GLM-5.3 API
Choose GLM-5.3
Z.ai's flagship reasoning model for coding agents and long-horizon engineering work, with a 1M-token context, always-on reasoning, tool calling, and prompt caching over both Chat Completions and Anthropic Messages.
GLM-5.3
Z.ai flagship reasoning model
glm-5.3Repository-scale coding agents, long-running task chains, tool-heavy workflows, and stable-prefix agent loops where cache reads carry most of the input volume.
GLM-5.3 pricing
Estimate a request before you route traffic. Every user group uses the same GLM-5.3 rate, and output tokens already include reasoning tokens.
Token calculator
Enter the token mix for one request.Estimated request cost
GLM-5.3Minimum charge: 0.01 credits per request. Output tokens already include reasoning tokens.
Budget guide
Approximate requests using the current mix.For quick testing
For regular development
For production evaluation
Model pricing
| Model | Context | Input tokens | Cache read tokens | Output tokens |
|---|---|---|---|---|
GLM-5.3glm-5.3 | All context sizes | $1.400 / 1M95.2 cr / 1M | $0.261 / 1M17.7 cr / 1M | $4.400 / 1M299.2 cr / 1M |
GLM-5.3
All context sizesUSD and credits are shown per 1M tokens, flat across the full 1M context.
Web search is billed separately at 0.68 credits per call.
GLM-5.3 API for coding agents and long-horizon engineering
Call Z.ai's flagship reasoning model for $1.40 input and $4.40 output per 1M tokens through EvoLink's unified API. Use model ID glm-5.3 with a 1M-token context, always-on low/high/max reasoning, tool calling, and prompt caching.
Where GLM-5.3 earns a place in a production model stack
GLM-5.3 costs more than GLM-5.2 and much more than Flash, so it should not become the default route just because it is newest. Its strongest fit is work where deeper reasoning removes a retry or a round of human correction.
GLM-5.3 for repository-scale coding
Hold connected modules, tests, and conventions in one 1M-token context instead of feeding a coding agent one file at a time. The gain shows up as fewer broken cross-file edits, not as a better single-file completion.
GLM-5.3 for long task chains
Plan-execute-verify loops that run for many steps benefit most, because a single early reasoning error compounds across the whole chain. Measure completed-chain rate rather than per-step quality.
GLM-5.3 for tool-heavy agents
Large tool schemas and long instruction blocks sit in the cached prefix, so the marginal cost of another step stays close to the cache-read rate rather than the full input rate.
GLM-5.3 through Anthropic-protocol clients
Requests to /v1/messages are passed through byte-for-byte, so prompt-cache prefixes and reasoning signatures survive. Claude-protocol coding CLIs can point at EvoLink and select glm-5.3 without a translation layer.
What GLM-5.3 changes — and what it breaks
GLM-5.3 is not a drop-in replacement for GLM-5.2. One parameter change is a hard break, and the pricing moved. Read these four before you switch a production route.
Breaking: reasoning can no longer be disabled
GLM-5.3 always reasons and rejects thinking.type: "disabled". Any client migrated from GLM-5.2 that still sends the disabled mode will fail outright. The official replacement is thinking.type: "enabled" with reasoning_effort: "low".
Three reasoning effort levels, defaulting to max
reasoning_effort accepts low, high, and max, and defaults to max. Because output tokens include reasoning tokens, the level moves cost directly — leaving the default in place on high-volume routine calls is the most common way to overspend on 5.3.
Cache reads are the cheapest lever on the page
Cache-read tokens bill at roughly a fifth of fresh input. In a long agent loop where the system prompt and tool schemas dominate the input, keeping that prefix byte-stable matters more to the bill than the model choice itself.
Priced above GLM-5.2, with no discount
Input runs about 40% higher than the current GLM-5.2 rate because 5.3 is billed at the upstream rate with no promotional discount. Route to 5.3 where the extra reasoning is measurably worth it, and leave the rest on 5.2 or Flash.
Why GLM-5.3 can handle these workloads
Three properties do most of the work. Each one also implies a way to use the model badly, so they are worth understanding before you route production traffic.
A 1M-token working context at one flat rate
There is no long-context price tier — the rate at token 900,000 equals the rate at token 1,000. That removes a common reason to truncate aggressively, but it does not make filling the window free or accurate.
Byte-faithful request passthrough
Apart from the model name, the request body reaches the upstream unchanged, so GLM-specific fields, prompt-cache prefix hashes, and reasoning signatures are preserved on both endpoints.
Reasoning tokens included in output billing
Reasoning is billed inside completion tokens rather than as a separate line, so a 131,072-token output ceiling covers reasoning and final answer together. Set output budgets by task, not by the ceiling.
What to verify before routing production traffic to GLM-5.3
Four checks catch nearly every migration problem we see on this model. The first two are hard failures; the last two are silent cost problems.
Use glm-5.3 as the model ID
The same ID works on Chat Completions and Anthropic Messages. It matches the upstream name exactly, so there is no alias to translate.
Remove any thinking.type: "disabled"
This is the one hard break from GLM-5.2. Replace it with thinking.type: "enabled" plus reasoning_effort: "low" and re-run your regression prompts.
Set reasoning_effort deliberately
The default is max. Drop routine and latency-sensitive calls to low and confirm the accepted-result rate holds before you keep the saving.
Confirm cache reads are actually hitting
Check that cached tokens appear in the returned usage. If the prefix shifts between calls, every request bills at full input rate and the change is invisible in the response itself.
Compare cost per accepted task, not token price alone
GLM-5.3 is more expensive per token than both GLM-5.2 and Flash. That only matters alongside how often each model produces a result you actually ship. Track these signals on your own workloads for a week before committing a route.
A route that costs 40% more per token but removes one retry in three is cheaper in practice. A route that costs 40% more and changes nothing is just a more expensive default. The only way to tell them apart is to measure both on the same tasks.
Compare against GLM-5.2 and Kimi K3 before switching
EvoLinkGLM-5.3 is priced above GLM-5.2, so the question is whether fewer retries and less human correction cover the difference on your own workloads. Measure first, then choose the production route.
| Model | GLM-5.3 | GLM-5.2 | Kimi K3 |
|---|---|---|---|
| Input / output | $1.4 / $4.4 | $1 / $3.5 | $3 / $15 |
| Context | 1M | 1M | 1.05M |
| Caching | Cache reads | Cache reads | Automatic cache reads |
| Best for | Repository-scale coding agents, long-running task chains, tool-heavy workflows, and stable-prefix agent loops where cache reads carry most of the input volume. | The previous GLM flagship, still cheaper per token. Keep it for workloads where 5.3 shows no measurable gain, and for clients that rely on disabling reasoning. | Moonshot's 1M-context reasoning route. A useful cross-vendor baseline for long-context coding and agent work. |
Related models
GLM-5.2
The previous GLM flagship, still cheaper per token. Keep it for workloads where 5.3 shows no measurable gain, and for clients that rely on disabling reasoning.
View model
Kimi K3
Moonshot's 1M-context reasoning route. A useful cross-vendor baseline for long-context coding and agent work.
View model
GLM-5.3 Flash
Same 5.3 generation at roughly a ninth of the price, plus native image, video, and file input. The right default for high-volume and multimodal work.
View model
DeepSeek V4 Pro
Cost-sensitive long-context baseline for bulk reasoning where per-token price dominates the routing decision.
View modelRelated reading
GLM-5.3 Flash vs GLM-5.3
Choose the family route by modality, task difficulty, and cost per accepted result — with a Flash-first escalation policy.
Read guide
GLM-5.3 Release: What Changed
What the August 14 release confirmed, what the live API now costs, and which migration checks matter before production traffic.
Read guide
GLM-5.3 vs GLM-5.2
Same base model, all gains from post-training — plus the breaking change that reasoning can no longer be disabled. When to stay on 5.2.
Read guide
GLM-5.3 vs Claude
Benchmarks, API contracts, and access reality compared before you route coding agents to either one.
Read guide
GLM-5.3 Cybersecurity Benchmarks
What the CyberGym and ExploitBench numbers claim, by Z.ai's own reporting, and how defenders should read them.
Read guide
One Gateway for 3 Coding CLIs
Config paths, environment variables, and a troubleshooting checklist for running coding CLIs through a single endpoint.
Read guideGLM-5.3 API FAQ
Is the GLM-5.3 API available through EvoLink?
Yes. GLM-5.3 is available as a production model, served over both Chat Completions and Anthropic Messages.
What model ID should I use for the GLM-5.3 API?
Use glm-5.3 for both Chat Completions and Anthropic Messages. The EvoLink model name matches the upstream name exactly.
Can I keep using the OpenAI SDK or Anthropic Messages?
Yes. Point either client at EvoLink with the same API key and select glm-5.3. Requests are passed through byte-for-byte apart from the model name, so prompt-cache prefixes and reasoning signatures survive.
Can I still disable reasoning like on GLM-5.2?
No. GLM-5.3 always reasons and rejects thinking.type: "disabled". If you are migrating from GLM-5.2, switch to thinking.type: "enabled" with reasoning_effort: "low" — that is the official replacement for the disabled mode.
How should I choose a reasoning_effort level?
The default is max. Use low for routine edits, classification, and latency-sensitive calls; high or max for architecture, debugging, and long task chains. Output tokens include reasoning tokens, so effort directly moves cost.
Why is GLM-5.3 more expensive than GLM-5.2?
GLM-5.3 is priced at the upstream rate without a discount, while GLM-5.2 carries a promotional rate. Input runs about 40% higher. Route to 5.3 where the extra reasoning removes retries, and keep 5.2 or GLM-5.3 Flash where it does not.
How is prompt caching billed?
Cache reads have their own rate, roughly a fifth of fresh input. There is no cache-write charge because the upstream does not report cache-creation tokens. Keep the prompt prefix byte-stable so it keeps hitting.
What is the context window and output limit?
1,000,000 tokens of context and up to 131,072 output tokens, at one flat rate across the whole window — there is no long-context price tier.
Should I use GLM-5.3 or GLM-5.3 Flash?
Flash costs roughly a ninth as much and accepts image, video, and file input. Start on Flash for high-volume, multimodal, and routine work; move to GLM-5.3 when a task needs deeper reasoning and the accepted-result rate justifies it.
Does GLM-5.3 accept images?
No. GLM-5.3 is text-only. Use GLM-5.3 Flash for image, video, and file input.
How should I compare GLM-5.3 with Claude, GPT, or Kimi?
Evaluate GLM-5.3 on cost per accepted task for coding agents and long task chains, treat Claude and GPT as frontier capability baselines, and use Kimi K3 as the cross-vendor long-context comparison.
What should a production GLM-5.3 evaluation measure?
Track first-pass success, accepted deliverables, retries, reasoning-token share of output, cache-hit ratio, valid tool calls, time to accepted result, and fallback rate.