in cards |
| AdvancedIF Models respond to over 1,600 prompts with complex, multi-turn, and system-level instructions. A judge model checks each response against expert-written rubrics. | Everyday | Graded by a model |  85% | In use |
| SysBench Models follow 500 system messages across five user turns each. A model verifier checks constraints, instructions and whole sessions. | Everyday | Graded by a model | 81.2% | Fading |
| Internal API instruction following (OpenAI) Models answer prompts drawn from real developer use cases that carry instructions on formatting, verbosity and length. OpenAI reports a hard subset. | Everyday | Graded by a model | 65.8% | Dropped |
| Inverse IFEval Models follow 1,012 Chinese and English instructions that conflict with habits learned in fine-tuning, such as writing code without comments. A model judge scores compliance. | Everyday | Graded by a model | 80.3% | Fading |
| MM-IFEval Multimodal models follow image-grounded instructions with exact output constraints. Checks mix rule-based verification and a judge model. | Everyday | Checked automatically |  61.9% | Fading |
| CFBench Models follow 1,000 real-world instructions with embedded constraints, across more than 200 scenarios and 50 NLP tasks. A judge checks each constraint and the overall requirement. | Everyday | Graded by a model |  81% | Dropped |
| IFStruct Models generate JSON or YAML that must follow a requested schema and phrasing constraints, without constrained decoding. A response passes only if every structural constraint holds. | Everyday | Checked automatically | 93.7% | In use |
| Complex Workflow (ByteDance Seed in-house) Models carry out step-by-step tasks that follow a given standard operating procedure, such as multi-turn after-sales service in support, sales and healthcare. | Everyday | Graded by a model | 64.7% | Fading |
| IFEval (Mistral internal version) Models follow prompts with verifiable format constraints. Mistral uses its own fixed version of IFEval. | Everyday | Checked automatically | 87.4% | Dropped |
| CELLO Models follow real-world complex instructions with multiple tasks, noisy inputs and format or length constraints. Rule-based checks score each answer against the instruction features. | Everyday | Checked automatically | 76.4% | Dropped |
| EIFBench Models follow extremely complex instructions that bundle several tasks and many constraints in one request, as in real product workflows. | Everyday | Graded by a model | 48.6% | Fading |
| HANDBOOK.md Agents do routine company work with mock email, chat, calendar, issue and commerce tools while following a 20 to 124 page handbook. Programmatic criteria check required and forbidden actions, and a trial passes only if all pass. | University | Checked automatically | 44.6% | In use |
| InsCtrl (Tencent) Models write text under detailed instructions in a Tencent in-house text creation test. Task and scoring details are not published. | Everyday | Graded by a model | 71.9% | Dropped |
| Robust IF Bench Models follow a diverse set of hard instructions from Microsoft AI's internal benchmark. | Everyday | Graded by a model | 61.2% | In use |
| RoleMRC Models play given roles in multi-turn chats, passage reading tasks and nested or prioritized instructions. The test split has 1.4k samples. | Everyday | Graded by a model | 79.1% | In use |
| SIFo Models follow a chain of dependent instructions, where each step uses the output of the one before. Only the final answer is checked, across 800 samples in four tasks. | Everyday | Checked automatically | 77.3% | Dropped |
| Chinese Complex Instruction Following (ByteDance Seed in-house) Models follow complex Chinese production prompts with constraints on format, conditions, content, phrasing, tone, emoji, few-shot style and length. It has 912 cases over 17 weighted dimensions. | Everyday | Graded by a model | 75.3% | System cards only |
| Hard Capability (MiniMax in-house) Models handle complex productivity requests with several levels of instructions that leading models often fail. | University | Graded by a model | 64.8% | System cards only |
| IF-Badcase Models follow 450 multi-constraint instructions drawn from real user failure reports. A checklist checks each constraint. | Everyday | Checked automatically | 83.2% | System cards only |
| IFEval-OOD Models answer unseen WildChat prompts with added verifiable constraints drawn from 52 types outside IFEval. Scored by prompt-level loose accuracy. | Everyday | Checked automatically |  33.1% | System cards only |
| InFoBench Models follow 500 instructions that are split into 2,250 yes or no requirement questions. A judge answers each question, and the score is the share met. | Everyday | Graded by a model |  94.9% | System cards only |
| Instruction Following (MiniMax in-house) Models follow multi-level instructions collected from user cases where current models fail. | Everyday | Graded by a model | 46.3% | System cards only |
| LengthCtrl (Qwen) Models do creative writing tasks with length requirements. The score depends on the gap between output length and target length. | Everyday | Checked automatically | 73.5% | System cards only |
| Meta Underspecified Intent evaluation Models answer ambiguous requests sourced from real Meta AI chats where users corrected the reply. Rubrics score whether the true intent is met. | Everyday | Graded by a model | 71.2% | System cards only |
| Agentic IF Index Composite of Meta internal evals on following tool-use restrictions, rubrics, compound constraints and long policies in agentic workflows. Scored 0-1. | Graded by a model | Not in cards | ||
| ComplexBench Models answer 1,150 hand-written Chinese instructions that compose several constraints by chaining, selection and nesting. LLM evaluators with rules check each constraint, and scores follow the dependency structure. | Everyday | Graded by a model | Not in cards | |
| ComplexConstraints Models answer 75 expert-written professional prompts whose constraints depend on each other, fire conditionally or must be inferred. An LLM judge grades 1,559 rubric criteria one at a time, and a task passes only if all criteria pass. | University | Graded by a model | Not in cards | |
| FollowBench Models answer 820 English and Chinese instructions where each level adds one more constraint of content, situation, style, format or example type. Rules and an LLM judge check each constraint. | Everyday | Graded by a model | Not in cards |
Multi-turn conversation
| Benchmark | Level | Scored | Best score | Status |
|---|---|---|---|---|
| MultiChallenge Multi-turn conversations testing instruction retention, inference memory, versioned editing and self-coherence. Graded by LLM judge with rubrics. | Everyday | Graded by a model | 82% | In use |
| EQ-Bench 3 Models respond in character in multi-turn role-plays and analyze transcripts of social conflict. A judge model gives rubric scores and pairwise Elo ratings. | Everyday | Graded by a model | 2,096 | Fading |
| Multi-IF Models follow verifiable instructions over three-turn conversations in eight languages, extending IFEval. Scored by instruction accuracy at each turn. | Everyday | Checked automatically | 89.1% | In use |
| MT-Bench 80 two-turn questions graded by GPT-4 as judge. Introduced LLM-as-judge methodology. | Everyday | Graded by a model | 9.35 | Fading |
| EQ-Bench 4 Models hold 16-message roleplay chats with 120 simulated user personas whose preferences compete and are hidden. A three-judge panel compares transcripts pairwise, and ratings are fitted as Elo scores. | Everyday | Graded by a model | 1,365 | Not in cards |
| MARS-Bench Models handle long multi-turn dialogues built from play-by-play sports commentary. Tasks test ultra multi-turn, interactive multi-turn and cross-turn reasoning. | Everyday | Checked automatically | 85.6% | Fading |
| Meeseeks Models follow complex instructions over several turns and get feedback on failed requirements after each turn. Scored by how many requirements they meet. | Everyday | Graded by a model |  43% | Dropped |
| Doubao Multi-Turn Bench Models answer the turn of a real Doubao conversation where the user was dissatisfied, with the earlier turns as context. Rubrics grade each answer. | Everyday | Graded by a model | 52.5% | In use |
| MT-Bench-101 Models hold 1,388 multi-turn dialogues with 4,208 turns across 13 tasks from a three-tier ability taxonomy. GPT-4 scores each turn from 1 to 10. | Everyday | Graded by a model |  8.59 | Dropped |
| DeepSeek complex instruction following and multi-turn writing evaluation Models answer 196 hard Chinese writing prompts with complex constraints or several turns. Pairwise judging against a baseline model gives a win rate. | Everyday | Ranked by matches | 45.9% | System cards only |
| ThinkFollow (Qwen) Models hold multi-turn dialogues with randomly inserted think and no-think flags and must switch reasoning mode correctly. | Everyday | Checked automatically | 98.9% | System cards only |
| RealCompanion Models read months of real chats between ten people and an AI companion, then answer questions about who each person is. The release has 27,218 messages over up to 120 days. | Everyday | Checked automatically | Not in cards |
Overall chat quality
| Benchmark | Level | Scored | Best score | Status |
|---|---|---|---|---|
| LMArena (Elo) People chat with two anonymous models side by side and vote for the better reply. Votes become Elo-style ratings for each model. | Ranked by votes | 1,528 | Fading | |
| EQ-Bench v2 Models read 171 short dialogues and rate how strongly a character feels each of four emotions on a 0 to 10 scale. The score compares the ratings with reference answers. | Everyday | Checked automatically | 86.4 | Not in cards |
| AlpacaEval 2.0 Models answer 805 open instructions, and a GPT-4 Turbo judge compares each reply with GPT-4 Turbo's own. The score is a win rate. | Everyday | Graded by a model | 87.6% | Fading |
| Arena-Hard v2.0 Models answer 500 hard real user prompts from Chatbot Arena and 250 creative writing prompts. GPT-4.1 or Gemini 2.5 judges each answer against a baseline answer from o3-mini, and the score is the win rate. | University | Graded by a model | 90.2% | Fading |
| Arena-Hard-Auto 500 hard arena prompts judged by an LLM against a baseline. Cheap proxy for arena rank. | University | Graded by a model | 97.1% | Fading |
| WildBench 1,024 challenging real-user tasks with checklist-guided LLM judging. | University | Graded by a model |  59.4 | Dropped |
| RewardBench Models act as judges and pick the preferred response in chosen and rejected pairs across chat, hard chat, safety and reasoning. Scored by accuracy. | Everyday | Checked automatically |  95.1% | Dropped |
| RewardBench 2 Reward models must score one correct answer above three wrong ones on 1,865 new human prompts. Subsets cover factuality, precise instruction following, math, safety, focus and ties, and the score averages them. | University | Checked automatically |  84.1% | Not in cards |
| MixEval Models answer real user queries matched to similar questions from existing benchmarks. Ground-truth answers make it a cheap proxy for Chatbot Arena rankings. | Everyday | Checked automatically | 89.9% | Dropped |
| AlignBench Models answer 683 real Chinese user queries across eight categories. A GPT-4 judge rates each answer from 1 to 10 against a reference. | Everyday | Graded by a model | 8.94 | Dropped |
| AlpacaEval Models answer 805 open instructions, and a GPT-4 judge compares each reply with text-davinci-003's reply. The score is the win rate. | Everyday | Graded by a model | 97.7% | Not in cards |
| CriticEval Models critique responses across four dimensions in nine task scenarios, giving both scalar scores and written critiques. GPT-4 grades written critiques against a large set of annotated reference critiques. | University | Graded by a model | 7.86 | Not in cards |
| Apple side-by-side human evaluation Apple's in-house graders compare two models' responses to locale-specific text and image prompts and pick the preferred one. Results are win, tie and loss fractions per comparison. | Everyday | Ranked by votes | 64.7% | Fading |
| JudgeBench Models act as judges and pick the correct response from hard pairs on knowledge, reasoning, math and coding. The label comes from objective correctness, not human preference. | Expert | Checked automatically | 79.1% | Dropped |
| NV Chat Benchmark Models answer chat prompts from an internal NVIDIA set. The Nemotron 3.5 Super VL card labels the score an LMArena v3 proxy. | Everyday | Graded by a model | 52.7 | In use |
| Brevity Benchmark NVIDIA evaluation of short, direct answers for voice and device assistants. The card gives no link or method. | Everyday | Graded by a model | 82.8% | In use |
| Infinite Generations (Mistral internal) Models answer challenging long or repetitive prompts from Mistral's internal set. The score is the share of responses that never stop or loop, so lower is better. | Everyday | Checked automatically | 1.29% | Dropped |
| Qwen Chinese human evaluation Three annotators rank chat model answers to 300 Chinese instructions on knowledge, language, writing, math and coding. Results are win rates against GPT-3.5. | Everyday | Ranked by votes | 30.7% | System cards only |
| RoleBench Models answer prompts while playing one of many fixed characters from films and books. Answers are compared with reference role answers by automatic text overlap. | Everyday | Checked automatically | 71.2 | In use |
| xDailyBench Models answer open-ended requests grounded in real user needs across more than 30 daily-life, study and office scenarios. Rubrics grade completion, correctness, coverage and usefulness. | Everyday | Graded by a model | 61% | In use |
| Chat All-in-One Bench Models handle conversational scenarios built around real online user needs at every stage of product usage. | Everyday | Graded by a model | 85.2% | System cards only |
| Consulting and Recommendations Benchmark (StepFun) Models answer 500 advice queries taken from Reddit, Stack Exchange and forums. Model judges score usefulness, logic, instruction following and tone against an expert reference set at 88. | Everyday | Graded by a model | 70.5% | System cards only |
| General Assistant (MiniMax in-house) Models answer real assistant queries from Hailuo AI users. Automatic assessments score the replies. | Everyday | Graded by a model | 73.9% | System cards only |
| MAI human side-by-side evaluation Paid raters compare two models' answers to 1,276 English prompts from experts and Copilot logs on a 7-point scale from -1.5 to 1.5. | Everyday | Ranked by votes | 0.07 | System cards only |
| Preference on ChatGPT and API prompts Human labelers compare model responses to 5,214 prompts sent to ChatGPT and the OpenAI API. Scored as win rate against a baseline model. | Everyday | Ranked by votes | 70.2% | System cards only |
| Antidote Experts such as doctors, lawyers and engineers compare model answers to prompts from their own history. They judge which answer would still help them a month later. | Expert | Graded by people | Not in cards | |
| CompassArena Users chat with two anonymous models side by side, mostly in Chinese, and vote for the better answer. A Bradley-Terry model turns the votes into Elo-style ratings. | Everyday | Ranked by votes | Not in cards | |
| Human feedback win rates (Anthropic) Raters chat with two models on tasks such as coding, documents and creative writing and pick the better reply. Score is the win rate. | Everyday | Ranked by votes | System cards only | |
| Kimi Chinese in-house open-ended benchmark Models answer held-out Chinese prompts drawn from real user queries. Outputs are compared pairwise against another model to give win, tie and loss rates. | Everyday | Ranked by matches | System cards only | |
| Llama 2-Chat helpfulness human evaluation (Meta) Three human raters compare Llama 2-Chat with a rival chat model on about 4,000 single and multi-turn helpfulness prompts. | Ranked by votes | System cards only | ||
| Llama 3 pairwise human evaluation (Meta) Human raters compare Llama responses with a rival model on about 7,000 prompts across six capabilities and three multiturn sets. | Ranked by votes | System cards only | ||
| LLMBoxing People see one question with two anonymous model answers and pick the better one. The site counts preferences between model pairs. | Everyday | Ranked by votes | System cards only | |
| Real-scenario prompts human evaluation (Z.ai) One human evaluator scores replies to 660 curated real user prompts in English, Chinese and other languages on a 0 to 10 scale. | Everyday | Graded by people | System cards only | |
| SEAL Showdown Users of Scale's Playground app chat with a model and sometimes compare its reply with a random opponent's reply. Scale ranks models with a Bradley-Terry model that controls for length and formatting. Rankings can be split by region, profession and age. | Everyday | Ranked by votes | Not in cards | |
| Yupp VIBE Score People use Yupp for real tasks, see answers from two or more models, and pick the one they prefer. The preferences become VIBE scores for each model. | Everyday | Ranked by votes | Not in cards |