Developers choice of LLM for a given task often hinges on which model excels in the relevant capability set. Philipp Dubach built Model Atlas with Codex to surface those differences, creating a tool that surfaces how published benchmark scores stack up when weighted by specific capabilities. The project emerged as a Codex artifact and now surfaces a benchmark catalog, published scores, inference settings, and coverage controls in a single index.
The snapshot dated 15 October 2026 maps 134 benchmark families across 12 capabilities. The index scores 37 evaluation profiles with enough comparable results, drawing 730 observations across 139 configurations. Comparison profiles draw on Artificial Analysis, Arena Creative Writing, MathArena, JobBench, and MCP-Atlas. JobBench tests professional work environments, while MCP-Atlas tests agents that invoke tools through the Model Context Protocol. The 15 profiles added in the latest review begin unselected, preventing the wider catalog from silently shifting the default comparison when a user opens a specific profile.
By default, each profile converts published scores to percentile ranks within its fixed group of tested configurations. Tied results share the middle rank, and percentiles preserve the ordering of results while discarding the magnitude of gaps between them. Filtering to open weights does not re‑rank against a smaller cohort. The index averages available profile scores within each capability, giving each profile equal weight. The capability weights the user applies then compose the combined score. When results are absent, coverage drops; coverage reflects the selected profiles and their weights. A lower coverage threshold allows a model to rank on partial evidence, which can advantage models whose reported results highlight their stronger suits.
The catalog draws from 14 release pages and model cards published by Google, Meta, OpenAI, Anthropic, DeepSeek, and NVIDIA. It logs test methods and limits, flagging benchmarks that currently cannot participate in comparison. An additional 366 release results stay outside the combined score. A benchmark appearing in a release table alone does not guarantee its scores are combinable with another source’s results. For instance, Meta’s Muse Spark Open scores 61.8 on JobBench and 89.2 on MCP‑Atlas. The snapshot’s leaderboards by model owner do not list Spark Open, so its JobBench result stands as a separate release record. Its MCP‑Atlas result joins a separate profile in Meta’s six‑model comparison, where it ranks first—but that result does not plug a gap in the model’s Artificial Analysis configuration nor dictate its rank across other capabilities.
Choosing and setting capability weights
The user selects weights that reflect the work at hand. Shifting weights reshapes the combined score by re‑balancing how each profile contributes within its capability. This lets a developer tune the comparison toward the skill set that matters—coding, writing, mathematical reasoning, professional tasks, or tool‑using agents—without altering the underlying percentile calculations.
How benchmark percentiles are computed
Each profile converts raw published scores into percentile ranks among the configurations it tested. Tied scores share the midpoint rank, and the resulting percentiles keep the order of results but lose the spacing between them. This means a difference of five points may matter more or less depending on how tightly the tested group clusters. The index then aggregates those percentiles, averaging them within each capability before the user’s weights lift or drag individual profiles up or down.
Seasonal and coverage considerations
Lowering the coverage threshold permits a model to place a rank using only the evidence that is present. This can benefit a model that reports strongly in a narrow domain while leaving broader capabilities untested. Because the system does not recompute percentile ranks when filtering to open weights, the relative standing of a model stays fixed within its original test group even when the weight set changes.
What the tool enables for developers
Model Atlas surfaces a pragmatic way to cut through the noise of dispersed LLM benchmark results. By weighting capabilities and tracking coverage, a developer can decide which model is worth a second look for a specific codebase, documentation workflow, or agent‑driven task. The catalog’s separation of release‑level results from the combined index means a single strong showing in one benchmark does not inflate a model’s overall standing. The 134 benchmark families and 730 observations give enough granularity to spot patterns—such as which models excel at MathArena versus which dominate JobBench—without over‑interpreting sparse data. The tool’s design, which began as a Codex artifact, stays lightweight enough to integrate into a developer’s existing review pipeline without a heavyweight setup.
Valve said the new achievements features are available in the latest Steam client beta. The full rollout to all users will follow the beta period, with no announced release date beyond that window.