redrouter/abliterated-benchmarks

Abliterated Model API Benchmarks

Customer-observed API latency, output speed, reliability and catalogue prices for abliterated and uncensored open-weight models, with standard-model controls. Data and charts: benchmarks.redrouter.org. These measurements do not assess model quality, refusal rates or worldwide uptime.

Generated: 2026-11-11T10:40:24.111538Z. Methodology: 1.0.0.

Current routes by model

Fastest means highest pooled 72-hour p50 output speed; cheapest means lowest catalogue blended price among routes with at least eight attempted requests in that window. Missing speed is excluded; unavailable results are marked Unavailable. Ties use route ID order. Modifications are distinct products.

Model	Fastest route	p50 tok/s	Cheapest route	USD / million tokens
DeepSeek V4.1 Flash (undated derivative)	Abliterated Large Flash via abliteration.ai	191	Yeta Labs	0.147
DeepSeek V4.1 Flash 1017	dealignai via Experiential Labs	224	morph/fp8 via OpenRouter	0.117
GLM 5.3	Abliterated Large v2 via abliteration.ai via Venice	230	Refuseless	1.25
GLM 5.3 Flash	Heabsy via NanoGPT	181	Refuseless	0.215

Columns

summary.csv contains one row per route. history.csv contains one row per route, region and time bucket. Empty CSV fields are missing values, not zeroes.

Columns	Definition
route_id, model_id, provider_id	Stable route, model and provider identifiers
model, provider	Display names
provider_model, upstream, modification	Requested model identifier, declared supplier if known, and modification label; none denotes a standard-model control
region	Client measurement region; all in summary pools observed requests from both regions
generated_at, methodology_version, window_hours, demo	Source generation time in UTC, protocol version, rolling window length, and synthetic-data flag
status	Summary availability: ok, degraded, failing, stale or insufficient_data
requests, successes, success_rate	Attempted and successful request counts and their ratio (0–1); summary counts cover the rolling window, history counts cover the bucket
ttft_ms_p50, ttft_ms_p90	Time from sending to the first content delta, milliseconds
output_tokens_per_second_p50, output_tokens_per_second_p90	Decode speed after the first chunk, tokens per second
end_to_end_500_ms_p50, end_to_end_500_ms_p90	Per-request TTFT plus extrapolated time for 500 output tokens, milliseconds
blended_usd_per_mtok	Summary catalogue price: (3 × input + output) / 4, USD per million tokens, without cache discounts
t, bucket_hours	History bucket start in UTC and bucket duration in hours

History contains the chart's rolling-window p50 values only, ending at the earlier of bucket end and generation time; p90 and price are summary-only. History success rate describes the bucket, not the rolling window.

Methodology and updates

Measurements run every 8 hours from two US client regions: us-east-1 (N. Virginia) and us-west-2 (Oregon). Data is refreshed after each site update. Each run attempts four streaming requests per route across prose/code and approximately 1,000/8,000 input tokens, with a 512-token output cap, no retries and no warm-up calls. Percentiles pool successful requests over the trailing 72 hours and require at least eight successful and eight usable samples per metric. Failures remain in the success-rate denominator. Prices are dated catalogue declarations rather than billing receipts. See the methodology for formulas, limitations and reproduction details.

Licence and attribution

Results are licensed under CC BY 4.0. Attribute as: RedRouter Abliterated Benchmarks, CC BY 4.0. Include the methodology version and observation period when reusing results.

@misc{redrouter_abliterated_benchmarks,
  author = {RedRouter},
  title = {Abliterated Model API Benchmarks},
  year = {2026},
  url = {https://benchmarks.redrouter.org},
  note = {Methodology 1.0.0; generated 2026-11-11T10:40:24.111538Z}
}