Loading benchmark catalogue
Compare performance benchmarks, coding intelligence, reasoning evaluations, context limits, and API economics between DeepSeek V4 Pro 0813 (DeepSeek) and Beam 1.0 501B A23B (Reflection AI).
| Metric / Capability | DeepSeek V4 Pro 0813 | Beam 1.0 501B A23B |
|---|---|---|
| Developer / Organization | DeepSeek | Reflection AI |
| Release Date | 2026-08-12 | 2026-10-16 |
| Deployment / Weights | Open Weights | Open Weights |
| Context Window | 1.0M tokens (1,048,576 tokens) | 1.0M tokens (1,048,576 tokens) |
| Input Pricing (/1M Tokens) | Free (Open Weights) | Free (Open Weights) |
| Output Pricing (/1M Tokens) | Free (Open Weights) | Free (Open Weights) |
| Benchmark / Domain | DeepSeek V4 Pro 0813 | Beam 1.0 501B A23B | Advantage |
|---|---|---|---|
Artificial Analysis Intelligence Index Composite index aggregating 10+ evaluations of general reasoning and knowledge. | 36.0% | — | — |
Artificial Analysis Coding Index Composite index of coding and software-engineering evaluations. | 68.8% | — | — |
Artificial Analysis Agentic Index Composite index of tool-use and long-horizon agentic tasks. | 41.3% | — | — |
SWE-bench Pro Harder, contamination-resistant successor to SWE-bench spanning multiple languages. | 55.4% | 66.1% | Beam 1.0 501B A23B (+19%) |
ARC-AGI-2 Novel visual grid puzzles requiring abstract rule induction. | 61.3% | — | — |
GPQA Diamond Graduate-level physics, chemistry, and biology questions written by domain experts. | 90.1% | 90.6% | Beam 1.0 501B A23B (+1%) |
Humanity's Last Exam Expert-written questions across hundreds of disciplines designed to resist saturation. | 42.7% | 36.0% | DeepSeek V4 Pro 0813 (+19%) |
FrontierMath v2 (Tiers 1-3) Unpublished, research-level mathematics problems across Tiers 1-3. | 45.3% | — | — |
Design Arena ELO Human-preference ELO across UI, 3D, data-visualization, and game-dev generations. | — | 1227 | — |
Beam 1.0 501B A23B outperforms DeepSeek V4 Pro 0813 across 2 evaluated benchmarks compared to 1 for DeepSeek V4 Pro 0813, excelling in SWE-bench Pro, GPQA Diamond.
Both models feature equivalent context windows of 1.0M tokens (1,048,576 tokens).
Both models have comparable pricing structures or open weight deployment options.
DeepSeek V4 Pro 0813 is an open-weights model. Beam 1.0 501B A23B is an open-weights model.
Analyze 130+ models side-by-side using interactive radar charts, cost-efficiency quadrant plots, and live dataset filters on the full NeuBench index.
Open Interactive Comparator