GLM-5.3-FLASH

Anatomy of $0.15

chapter 6 of 8 · about 7 minutes

This chapter turns the architecture into a bill. GLM stands for General Language Model. An application programming interface (API) charges separately for input tokens, which it reads, and output tokens, which it generates. When a repeated prompt prefix is reused from memory, the provider can charge a lower cache-read rate. Prices here are quoted per million tokens, a billing unit rather than a required request size. The list price is $0.15 input / $0.50 output per million tokens, with cache reads at $0.03. At launch Z.ai ran a 50% discount ($0.075/$0.25, promotion ended Sep 9, 2026). GPT, where it appears in the table, means Generative Pre-trained Transformer. The comparison uses live OpenRouter data:

per 1M tokens          input      output     cache-read
GLM-5.3-Flash          $0.15      $0.50      $0.03     this article's subject
DeepSeek V4 Flash      $0.089     $0.18      $0.018    the one cheaper rival
Kimi K2.5              $0.45      $2.25      $0.07
Gemini 3.7 Flash       $0.75      $3.75      $0.075
GLM-5.3 (full)         $1.40      $4.40      $0.26     9.3× / 8.8× Flash
Claude Sonnet 5        $2.00      $10.00     $0.20
GPT-5.2                $1.75      $14.00     $0.175
Claude Opus 5          $5.00      $25.00     $0.50
GPT-5.5                $5.00      $30.00     $0.50

Which mechanisms lower the bill?

The price reflects several separate costs. Mixture-of-Experts (MoE) activates a small subset of specialist subnetworks for each token instead of all model parameters. Kimi Delta Attention (KDA) carries a fixed state instead of rereading all earlier tokens. Sparse multi-head latent attention (MLA) retains compressed token records for precise lookup. 8-bit floating point (FP8) stores each number in one byte, and multi-token prediction (MTP) lets a verified pass deliver more than one output token when the draft is accepted. In the diagram, B means billion model parameters. The bars show an illustrative blended price under a 3:1 input-to-output mix, not a measurement of one provider's invoice. Click one lever at a time in the builder below to see what each published factor changes.

blended $/M tokens, one discount per step (log-ish scale, illustrative)$12.00closed frontier (blended)$2.50MoE: 18B of 320B awake$1.1034 KDA layers, no lookback$0.50sparse MLA: read 0.2%$0.15FP8 + MTP + cacheeach step is a published mechanism with receipts in the builder below

Build a price estimate

Click a factor to include it. Watch the estimated price change.

blended reference price:$10.00$10.00per M tokens
Show sources for each factor
  • Activate 18B of 320B parameters: GLM-5.3-Flash has 320B total parameters and 18B active per token (5.6%). A dense 320B model uses ≈640 GFLOP per token, compared with ≈36 for Flash. The same pattern lets GLM-5 use 744B/40B (report §2.1).
  • Reduce attention work: The design combines a lightning indexer, top-2,048 sparse attention, and 34 KDA linear layers. Z.ai measured 3.01× less attention compute than GLM-5.3; the GLM-5 report measured DSA at 1.5–2× with ~90% of entries redundant (§2.1.1).
  • Use FP8 to move fewer bytes: The native FP8 e4m3 checkpoint is 331 GB versus 640 GB for BF16. Decode is memory-bandwidth-bound, so half the bytes can increase throughput or concurrency. vLLM's FP8 KV pool fits 14.92M tokens versus 8.87M in BF16 (1.68×).
  • Accept ~2.7 tokens per forward pass, MTP: Multi-token prediction (MTP): GLM-5 measured a 2.76-token acceptance length versus DeepSeek-V3.2's 2.55 (report Table 2). Flash includes an in-weights draft layer, and vLLM runs it with 5 speculative tokens. This estimate uses 1.6× for verification overhead.
  • Lower deployment cost and self-hosting: Report §5 estimates a single Chinese node at about a dual-GPU international cluster, with 50% lower long-sequence deployment cost. The MIT license allows self-hosting, which sets a public lower price bound for APIs. The report says DeepSeek V4 Flash is cheaper.

What does repeated context cost?

An agent is software that asks a model to plan, use tools, and continue a multi-step task. Its shared context can include system instructions, a codebase, and tool results. Later requests often reuse that same prefix. When a large prefix is read from cache hundreds of times, the cache-read price can dominate the total bill even when the input and output rates look small. The worked example in the research receipts uses an agentic coding session with 1 million input tokens, 80% cache hits, and 200,000 output tokens:

modelsession costvs Flash promo
GLM-5.3-Flash (promo)$0.077
GLM-5.3-Flash (list)$0.154
DeepSeek V4 Flash$0.0670.9×
Kimi K2.5$0.5967.7×
Gemini 3.7 Flash$0.9612.5×
GPT-5.4-mini$1.1114.4×
GLM-5.3 (full)$1.36817.8×
Claude Sonnet 5$2.5633×
GPT-5.5$7.4096×
Claude Opus 5$6.4083×

Same traffic. 83–96× apart at the frontier tier. At 10,000 sessions/day that's $770 vs $64,000 per day.

Two structural conditions sit outside the waterfall. First, the Massachusetts Institute of Technology (MIT) license lets anyone download the weights and estimate a self-hosted cost. That does not set an API price by itself, because serving also requires hardware, operations, and capacity. It does give customers a visible comparison point instead of a price set only by one provider. Second, DeepSeek V4 Flash lists a lower price than Flash. The published mechanisms explain part of Flash's cost structure, while competition and the cost of serving open weights shape the price a provider can charge.

receipts, every claim in this chapter, checked 2026-08-31