DeepSeek V4.1 Flash Abliterated: cost, capacity and observed service behavior

Record the integration target and baseline

The request model ID is deepseek-v4.1-flash-abliterated. The catalog records context capacity of 1,048,576 tokens and caps generated output at 131,072 tokens. The service documents POST https://uncensuredai.com/v1/chat/completions with a Bearer API key.

Those facts identify the offered API surface. They do not verify performance at the maximum limits, identical behavior to another DeepSeek edition, or every feature expected by a coding agent. This research used public documentation and telemetry without live inference.

Prices per million tokens are $0.27 input, $0.95 output, $0.06 cache read and $0.27 cache write. Six million uncached input tokens plus one million output tokens produce a $2.57 inference subtotal: $1.62 plus $0.95.

This baseline excludes cache operations, retries and additional charges. Measure reusable input and actual cache behavior before adding a savings projection. Include failed attempts and retries when estimating cost per accepted result.

Keep throughput and success in the same frame

The saved snapshot reports 151.69 token/s with 99.7% success and 7678 ms average request latency. Its window runs from 10 November 2026, 00:00:00 to 23:35:03 UTC. The requested 24-hour series includes one partial hour.

Sample counts, prompt lengths and concurrency are not published. Average request latency cannot establish time to first token. Throughput does not promise the same rate for another workload, while success documents observed outcomes rather than an uptime commitment.

Resolve capability questions before adoption

The reviewed evidence does not establish function calling, strict structured output, image input or compatibility with every agent client for this exact model ID. Turn required features into test cases with pass criteria, rather than inheriting assumptions from another endpoint.

Use a versioned prompt set and fixed settings. Record token usage, full-request duration, first-token time for streams, errors and retry count. Compare accepted outputs and total spend under comparable concurrency. Retain the sample count so later runs can be interpreted.

Shared capacity can return HTTP 429, and the catalog guarantees no fixed RPM or TPM quota. Test backoff and the concurrency your application needs.

If the pilot meets its quality and delivery criteria, the catalog supplies a transparent basis for planning spend. Record the current model terms and the documented endpoint with the evaluation.