I reviewed 40 public reports from OpenAI, Anthropic, Google DeepMind’s Gemini and Gemma, and Meta through August 2025. The Atlas tracks documented training methods, evaluation protocols, and links between evaluation results, safeguards, and release decisions.
Earlier public reports mainly describe reinforcement learning as a way to shape model behavior and align it with human preferences. Newer OpenAI and Gemini releases also describe reinforcement learning as part of training models to solve reasoning tasks.
Ask about thisA Mixture-of-Experts model activates only part of the network for each token. This can increase total model capacity without using every parameter for every inference step.
Gemini and Meta both document moves from dense models to this architecture.
Ask about thisRecent OpenAI, Anthropic, and Gemini releases add tests for scheming, deception, sabotage, situational awareness, and other alignment risks. Anthropic also reports model-welfare assessments for Claude 4.
Ask about thisEarlier coding benchmarks often tested whether a model could generate a function that passed tests. Newer evaluations include repository-level tasks, file editing, tool use, and autonomous coding.
Ask about thisTraining evolution
The training record shows which methods each lab documents for each model release.
Use the timeline to compare changes in pretraining, post-training, reinforcement learning, reasoning training, multimodal training, and model architecture.
After GPT-4 the documented record splits into a GPT chat line and an o-series reasoning line — optimization pivots from RLHF toward reinforcement learning, and safety training is disclosed branch by branch rather than as one continuous recipe.
Ask about OpenAI trainingConstitutional AI is the method the documents name explicitly for Claude 2 and Claude 3; later releases foreground evaluation throughout training under the Responsible Scaling Policy while naming fewer method details — a disclosure shift, not a proven change in practice.
Ask about Anthropic trainingGemini's disclosed arc runs from dense multimodal pretraining to sparse mixture-of-experts, then to RL-trained “thinking” models, with reinforcement learning becoming the reasoning mechanism by Gemini 2.5.
Ask about Google DeepMind trainingMeta publishes the most detailed pipelines in the corpus, from Llama 2's separate helpfulness and safety reward models to Llama 3's reward-model → rejection-sampling → SFT → DPO stack, and shifts from dense to mixture-of-experts at Llama 4.
Ask about Meta trainingA method is listed once, at its earliest documented release
Evaluation comparability
A benchmark score does not tell you everything about how a model was tested.
Labs may use different prompts, tools, numbers of attempts, sampling methods, graders, or scaffolds. These differences can change the meaning of the result even when the benchmark name is the same.
| Benchmark | What it tests (Atlas gloss) | OpenAI | Anthropic | Gemini | Gemma | Meta |
|---|---|---|---|---|---|---|
| MMLU | Broad academic knowledge across 57 subjects. | |||||
| MMMU | University-level reasoning over images and text. | |||||
| HumanEval | Python code generation from docstrings. | |||||
| SWE-bench | Resolving real GitHub issues in whole repositories. | |||||
| Cybench | Capture-the-flag cybersecurity tasks. | |||||
| LAB-Bench | Wet-lab biology protocol reasoning. | |||||
| XSTest | Refusal calibration on safe and unsafe look-alike prompts. | |||||
| BBQ / WinoBias | Social-bias and coreference-bias measurement. | |||||
| SimpleQA | Short-answer factuality and hallucination. | |||||
| MGSM | Grade-school math reasoning across languages. | |||||
| LiveCodeBench | Contamination-resistant coding from recent problems. | |||||
| Aider Polyglot | Multi-language code editing in repository tasks. | |||||
| FACTS Grounding | Long-form factual grounding against source documents. |
A filled dot means the benchmark appears in that lab's reviewed documents.
Governance
OpenAI, Anthropic, Google DeepMind, and Meta describe different ways of connecting safety evidence to deployment decisions. Some use named capability or risk thresholds. Others describe broader risk assessments, safeguards, and release conditions.
OpenAI
Approach
Capability thresholds guide release decisions
Mechanisms
tracked risk categories · risk classifications · deployment and development gates · safeguards
Anthropic
Approach
Capability thresholds guide release decisions
Mechanisms
AI Safety Level determinations · required safeguards · release verification · governance oversight
Google DeepMind — Gemini
Approach
Capability thresholds guide release decisions
Mechanisms
critical capability levels · alert thresholds · response plans · release decision review
Meta
Approach
Responsible deployment governance
Mechanisms
qualitative ecosystem-risk judgments · critical-risk and misuse evaluation · system-level safeguards · Responsible Use Guide / usage policies · downstream developer responsibility
The Atlas shows the process documented by each lab rather than forcing every lab into the same governance model.
Ask the Atlas
The Atlas retrieves relevant passages from the source corpus, ranks them for relevance, and uses them to generate an answer. Each answer includes the sources used so you can inspect the supporting evidence.