Abliterated API capabilities: confirmed facts and an evaluation checklist

Exact IDs and published ceilings

Use glm-5.3-abliterated, glm-5.3-flash-abliterated or deepseek-v4.1-flash-abliterated as the exact request ID. For each edition, the published context ceiling is 1,048,576 tokens and the output ceiling is 131,072 tokens. The documented endpoint is POST https://uncensuredai.com/v1/chat/completions, authenticated with a Bearer API key.

Capabilities that still need a test

No live inference was performed, and the published limits were not load-tested. Function calling, strict JSON-schema output, image input and compatibility with particular agent clients remain unverified for these exact editions.

Read the economics by token category

Per million tokens, GLM 5.3 lists $0.70 input, $1.95 output, $0.28 cache read and $0.70 cache write. GLM 5.3 Flash lists $0.22, $0.80, $0.08 and $0.22 respectively. DeepSeek V4.1 Flash lists $0.27, $0.95, $0.06 and $0.27.

At five million uncached input tokens plus one million output tokens, the respective inference subtotals are $5.45, $1.90 and $2.30. These examples exclude cache operations, retries and additional charges. They compare published token costs, not equal-quality completed tasks.

Dated throughput and success observations

The snapshot reports GLM 5.3 at 134.84 token/s and 99.29% success; GLM 5.3 Flash at 45.84 token/s and 98.39% success; DeepSeek V4.1 Flash at 151.69 token/s and 99.7% success.

Interpret the snapshot within its limits

The exact observation window is 10 November 2026, 00:00:00–23:35:03 UTC: the current partial hour plus 23 complete hours. The summary does not expose request totals or workload mix. These are not controlled comparative results or guaranteed rates, and request-latency averages do not establish time to first token.

A reproducible decision protocol

Version your prompt set and acceptance criteria. Hold model ID, settings, prompt sizes and concurrency constant within each run. Log input/output usage, errors, retries and cost per accepted result. Measure first-token and full-request time separately for streams; report sample counts and percentiles.

Test required capabilities individually. Shared capacity may produce HTTP 429, and fixed RPM/TPM quotas are not guaranteed. Include backoff behavior and realistic concurrency before accepting a model for production.

Take a qualified shortlist into evaluation

A qualified shortlist is the starting point for an application-specific evaluation, not the conclusion. Check the current catalog, prepare the API integration, and retain only the capabilities supported by the test record.