ModelsBenchmarksCoding AgentsAPI PricingGLM

GLM-5.3-FlashX: 200 Tokens/sec Throughput vs 2.5x Price Analysis

Review GLM-5.3-FlashX provider speed claims, current API rates, task-cost math, and how to measure whether faster serving helps your coding workflow.

Published: Last checked:

GLM-5.3-FlashX: 200 Tokens/sec Throughput vs 2.5x Price Analysis
Analyzed Foundation Models (3)All Models Directory→

GLM-5.3-FlashX is Z.ai’s faster serving option for GLM-5.3-Flash. Z.ai advertises inference speed of 200 tokens per second. The practical question is whether that speed shortens your complete coding workflow enough to justify a higher API bill. This review uses provider documentation and a transparent cost example; it is not a RankLLMs latency test.

What FlashX is and what is documented

Z.ai lists the model codes glm-5.3-flash and glm-5.3-flashx in its model documentation. Both are described as part of the GLM-5.3-Flash family, with a 1M-token context window, up to 128K output tokens, and text output from text, image, video, and file inputs. The page describes function calling, structured output, caching, and thinking. Thinking can be enabled but not disabled through the documented parameter.

Z.ai describes FlashX as delivering inference speeds of 200 tokens per second. This is a provider claim, not a guarantee for every request. The documentation does not establish a universal time-to-first-token, throughput under concurrency, or end-to-end agent completion time. FlashX is also not included in the GLM Coding Plan according to the same page, so plan access should not be assumed from standard Flash availability.

Current native API pricing

The Z.ai pricing page lists these USD rates per million tokens, checked October 9, 2026:

Token categoryGLM-5.3-FlashGLM-5.3-FlashXFlashX premium
Uncached input$0.15$0.37About 2.47x
Cached input$0.03$0.0752.5x
Output$0.50$1.252.5x

The prices are for Z.ai’s native API rate card. A subscription, reseller, or gateway may use a different rate or display promotional pricing. Check the route and invoice associated with your key before estimating costs. Promotional and third-party rates can differ from this native API snapshot; compare the endpoint shown in your account with the provider rate card.

Cost for a representative task

Suppose a coding task uses 50,000 uncached input tokens and generates 10,000 output tokens, with no cache hits. At the listed rates:

  • Flash: 0.05 x $0.15 + 0.01 x $0.50 = $0.0125.
  • FlashX: 0.05 x $0.37 + 0.01 x $1.25 = $0.031.
  • Difference: $0.0185 per task, or $18.50 for 1,000 tasks of this exact size.

This is token cost only. It excludes tool charges, retries, and larger or smaller context on real agent runs. If 80% of the input is cached, recalculate using the cached rates. Read usage fields from the provider after a run instead of assuming that a repeated prompt prefix was fully cached.

Speed is not total task time

Token generation speed describes only one part of a request. A coding agent also spends time waiting for the first token, reading files, invoking tools, running tests, and waiting for network services. If the agent emits a long patch, generation speed may matter more than it does for a short classification. If a tool call or test suite dominates, a faster model decoder may barely change the total.

A simple estimate helps frame the upper bound: generating 2,000 output tokens at a sustained 200 tokens per second would take 10 seconds of decoding. That arithmetic is not an end-to-end latency estimate. It excludes prompt processing, time-to-first-token, tool cycles, network delay, and the possibility that your request does not reach the advertised rate.

Measure first useful output, decode time, tool wait, and finished-task time separately; a single throughput figure cannot characterize an entire agent loop.

Testing FlashX in a coding workflow

For an IDE or CLI agent, run the same bounded tasks against both model codes from the same repository commit. Use equal instructions, tool permissions, reasoning settings, context, and spending caps. A useful task set includes a small bug fix, a multi-file change, and a test failure that requires reading logs. Record:

  • Whether the final patch passes the existing tests.
  • Time to first useful output and time to a validated result.
  • Input, cache-hit, and output token counts.
  • Retries, unrelated changes, and manual corrections.
  • Total cost per accepted task, including tools and failed attempts.

Repeat tasks so a single quick response does not define the result. If you use VS Code, OpenCode, or another client, first check its current Z.ai provider configuration and confirm that streaming, function calls, and thinking parameters work on the endpoint you selected. A model name shown in a picker does not by itself confirm which provider or price is being used.

Where the premium may make sense

FlashX is worth testing for interactive work where people wait on multiple model responses: short turnarounds, coding-agent loops, and multimodal tasks that use screenshots or files. The value depends on measured time saved and equal task quality. For overnight batch processing, background classification, or jobs dominated by browser and test execution, standard Flash’s lower rates may matter more than generation speed.

A faster model can also cause an agent to run more iterations in a fixed period. Monitor total token use rather than assuming higher speed lowers the bill. The right comparison is cost and wall-clock time per completed, reviewed task.

Verdict

Z.ai documents FlashX as the 200 tokens-per-second option and currently lists it at roughly 2.5 times standard Flash’s rates across uncached input, cached input, and output. The documentation does not show that FlashX is more capable, that every request reaches 200 tokens per second, or that the higher price reduces cost per task. Test both endpoints on your actual agent loop and keep FlashX only if the saved time is worth the measured premium.

See the GLM-5.3-Flash profile, GLM-5.3 profile, and coding model guide.

Frequently Asked Questions

How fast is GLM-5.3-FlashX?

Z.ai advertises 200 tokens per second for FlashX. This is a provider specification, not a guarantee for every request or a RankLLMs measurement.

What is the price of GLM-5.3-FlashX compared to standard Flash?

Z.ai currently lists FlashX at $0.37 uncached input, $0.075 cached input, and $1.25 output per million tokens; standard Flash is $0.15, $0.03, and $0.50.

Sources & disclosures

Source links for data and claims discussed in this article are listed below. Refer to each source for its current figures and methodology.

Share Article

Recommended Reading

Related by topic and recency.

All articles→
DealsFree Offers

TypeSafe AI Jev: Access, Pricing, and Structured Decisions

Jev is TypeSafe AI's structured decision model. See its current Vercel AI Gateway listing, published input price, suitable tasks, and limits.

Sep 20, 2026 · 2 min read
GuidesModels

Jev by TypeSafe AI: System One Decision Models Explained

How TypeSafe's Jev turns structured state into typed choices, scores, and probabilities, and where a specialized decision model may fit in software.

Sep 20, 2026 · 3 min read
ComparisonsModels

Step 5 Preview vs GLM-5.3 vs DeepSeek V4.1 Flash: What Can Be Verified?

A verification-first comparison framework for Step 5 Preview, GLM-5.3, and DeepSeek V4.1 Flash, with guidance on checking model claims and benchmark methodology.

Sep 20, 2026 · 2 min read