Who Performs Model Evaluations?
Model evaluations are performed by three major groups, each with different goals and evaluation methodologies.
Understanding who is evaluating a model helps you correctly interpret benchmark results and choose the right model for your application.
1. Frontier Labs (Model Builders)
Who are Frontier Labs?
Frontier labs are organizations that build and release large language models.
Examples include:
- OpenAI
- Anthropic
- Google DeepMind
- Meta
- Mistral AI
- DeepSeek
- Qwen
- Cohere
- xAI
These organizations develop foundation models and continuously evaluate them throughout the development lifecycle.
Why Do Frontier Labs Evaluate Models?
Frontier labs perform evaluations for three main reasons.
1. Internal Development
The vast majority of evaluation happens before the model is released.
Engineers use evaluations to guide every stage of model development.
Evaluations help answer questions such as:
- Is the latest checkpoint better than the previous one?
- Did the new training data improve reasoning?
- Did post-training improve instruction following?
- Did safety tuning reduce harmful outputs?
- Which checkpoint should be released?
Without evaluation, model development becomes guesswork.
You can't improve what you can't measure.
2. Release Gating
Before releasing a new model, labs evaluate whether it is:
- Accurate enough
- Reliable enough
- Safe enough
- Stable enough
If evaluation results are poor, the model is not released.
This process is called release gating.
New Checkpoint
│
▼
Run Evaluations
│
▼
Pass?
┌────┴────┐
│ │
Yes No
│ │
Ship Continue Training
3. Marketing
The benchmark tables shown in launch blogs are also produced through evaluations.
Example:
| Model | MMLU | SWE-bench | GPQA |
|---|---|---|---|
| Model A | 90% | 72% | 81% |
| Model B | 87% | 65% | 76% |
These benchmark numbers help demonstrate improvements over competing models.
This is the evaluation work most users see publicly.
How Should You Interpret Frontier Lab Results?
Frontier lab benchmark scores should be viewed as a best-case performance estimate.
These results are often measured under carefully optimized conditions, such as:
- Carefully engineered prompts
- Few-shot examples
- Optimal decoding settings
- Large context windows
- High compute budgets
As a result:
- Scores may not reflect real-world production performance.
- Benchmark selection may favor the model's strengths.
- Configuration choices can significantly influence results.
Think of frontier lab numbers as a performance ceiling rather than an average real-world outcome.
2. Third-Party Evaluators
Who Are They?
Third-party evaluators are independent organizations whose primary purpose is to evaluate AI models.
Examples include:
- LMArena (Chatbot Arena)
- Artificial Analysis
- Epoch AI
- Vals AI
Unlike frontier labs, they do not build foundation models.
Their goal is to evaluate models fairly and independently.
Why Do They Evaluate Models?
Their business depends on being trusted as neutral evaluators.
They:
- Test models independently.
- Use standardized evaluation methods.
- Publish public leaderboards.
- Compare models under identical conditions.
Their focus is on fair comparisons rather than marketing.
Why Are Third-Party Evaluations Valuable?
For AI engineers, third-party evaluations are often the most useful because:
- Every model is tested using the same methodology.
- Results are directly comparable.
- Evaluations are less influenced by vendor-specific optimizations.
This makes it easier to compare competing models objectively.
What Do Third-Party Evaluators Measure?
In addition to benchmark accuracy, they often evaluate practical deployment metrics that model builders may not emphasize.
These include:
- Cost (price per million tokens)
- Latency (response time)
- Throughput (requests handled per second)
- Reliability
- Availability
- Real-world performance
These operational metrics are crucial when selecting a model for production systems.
How Should You Interpret Third-Party Results?
Because every model is tested under the same conditions:
- Comparisons are generally fairer.
- Benchmark scores are easier to trust.
- Performance differences are more meaningful.
For engineering decisions, third-party evaluations are often more informative than vendor-reported benchmark tables.
3. Companies and AI Engineering Teams
Who Are They?
This group includes organizations that use AI models to build real-world applications rather than creating foundation models.
Examples include:
- Startups
- SaaS companies
- Enterprises
- Internal AI teams
- Product engineering teams
Their goal is not to determine the world's best model but to identify the model that works best for their specific use case.
Why Do Companies Evaluate Models?
Public benchmarks measure general capabilities, but businesses need answers to application-specific questions.
For example:
- Can the model classify support emails correctly?
- Does it generate valid SQL queries?
- Can it summarize medical reports accurately?
- Does it follow our company's formatting rules?
- Is it fast enough for production?
- Is it affordable at our expected traffic volume?
These questions cannot be answered by generic benchmarks like MMLU or GSM8K.
Example: Customer Support Email Routing
Suppose a company is building an AI system to categorize incoming customer emails into:
- Billing
- Bug Report
- Feature Request
- Cancellation
Public benchmarks might show that Model A is the most capable model overall.
However, after evaluating both models on the company's own dataset:
| Metric | Model A | Model B |
|---|---|---|
| Category Accuracy | 94% | 91% |
| Cost / 1,000 Emails | $6.20 | $0.21 |
| p95 Latency | 4.1 s | 0.9 s |
Although Model A is slightly more accurate, Model B is dramatically cheaper and faster.
For this business, Model B may provide the better overall trade-off.
This illustrates why companies perform custom evaluations instead of relying solely on public benchmarks.
Comparing the Three Evaluators
| Evaluator | Primary Goal | Measures | Best For |
|---|---|---|---|
| Frontier Labs | Improve and release models | Capability, safety, benchmark performance | Tracking model development and launch performance |
| Third-Party Evaluators | Fair comparison across models | Accuracy, cost, latency, reliability | Independent model comparison |
| Companies / AI Teams | Solve business-specific problems | Task accuracy, cost, latency, business KPIs | Selecting the best model for production |
Complete Picture
Who Evaluates Models?
┌────────────────────────────────┐
│ 1. Frontier Labs │
│ Build and improve models │
└────────────────────────────────┘
│
▼
┌────────────────────────────────┐
│ 2. Third-Party Evaluators │
│ Compare models fairly │
└────────────────────────────────┘
│
▼
┌────────────────────────────────┐
│ 3. Companies & AI Teams │
│ Test models for their own use │
└────────────────────────────────┘
Key Takeaways
- Frontier labs evaluate models during development, release gating, and for public benchmark reporting.
- Their published benchmark numbers represent best-case performance under optimized conditions.
- Third-party evaluators independently test models using standardized methods, making comparisons fairer and often including operational metrics like cost and latency.
- Companies and AI engineering teams perform custom evaluations to determine which model best meets their specific business requirements.
- The "best" model depends on the use case—not just on leaderboard rankings. Yes. Below are production-ready, structured notes in the same style as the previous sections.