GLM-5.3-Flash Pricing: What It Costs Before and After the Launch Discount

GLM-5.3-Flash is currently half price. That ends on September 9, 2026, and every cost projection based on today’s rates doubles after that date.

Try Apidog today

This guide covers current and post-promotion pricing, practical workload costs, cache and reasoning controls, reseller differences, and when self-hosting makes sense. Figures were verified on August 27, 2026; confirm provider pricing before committing a budget.

Rate card

	Launch price (through Sept 9, 2026)	List price (after)
Input	$0.075 / 1M tokens	$0.15 / 1M tokens
Output	$0.25 / 1M tokens	$0.50 / 1M tokens
Cached input	$0.015 / 1M tokens	$0.03 / 1M tokens

Artificial Analysis reports a blended cost of $0.10 per million tokens, based on a 7:2:1 cached-input, input, output ratio. Use it for comparisons, but budget from your own token mix.

Practical monthly costs

Use list price for planning after September 9.

Support classifier: 100,000 tickets/month at roughly 800 input and 100 output tokens each:

• 80M input tokens: $12.00
• 10M output tokens: $5.00
• Total: $17/month

Set reasoning_effort to low to reduce output costs further.

Coding assistant: 500 sessions/day at 15,000 input tokens and 2,000 output tokens per session:

• 225M monthly input tokens: $33.75
• 30M monthly output tokens: $15.00
• Without caching: $48.75/month
• With 70% cached input: about $30/month

Document pipeline with images: 10,000 documents/month at 40,000 input tokens and 1,500 output tokens each:

• 400M input tokens: $60.00
• 15M output tokens: $7.50
• Total: $67.50/month

These numbers are low by design: GLM-5.3-Flash targets inexpensive high-volume workloads.

Maximize cached input

Cached input costs $0.03 per million tokens, versus $0.15 for fresh input: one fifth the cost.

Keep stable prompt content first:

1. System prompt
2. Shared instructions
3. Repeated document or code context
4. Variable user input last

Do not put timestamps, request IDs, or user names near the start of the prompt. Any variation can invalidate the cache for everything after it.

Z.ai has listed cached-input storage as free for a limited period. Treat that as promotional and verify current terms before relying on it.

Set reasoning effort deliberately

reasoning_effort defaults to max, the most expensive setting.

Available modes:

• low
• high
• max

Reasoning tokens are billed as output tokens. For classification, extraction, routing, and formatting, use low unless evaluation proves a higher setting is necessary.

See the setup in our API guide.

How GLM-5.3-Flash compares

Model	Blended per 1M
GLM-5.3-Flash	$0.10
GLM-5.3	$0.90

That is roughly a nine-fold price gap for a three-point difference on the Artificial Analysis Intelligence Index: 57 versus 60.

Choose GLM-5.3 instead when long streamed responses are waiting on a human; it generates almost twice as fast. See the full comparison.

Against GLM-5.2, Z.ai claims GLM-5.3-Flash is roughly one tenth the price, with a cost-per-task figure of $0.045. See the GLM-5.2 pricing breakdown when budgeting a migration.

GLM-5.3-Flash is also competitive with cheap multimodal models from other vendors and is unusual in being MIT-licensed, leaving self-hosting open. For the closest Google comparison, see Gemini 3.7 Flash pricing explained.

Compare reseller rates

GLM-5.3-Flash is available through Z.ai, OpenRouter, Cloudflare Workers AI, Vercel AI Gateway, DeepInfra, Novita, GMICloud, Baseten, and io.net.

Rates vary. AIHubMix, for example, has listed approximately:

• $0.113 / 1M input tokens
• $0.394 / 1M output tokens

That falls between Z.ai’s promotional and list prices. Check the price at the provider you call; aggregators may pass through the discount or add a margin.

When self-hosting wins

The weights are MIT-licensed, but low API pricing makes self-hosting difficult to justify on cost alone.

A full-precision deployment needs roughly an 8x H200-class node, costing around $24–$48/day, or approximately $1,000/month. At list API pricing, $1,000 buys roughly 6.7 billion input tokens.

Self-host when you need:

• Data residency
• Infrastructure control
• Licensing flexibility
• Very large, sustained volume

The main cost exception is a quantized model running on hardware you already own, where marginal cost is largely electricity. See the local-run guide for hardware-tier guidance.

Consider the GLM Coding Plan

For developers using Claude Code or Cline rather than calling an API directly, a flat subscription can be a better fit.

The GLM Coding Plan starts around $18/month, includes GLM-5.3-Flash, and reportedly provides three times the usable quota of GLM-5.3. Z.ai’s documentation also states that off-peak calls consume half the standard points.

For a developer coding daily, a flat plan is usually cheaper and more predictable than metered API access. See the harness setup guide and coding-plan comparison.

Watch output costs

Output costs more than three times as much as input: $0.50 versus $0.15 per million tokens.

Reasoning tokens are also billed as output. With reasoning_effort: max, the model can generate several times more reasoning tokens than visible answer tokens.

Measure actual usage instead of estimating:

1. Log the usage object from every completion.
2. Aggregate input, output, and cached-token counts.
3. Compare your actual mix with the 7:2:1 blended-rate assumption.
4. If output exceeds roughly 15% of total tokens, reduce reasoning_effort first and prompt length second.

Before September 9

• Measure token mix. Output-heavy workloads cost closer to the $0.50 output rate than the $0.10 blended figure.
• Check cache hit rate. Cached input is five times cheaper than fresh input.
• Recalculate at list price. Costs double on September 10. Solve any pricing problem before the promotion ends.

Each completion’s usage object includes the input, output, and cached-token counts you need. Run representative requests as a saved collection in Apidog, store the model ID in an environment variable, and rerun the same suite against GLM-5.3 to compare actual bills rather than marketing claims.

FAQ

Is GLM-5.3-Flash free? No. It was free for about a week while running anonymously as “ox-alpha” before launch, but that ended at announcement. The current 50% discount is not free access.

When does the discount end? September 9, 2026. List pricing applies afterward.

Why do sites quote different prices? Some show promotional rates, some show list rates, and resellers set their own margins. Check the provider you actually call.

Does image input cost extra? Images consume context tokens and are billed as input. There is no separate image surcharge, but high-resolution images use meaningful token volume.

Is self-hosting cheaper? Usually only at very large sustained volume, or when you run a quantized model on hardware you already own. At $0.15 per million input tokens, renting an 8x H200 node is difficult to justify on cost alone.