Article cover: the title GLM-5.3 Flash: specs, benchmarks, and how to run it on Lyceum, labelled Model library, with a long purple-topped bar and a small outlined block, suggesting a long prompt reused with a small new part.

The low price turns the model choice into a quality test. In our example agent turn below, GLM-5.3 Flash costs about a sixth of the full GLM-5.3.

Is GLM-5.3 Flash good enough to replace a bigger model in your agent, and what will it cost to run? That question comes up as soon as an agent runs thousands of turns a day. Each turn resends the same history, so a small price gap grows fast.

What's new

Z.ai released GLM-5.3 Flash on 26 August 2026. Z.ai calls it the first model in the GLM-5 series that was trained on text and images from the start.

It is a mixture of experts (MoE). That means each token passes through only a small part of the network. GLM-5.3 Flash has 320 billion parameters, and 18 billion of them are active for each token.

Z.ai also changed how the model reads long prompts. Attention is the step where a model looks back over earlier tokens. GLM-5.3 Flash mixes two cheaper kinds of attention to cut the cost of serving long prompts. One kind looks back at only some earlier tokens, the other keeps a running summary of them.

Z.ai reports that GLM-5.3 Flash beats GLM-5.2 on all four of the benchmarks below. These are Z.ai's own numbers, from the chart on its model card. We have not re-run them.

Benchmark (what it tests)GLM-5.3 FlashGLM-5.2
Terminal-Bench 2.1 (tasks in a command-line terminal)84.381.0
DeepSWE v1.1 (fixing bugs in real code repositories)63.446.2
AutomationBench v1.0.6 (multi-step business workflows)48.826.2
Humanity's Last Exam with tools (hard expert questions)55.354.7

The biggest gains are on repository fixes and workflows. The scores come from Z.ai's own test setups, with reasoning at its maximum setting. They do not show how the model handles your prompts, so test it on a sample of your own tasks.

Specs and context window

The context window is how many tokens the model can read in one request: prompt, history and documents together. For GLM-5.3 Flash, it is 1M tokens.

SpecGLM-5.3 Flash on our API
Model IDz-ai/glm-5.3-flash
InputText and images
OutputText
Context window1M tokens
Longest answer per request131,072 tokens
ReasoningOn by default
Tool callingSupported
Parameters320B total, 18B active
LicenceMIT
Processing regionEU-hosted

Reasoning means the model writes out its working before it answers. That working, the reasoning trace, counts toward your output tokens. You can switch reasoning off with reasoning_effort set to "none", or by calling the Instant variant. Today that does not remove the working. The working moves into the answer text instead.

Two rows show one GLM-5.3 Flash response. Reasoning on, the default: the working comes back in a separate reasoning field and the answer text holds only the answer. Reasoning off, or the GLM-5.3 Flash Instant variant: the answer text starts with the working and a </think> marker, then the answer, so your code removes everything up to the marker. In both rows the working counts as output tokens. Behaviour as documented on our model pages and docs, read 23 October 2026.
With reasoning on, GLM-5.3 Flash returns its working in a separate field; with reasoning off or on the Instant variant, the working sits in the answer text up to a </think> marker (our docs, read 23 October 2026).

We also measure speed in our own status checks. Over the seven days before 23 October 2026, the median time to first token (TTFT) was 280 ms. TTFT is the wait before the first word of the answer arrives. Output then ran at 45 tokens per second, reasoning included. At that rate, a 1,000-token answer takes about 22 seconds. Current numbers are on the GLM-5.3 Flash model page, and incidents show on our status page.

Pricing

GLM-5.3 Flash is billed per token, with no base fee. Prices are in US dollars per million tokens, as listed on 23 October 2026.

Token typePrice per 1M tokens
Input$0.20
Cached input$0.05
Output$0.50

Cached input is text that matches the start of an earlier request. We reuse the work already done on it, a feature called prompt caching. That part bills at the cached rate listed on the model page. It is automatic but best effort. A request that misses pays the full input price.

Here is what that means for one agent turn. The token counts are a hypothetical example. The turn resends 40,000 tokens of system prompt, tool definitions and history, adds 2,000 new tokens and writes 1,000 tokens of output.

Cache hit:  40,000 x $0.05/1M + 2,000 x $0.20/1M + 1,000 x $0.50/1M = $0.0029
Cache miss: 42,000 x $0.20/1M + 1,000 x $0.50/1M                    = $0.0089
Same turnCost per turnCost per 1,000 turns
GLM-5.3 Flash, history cached$0.0029$2.90
GLM-5.3 Flash, no cache$0.0089$8.90
GLM-5.3, history cached$0.0176$17.60

A miss costs about three times as much as a hit. The full GLM-5.3 costs about six times as much, even with its cache. The example assumes 1,000 output tokens. A turn with long reasoning writes more and costs more, and the example says nothing about answer quality. Our post on cache-aware routing explains how we send repeat turns to the copy of the model that holds their history.

How to call it

Stacked bars for one GLM-5.3 Flash agent turn: 40,000 tokens of history resent, 2,000 new input tokens and 1,000 output tokens. No cache: 42,000 input tokens at $0.20 per 1M is $0.0084, plus $0.0005 output, $0.0089 in total. History cached: $0.0020 cached input, $0.0004 new input and $0.0005 output, $0.0029 in total. Calculated example, prices read 23 October 2026.
In our calculated example, one GLM-5.3 Flash agent turn costs $0.0029 with its history cached and $0.0089 without, at our prices read 23 October 2026.

Lyceum is a sovereign European AI inference provider for open-source models. The GLM-5.3 Flash model page lists its processing region as EU-hosted. We don't keep your prompts or outputs, and we never use them for training.

Running GLM-5.3 Flash yourself takes about four GPUs, and our API removes that step. The recipe from vLLM, an open-source serving engine, puts its weights at about 306 GiB of GPU memory, before any cache.

You call it with the OpenAI SDK. Set LYCEUM_API_KEY to your API key, then point the client at our base URL. The example sends a screenshot and a question, then prints how many input tokens came from the cache.

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["LYCEUM_API_KEY"],
    base_url="https://api.lyceum.technology/openai/v1",
)

response = client.chat.completions.create(
    model="z-ai/glm-5.3-flash",
    messages=[{
        "role": "user",
        "content": [
            {"type": "text", "text": "What does this error message say?"},
            {"type": "image_url", "image_url": {"url": "https://example.com/error.png"}},
        ],
    }],
    max_tokens=8000,
)

print(response.choices[0].message.content)
details = response.usage.prompt_tokens_details
print(details.cached_tokens if details else 0)

Three settings matter in practice:

  • max_tokens: leave room for the working and the answer together. A small limit can end the request before the answer starts.
  • Images: send them as image_url parts, from a public URL or a base64 data URL. Leave reasoning on for image requests. With reasoning off, the model writes its working into the answer before the answer itself.
  • Tools: pass tools in the standard OpenAI format. Parallel calls work, and so does tool_choice set to "required".

The usage field for cached tokens is missing when nothing came from the cache, so read it with a fallback, as above. The docs list every request field and limit.

Running GLM-5.3 Flash in Claude Code

We also serve an Anthropic-compatible endpoint, so Claude Code can run on GLM-5.3 Flash without an adapter. The easiest way is our CLI. lyceum code starts Claude Code in its own profile, so an existing Anthropic login stays untouched.

pip install lyceum-cli
lyceum auth login --api-key lk_<your_api_key>   # one-time
lyceum code

To try it in one terminal without the CLI, set two variables and start Claude Code. Then switch to GLM-5.3 Flash inside the session. The switch applies from your next message, and the conversation stays.

export ANTHROPIC_BASE_URL=https://api.lyceum.technology/anthropic
export ANTHROPIC_AUTH_TOKEN=lk_<your_api_key>
claude

# inside Claude Code:
/model z-ai/glm-5.3-flash

Use ANTHROPIC_AUTH_TOKEN, not ANTHROPIC_API_KEY. With ANTHROPIC_API_KEY, Claude Code asks you to approve the key on first run. In non-interactive use, such as claude -p or CI, that check fails with "Not logged in". Our Claude Code docs also cover a lasting setup in settings.json and the VS Code extension.

Where it fits in the Lyceum catalogue

GLM-5.3 Flash sits at the low-cost end of our text models. The table compares it with its closest neighbours on our models page, as listed on 23 October 2026.

ModelInput / cached / output per 1M tokensContext window
GLM-5.3 Flash$0.20 / $0.05 / $0.501M
GLM-5.3 Flash Instantsame as GLM-5.3 Flash1M
GLM-5.3$1.40 / $0.26 / $4.401M
Qwen3.8 Flash Nextsame as GLM-5.3 Flash256K
DeepSeek V4 Flash$0.25 / $0.06 / $0.301M

GLM-5.3 Flash Instant is the same model without a separate reasoning phase, at the same price. Its answers still begin with a short line of working that ends in a </think> marker. That marker closes the working. Remove everything up to the marker, or use GLM-5.3 Flash with reasoning on. We are fixing it.

GLM-5.3 is the larger model, with 40 billion active parameters per token. Qwen3.8 Flash Next costs the same as GLM-5.3 Flash but reads at most 256K tokens per request.

For high-volume agent and coding work, GLM-5.3 Flash is the GLM model we would test first. Move a task to GLM-5.3 when your own tests show Flash gets it wrong too often. To start, create an API key and run the example above on Serverless Inference.

FAQ

Sources

  1. huggingface.co
  2. github.com
  3. recipes.vllm.ai