Free GLM-5.3-Flash API: Instant Signup, No Credit Card

Sign up for AnRouter and you can call GLM-5.3-Flash right away. No credit card, and your account starts with $1 of model credit.

curl https://api.anrouter.com/v1/chat/completions \
  -H "Authorization: Bearer $ANROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"glm-5.3-flash","messages":[{"role":"user","content":"hi"}]}'

What $1 gets you

On the Default plan, GLM-5.3-Flash costs $0.045 for input and $0.16 for output, per million tokens. Converted into real work, at the token counts shown:

Thirty-eight agent sessions let you run the same real task a dozen times over, varying the prompt and the reasoning budget, and watch how steady it actually is on your own codebase — which is exactly what a benchmark cannot tell you.

Prompt caching stretches that further. When you iterate on the same code, the repeated portion bills at the cache rate of $0.012, which is 73% below the input rate. On that same agent session, a 90% cache hit rate drops the cost from $0.0257 to $0.0109, so $1 covers 92 runs.

Against the Z.AI list price ($0.15 input / $0.50 output), the same agent session costs $0.085. For the same dollar: 11 runs there, 38 runs here.

Pricing and the three plans

Per million tokens, in US dollars:

All three plans serve the same models with the same features, give you every model on the platform, and set no plan-level token cap:

• Default — $0/month, usable the moment you sign up, billed by usage
• Pro — $5.99/month, 20% off input and output
• Max — $39.99/month, 50% off input and output

If you plan to run agents seriously, going straight to Pro is the better deal. Do the math: Pro costs $5.99 a month, and at 20% off you need $29.95 of monthly usage at Default rates to earn that back — which, at $0.0257 per agent session, means 39 sessions a day. Anyone who uses a coding agent daily runs well past 39.

The Max break-even sits at roughly 104 sessions a day, which suits high-frequency production traffic and automated workflows. If you are genuinely just trying things out, stay on Default — a zero monthly fee is the lowest possible cost.

GLM-5.3-Flash or DeepSeek

On AnRouter the two share one key and one base URL. Switching means changing a single model line, so you do not have to commit up front.

On paper they are nearly the same model: glm-5.3-flash and deepseek-v4.1-flash-1017 both have a context window of about one million tokens and a maximum output of 131,072 tokens, and both support streaming, vision, tool calls, structured output, and prompt caching.

The difference is mostly price:

For the same context window and the same output ceiling, GLM-5.3-Flash is 44% cheaper on input and 47% cheaper on output. So start there, and switch to DeepSeek for comparison if a specific task disappoints you — it is one line either way.

Specs, and one habit that trips people up

GLM-5.3-Flash is a natively multimodal mixture-of-experts model from Z.AI, 320B total parameters with roughly 18B active. What you actually get on AnRouter:

• 1,048,576-token context window (about one million)
• 131,072-token maximum output per request
• Text and image input, text-only output
• Streaming, vision, reasoning, tool calls, function calling, structured output, JSON mode, and prompt caching
• All three endpoints work: /v1/chat/completions, /v1/responses, /v1/messages

It likes to stop early, so say so when you want long output. A single reply can run to 131,072 tokens, but GLM-5.3-Flash tends to wrap up early: asked to rewrite a file several thousand lines long, it may stop around 8,000 tokens and say "I must stop here". That is the model's habit, not an API limit. Add one line to the system prompt (this reply may be very long; do not stop early) and in our tests it writes all the way to 131,072.

Connecting your client

AnRouter speaks three protocols, and which one you configure depends on which one your client speaks. This is not interchangeable — Claude Code only ever sends Anthropic-format requests, so /v1/chat/completions is of no use to it.

The base URL on the third row carries no /v1, and it should not be the full /v1/messages either — the Claude-family SDKs append that themselves.

Full configuration for each client:

• Codex CLI — ~/.codex/config.toml needs wire_api = "responses"; see the setup guide
• Claude Code — use ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN; see the setup guide
• OpenCode — add a custom provider in settings; see the setup guide
• Hermes Desktop — point it at a custom endpoint; see the setup guide

Common errors

401 invalid_api_key — two causes account for most of these: the key picked up a space or newline when it was copied, or the client adds the Bearer prefix itself and you wrote it in by hand as well, producing Bearer Bearer sk-....

404 Invalid URL (POST /v1/v1/messages) — a Claude-family client has its base URL set to https://api.anrouter.com/v1. The SDK appends /v1/messages, which yields /v1/v1. Drop the trailing /v1.

Model not found — the model name is exactly glm-5.3-flash, with no -free or similar suffix. Some relay sites use IDs like glm-5.3-flash-free; that is their own naming, not ours.

Image requests failing — GLM-5.3-Flash accepts image input, but not every model does. Confirm before you send an image after switching model.

After switching models, retest tool calls and similar features: a route existing does not mean every model supports that capability. The model, upstream channel, and request fields must all be compatible.