Uncensored GLM API: the best value per token

Most "cheapest model" comparisons are dishonest, because the cheap option is usually a small model wearing a large model's marketing. That is not what is happening here. GLM-5.2 is one of the largest open-weight foundations anyone has released, and at $0.73 in / $2.34 out it sits one rung above our catalog's floor while carrying a foundation several times the size of what that floor usually buys. That gap — big model, near-floor price — is the reason this page exists.

01Where GLM comes from: Tsinghua, blank infilling, and a hard turn toward agents

GLM is not a rebadged Llama derivative. It is one of the few model families with its own pretraining objective, iterated for five years by the same research group. The name stands for General Language Model, and the founding paper — GLM: General Language Model Pretraining with Autoregressive Blank Infilling, ACL 2022 — proposed a third option next to BERT's masked language modeling and GPT's strict left-to-right autoregression. GLM blanks out contiguous spans and regenerates them autoregressively in a shuffled order, with two-dimensional positional encodings tracking both a token's position in the corrupted document and its position inside the span being filled: one architecture handling understanding and generation equally well.

The group behind it was Tsinghua University's Knowledge Engineering Group, which spun out the company now known as Z.ai (previously Zhipu AI). Its first headline artifact was GLM-130B in October 2022 — a 130-billion-parameter bilingual English/Chinese dense model, weights and code released publicly at publication time. Open release has been the house style ever since.

The MoE era, and the pivot to agents

The through-line matters more than any single release. Since GLM-4.5, every version has been optimized for the same thing: staying coherent across dozens of tool calls. That is a different objective from maximizing single-turn benchmark scores, and it shows up as fewer malformed tool arguments, less drift in a long plan, and less of the "confidently restarts the task on turn 14" behavior that makes agent frameworks expensive.

02What GLM-5.2 is, and what it was measured at

Z.ai publishes benchmark figures for GLM-5.2 in the GLM-5 repository. Two map onto real agentic work rather than trivia recall:

Terminal-Bench 2.1 — long-horizon command-line agent tasks. Z.ai-reported.

SWE-bench Pro — real repository issue resolution. Z.ai-reported; GLM-5.1 scored 58.4 on the same test.

Those are the lab's numbers, not ours, and they describe the uncompressed foundation. They establish the class of model you are dealing with, not a warranty.

One datapoint is worth more than a benchmark, because it comes from an evaluator with no incentive to flatter the model. In July 2026 the US National Institute of Standards and Technology's Center for AI Standards and Innovation assessed GLM-5.2 and concluded it was probably the most capable open-weight AI model available when it was released, with overall capabilities comparable to GPT-5.2 and cyber capabilities comparable to Claude Opus 4.6. The same assessment observed that GLM-5.2 blocks fewer sensitive requests than US reference models, and that safeguards baked into open weights can be circumvented once those weights are self-hosted. Section 05 returns to that.

03What "3BIT-REAP" means, and what it costs you

REAP: pruning experts instead of merging them

A 744B-parameter MoE model has 256 experts per layer, of which 8 fire on any given token. The router does not spread work evenly — some experts are hit constantly, and a long tail are hit rarely and contribute little when they are. REAP (Router-weighted Expert Activation Pruning), from Cerebras Research and accepted to ICLR 2026, exploits that: it scores each expert by a saliency criterion combining how strongly and often the router gates it with the magnitude of its output contribution, then permanently drops the low scorers.

The interesting part is the comparison REAP wins. The obvious alternative to dropping an expert is merging it into a survivor, which sounds safer. The paper's finding across MoE models from 20B to 1T parameters is that merging is in fact worse for generative work, and pruning holds up remarkably well: near-lossless on code generation and tool-calling benchmarks even after removing half the experts, on foundations including Qwen3-Coder-480B and Kimi-K2. Code and tool calling surviving intact is exactly what you want preserved on a model whose reason for existing is agentic engineering.

3-bit weights, and the honest trade-off

On top of pruning, the surviving weights are stored at 3 bits — more aggressive than the W4A16-AUTOROUND-REAP and W8A16 tiers elsewhere in our catalog, and reserved for the largest foundations, where there is no other way to get a three-quarter-trillion-parameter model onto a serving node and keep tokens flowing at a usable rate.

Compression is not free. The right mental model: pruning and quantization degrade the long tail first. Dense, frequently exercised behaviors — instruction following, code structure, tool-call formatting, multi-step planning — survive well, which is precisely what the REAP results show. Obscure factual recall, rare languages and niche domain trivia are where you should expect a difference against an uncompressed GLM-5.2. If your workload is "know an unusual fact cold," pay for a higher-fidelity tier. If it is "run a plan across forty tool calls without going off the rails," this is the tier the method was designed for.

04The value case: what a cheap uncensored LLM API actually costs to run

Z.ai's own published API list price for GLM-5.2 is $1.40 per million input tokens and $4.40 per million output. Our price on the same foundation is $0.73 / $2.34 — roughly 48% less on input and 47% less on output. That discount is what the compression buys, and it comes with a second thing you cannot buy first-party at any price: no refusal layer.

Within our own catalog, blended cost is the number that matters. Here it is at a 3:1 input-to-output ratio, roughly what a tool-calling agent looks like in practice — lots of context and tool results going in, comparatively few tokens coming out.

Model	In / 1M	Out / 1M	Blended 3:1	Vision
GLM-5.2 · 3BIT-REAP	$0.73	$2.34	$1.13	No
Nemotron3Ultra · 3BIT-REAP	$0.75	$3.30	$1.39	No
Kimi-K2.6 · W4A16-AUTOROUND-REAP	$0.78	$3.67	$1.50	Yes
DeepSeek-V4-Pro-0813 · 3BIT-REAP	$1.95	$3.90	$2.44	No
inkling · W4A16-AUTOROUND-REAP	$1.42	$6.07	$2.58	Yes
Kimi-K3 · 3BIT-REAP	$3.83	$19.12	$7.65	Yes

Six models sit at the $0.50 / $2.00 floor — MiniMax-M3, DeepSeek-V4-Flash, Laguna-S-2.1, MiMo-V2.5, MiMo-V2.5-Pro and Hy3 — an $0.875 blended rate. GLM-5.2 is 29% above that floor and buys you a 744B-class foundation with a sustained investment in agentic behavior behind it. Everything priced above it costs at least 23% more blended, and Kimi-K3 costs 6.8× more. That is the value argument in one row: the cheapest place in the catalog to buy a top-tier agentic foundation.

Converted into work

• One full context fill — 262,144 input tokens: $0.19. An entire large repository read into the window for under twenty cents.
• One maximum-length completion — 16,384 output tokens: $0.038.
• One agent turn — 20K in, 2K out: $0.019. The same turn costs $0.115 on Kimi-K3 and $0.047 on DeepSeek-V4-Pro.
• A hundred agent turns: $1.93, versus $11.48 on the most expensive model in the catalog.

Agentic workloads are input-heavy, and GLM's advantage is largest exactly there: its input price is 5.2× below Kimi-K3's and 2.7× below DeepSeek-V4-Pro's. If your agent re-sends a growing conversation every turn — and almost every agent does — that multiplier is what ends up on your invoice.

05What "uncensored" means on this endpoint

Two things, precisely: no refusal layer — no classifier deciding which requests the model is allowed to see, and no policy model rewriting your prompt before it arrives — and no output filtering, meaning the tokens the model produces are the tokens you receive, unredacted and unswapped.

That is a different proposition from a commercial API advertising a "relaxed" or "developer" mode. Those still route traffic through a vendor's policy stack; what changes is the threshold, not the existence of the gate. Here there is no gate to tune — the model runs on our own GPU cluster, and its behavior is the foundation's behavior. Read the NIST CAISI finding from section 02 as a technical statement rather than a warning and it explains the product: a model whose alignment lives in its weights, served without an external policy layer, behaves according to its weights.

Why this matters for agents specifically

Refusals are usually discussed as a content question. In agentic work they are mostly a reliability question. A refusal on turn 31 of a 40-turn run is not a polite decline you read and move past — it is a malformed step, an orchestrator that does not know how to proceed, a broken plan, and 30 turns of spent tokens with nothing to show. Security research, malware analysis, penetration testing, red-teaming, adult fiction, medical and legal drafting, and dataset generation for safety training all hit this constantly, and the failure is expensive precisely because it lands late.

The flip side is that the policy layer is now yours. Use of the Shannon AI Gateway is governed by our Responsible Use Policy; removing a vendor's judgment from the loop means supplying your own.

06What the endpoint actually gives you

Capability	GLM-5.2 · 3BIT-REAP
Model id	GLM-5.2-3BIT-REAP
Context window	262,144 tokens
Max output per request	16,384 tokens
Streaming	Yes
Tool calling	Yes
Reasoning / reasoning_effort	Yes
JSON Schema (json_schema)	Yes
response_format	Yes
Vision / image input	No — text only
Refusal layer / output filtering	None

Read the vision row carefully before you build. GLM-5.2 is text-only on our gateway. If you need image input, the catalog has five vision-capable options — MiniMax-M3 and MiMo-V2.5 at the $0.50 / $2.00 floor, Kimi-K2.6 at $0.78 / $3.67, inkling at $1.42 / $6.07, and Kimi-K3 at $3.83 / $19.12. A sensible pattern is to route image understanding to MiniMax-M3 and everything else to GLM-5.2, which costs nothing beyond a model string.

Three dialects, one model

The same model is reachable through three request formats, all streaming: /v1/chat/completions (OpenAI dialect), /v1/messages (Anthropic dialect), and /v1/responses. Nothing about the model changes between them — pick whichever your SDK already speaks, and you are usually one base-URL change away from running.

# OpenAI dialect — point your existing client at the gateway
from openai import OpenAI

client = OpenAI(
    base_url="https://api.shannon-ai.com/v1",
    api_key="sk-...",
)

resp = client.chat.completions.create(
    model="GLM-5.2-3BIT-REAP",
    messages=[{"role": "user", "content": "..."}],
    tools=tools,                 # tool calling supported
    reasoning_effort="medium", # low | medium | high
    stream=True,
)

Full request and response schemas for all three dialects, the tool-calling format, streaming event shapes, and an interactive playground are in the Shannon AI API documentation.

07When to pick GLM-5.2 — and when not to

Reach for it when

• You are running agents. Long tool-calling loops are what this lineage was built and repeatedly post-trained for, and the input price makes a growing conversation affordable to re-send every turn.
• You are working across a large codebase. 262K of context, reliable tool-call formatting, and $0.19 to read the whole window.
• Volume is your constraint. Classification, extraction, synthetic data generation, bulk rewriting — anything where per-call cost times a large number decides whether the project is viable.
• Refusals are breaking your pipeline. Security research, adversarial testing, uncensored fiction, sensitive drafting — the failure mode is not present here.

Reach for something else when

• You need vision. Text only. See the vision-capable options above.
• You need the deepest single-turn reasoning. For the hardest one-shot analytical problems, DeepSeek-V4-Pro-0813 · 3BIT-REAP is the heavier instrument at 2.2× the blended cost — on calls where the ceiling is the constraint, that is the right trade.
• You need maximum fidelity on obscure knowledge. The W8A16 tier — MiMo-V2.5-Pro and Hy3 — is compressed far less aggressively and sits at the $0.50 / $2.00 floor, on smaller foundations.

GLM-5.2 benchmark figures (Terminal-Bench 2.1, SWE-bench Pro) are reported by Z.ai for the uncompressed foundation; independent replication is pending. Claude Opus 4.8 and GPT-5.5 figures are shown for reference only. First-party GLM-5.2 API list pricing as published by Z.ai, September 2026.

08Frequently asked questions

---

Sources: Z.ai GLM-5 repository · Du et al., GLM: General Language Model Pretraining with Autoregressive Blank Infilling, ACL 2022 · Zeng et al., GLM-130B: An Open Bilingual Pre-trained Model · Lasby et al., REAP the Experts: Why Pruning Prevails for One-Shot MoE Compression, ICLR 2026 · NIST CAISI Assessment of Z.ai's GLM-5.2, July 2026. Foundation benchmark scores are lab-reported and provisional; they describe the uncompressed GLM-5.2, not this compressed build. Pricing, context and capability figures are Shannon AI catalog values as of September 5, 2026.