A comparison tool, not a universal winner
Choose a model.
See the trade-offs.
Compare current API listings with a dated Text Arena snapshot. Quality, price, context and weight memory stay separate.
Checking source freshness…
Arena crowd preference is not task accuracy. A newer API model without a verified match is listed as unranked—not assigned another model’s score.
Base rates only. Long-context tiers, caching, tools and reasoning-token usage can change the bill. Unknown prices are excluded when a cap is set.
Decimal GB for weights only. KV cache, activations, batching and runtime need additional memory. A cap excludes unknown estimates and is not a promise that a model will fit.
Loading model data…
All columns remain available on small screens. Swipe sideways, or focus the table region and use the arrow keys.
| # | Model / tested setting | Text Arena rating | Input $/1M | Output $/1M | Context | Refusal audit | Weight GB | Evidence |
|---|
Read the evidence, not just the rank.
One quality source, with uncertainty
Ratings and confidence intervals come from the official Text Arena dataset (CC BY 4.0), overall category, latest published split, top 100. These are the dataset’s raw ratings, without the website’s optional style or factuality adjustments. They may differ from the adjusted leaderboard view. Overlapping intervals do not establish a clear winner. API listings outside this snapshot are unranked here; absence is not a bad score.
Exact identities before prices
API prices and context come from the OpenRouter model catalog. Joins use exact names/IDs or an explicit reviewed base-model alias. Versions, image variants and reasoning settings are never fuzzily substituted. The evidence cell names the actual API ID and any tested setting. A listed base rate does not reproduce Arena’s reasoning-token budget. OpenRouter passes through provider inference prices; credit-purchase fees are separate. See its pricing explanation.
Historical refusal evidence, not a safety ranking
Only exact audited API IDs receive rates from moderationbias.com. The cell shows the highest rate across that audit’s available substantive categories, excluding its false-positive control; this is LLM-judged historical evidence, not a current safety benchmark. The audit’s publication date is unverified, and its tested setting may differ from Arena’s. Unknown does not mean zero refusal.
Weight storage is not required VRAM
Where a public Hugging Face repository reports a parameter count, weight storage is parameters × 2, 1, or 0.5 bytes for FP16, INT8, or INT4. This estimate excludes KV cache, activations, runtime and quantization metadata. A weights-repository link is not a license or hardware-compatibility guarantee. Verify the model card and your deployment stack.
Freshness and failure
The benchmark publication date and price retrieval time are displayed separately. The updater refuses a missing, partial, mixed-date or more-than-14-day-old Arena snapshot and keeps the previous file. The page independently warns if a saved snapshot ages. A successful download is not proof that the underlying rankings changed.