GLM 5.3 Abliterated: a decision record for API buyers

What the catalog actually confirms

Use the exact model ID glm-5.3-abliterated. UncensuredAI lists a 1,048,576-token context window and a 131,072-token maximum output. Its API guide documents Chat Completions through POST https://uncensuredai.com/v1/chat/completions, using a Bearer API key. These are published integration details; this research did not send a live inference request or validate the maximum limits.

The label identifies the offered Abliterated edition. It does not establish identical behavior or feature support to another GLM deployment.

A bill you can reproduce

Published prices per million tokens are $0.70 input, $1.95 output, $0.28 cache read and $0.70 cache write. For 10 million uncached input tokens and 2 million output tokens, the inference subtotal is $10.90: 10 × $0.70 + 2 × $1.95.

This example assumes no cache operations or additional charges. Budget retries separately; cache economics require actual reuse and the applicable write behavior.

Read speed together with outcomes

The saved public snapshot reports 134.84 token/s with 99.29% success and 5294 ms average request latency. Its exact window is 10 November 2026, 00:00:00–23:35:03 UTC. The 24-hour query includes the current partial hour and 23 complete hours.

Sample counts, prompt lengths and concurrency are not exposed. Request latency does not reveal time to first token. The throughput is an aggregate of observed traffic, not a guaranteed per-request rate; the reported non-success outcomes remain part of the purchasing decision.

What the aggregate leaves open

The public summary omits sample sizes, request lengths and simultaneous-load details. Request latency does not reveal time to first token. Throughput is an aggregate of observed traffic, not a guaranteed per-request rate, and the reported non-success outcomes remain part of the purchasing decision.

What a pilot still needs to establish

Function calling, strict JSON-schema output, image input and client-specific agent compatibility were not established for this exact endpoint by the evidence reviewed. Put each required capability on an acceptance checklist. Use a fixed, versioned prompt set, record generation settings and token counts, and separate successful responses from errors and retries.

For streamed responses, record request start, first token and final token separately. Report sample size and latency percentiles alongside throughput and task correctness.

Turn the record into a purchasing decision

The catalog describes shared capacity, possible HTTP 429 responses and no guaranteed fixed RPM or TPM quota. Include realistic concurrency and backoff in your pilot. If the model clears your acceptance criteria, its published rates give you a concrete starting budget.

The decision should rest on the acceptance results and the reproducible budget. Check the current catalog entry and use the connection guide to document the integration under test.