GLM-5.3 Flash API
Choose GLM-5.3 Flash
The natively multimodal member of the 5.3 generation: text, image, video, and file input at roughly a ninth of the GLM-5.3 price, with the same 1M-token context, always-on reasoning, tool calling, and prompt caching.
GLM-5.3 Flash
Z.ai multimodal high-volume model
glm-5.3-flashHigh-volume classification and extraction, screenshot and document understanding, video and file analysis, and routine agent steps where per-token price decides the routing.
GLM-5.3 Flash pricing
Estimate a request before you route traffic. Every user group uses the same rate, and output tokens already include reasoning tokens.
Token calculator
Enter the token mix for one request.Estimated request cost
GLM-5.3 FlashMinimum charge: 0.01 credits per request. Image, video, and file input is billed as input 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-5.3 Flashglm-5.3-flash | All context sizes | $0.150 / 1M10.2 cr / 1M | $0.031 / 1M2.1 cr / 1M | $0.500 / 1M34 cr / 1M |
GLM-5.3 Flash
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 Flash API for multimodal and high-volume work
The volume tier of the 5.3 generation on EvoLink's unified API. Text, image, video, and file input at roughly a ninth of the GLM-5.3 price, with the same 1M-token context, always-on reasoning, tool calling, and prompt caching.
Where GLM-5.3 Flash belongs in a production model stack
Flash is priced for volume, so the question is rarely whether you can afford it — it is whether it clears your accuracy bar. Run it first, escalate only what fails, and the price gap against the flagship starts working for you.
High-volume classification and extraction
Ticket routing, tagging, structured extraction, and content triage run at a price where a full re-run costs less than the engineering time to avoid one. Batch size stops being the constraint.
Screenshot and document understanding
Send UI screenshots, scanned pages, or diagrams as image_url blocks and get structured output back. This is the capability GLM-5.3 does not have at all — the flagship is text-only.
Video and file analysis
Video and file input are accepted natively rather than through a separate pipeline, so one request can carry mixed evidence instead of being split across a transcription step and a reasoning step.
Routine steps inside a larger agent
Summarising a tool result, deciding a branch, reformatting output — the steps that dominate call volume but not difficulty. Keep the flagship for the steps where reasoning depth actually decides the outcome.
What Flash gives you, and what it asks in return
Flash is not a cut-down 5.3 with a smaller window — the context, protocols, and caching are identical. Two things differ, and one is still worth measuring yourself.
Native multimodal input, which the flagship lacks
Text, image, video, and file input all work through the standard messages structure. Images go in as image_url blocks with a public URL (officially recommended) or a Base64 data URL, one block per image.
Roughly a ninth of the flagship price
Input and output both sit near one ninth of the GLM-5.3 rate, and cache reads are cheaper again. That gap is wide enough to change the shape of a workload, not just its bill.
Same reasoning constraint as the rest of 5.3
Reasoning is always on here too, and thinking.type: "disabled" is rejected. Z.ai additionally recommends clear_thinking: false for Flash; the parameter is passed through unchanged.
Media token accounting is worth verifying yourself
Images and video count into prompt_tokens and bill at the input rate, but the upstream publishes no per-image token formula. Run a representative sample, read the returned usage, and size your batch from that rather than from an estimate.
Why Flash can carry this much volume
Flash keeps the platform properties of the 5.3 generation and changes only the price and the input modalities. That combination is what makes it a default rather than a fallback.
Mixed text and media in one request
One message can carry instructions, several images, and a file together, so evidence that belongs to one decision stays in one call instead of being stitched across a pipeline.
The same 1M context at one flat rate
Flash is not a short-context model. The full 1M window is available at a single rate with no long-context tier, which is what makes bulk document work practical at this price.
Cache reads at a fraction of input
A stable system prompt and tool schemas bill at the cache-read rate across a long loop. On an already-cheap model this is what pushes the marginal cost of an extra step close to nothing.
What to verify before routing volume to Flash
Two of these are correctness checks and two are cost checks. The media-token check is the one most teams skip and later regret on a large batch.
Use glm-5.3-flash as the model ID
The same ID works on Chat Completions and Anthropic Messages and matches the upstream name exactly.
Remove any thinking.type: "disabled"
The whole 5.3 generation rejects it. Use thinking.type: "enabled" with reasoning_effort: "low" for the cheapest, fastest path.
Measure media tokens on a real sample
Send representative images or video, read prompt_tokens from the response, and derive your own per-item cost before committing to a batch size.
Define the escalation rule to GLM-5.3
Decide in advance what counts as a Flash failure and what it triggers. Without a rule, teams either escalate nothing and ship errors, or escalate everything and lose the price advantage.
Measure the accuracy gap, then price it
The useful comparison is not Flash against a benchmark, it is Flash against GLM-5.3 on your own tasks. If Flash matches the flagship on nine requests in ten, running both and escalating the tenth is far cheaper than running the flagship on all ten.
Price the accuracy gap instead of assuming it. A model at one ninth the cost only has to be right most of the time for a run-then-escalate setup to beat routing everything to the flagship — but "most of the time" is a number you have to measure on your own workload, not inherit from a benchmark.
Compare against GLM-5.3 and DeepSeek V4 Flash
EvoLinkFlash is the volume tier of the 5.3 generation. Check whether it clears your accuracy bar first; if it does, the price gap against the flagship is large enough to change how you route everything routine.
| Model | GLM-5.3 Flash | GLM-5.3 | DeepSeek V4 Flash |
|---|---|---|---|
| Input / output | $0.15 / $0.5 | $1.4 / $4.4 | $0.442 / $1.324 |
| Context | 1M | 1M | 1M |
| Caching | Cache reads | Cache reads | Cache reads |
| Best for | High-volume classification and extraction, screenshot and document understanding, video and file analysis, and routine agent steps where per-token price decides the routing. | The 5.3 flagship: text-only, roughly nine times the price, and the right escalation target when Flash misses the accuracy bar on hard reasoning. | DeepSeek's high-volume route with a 1M context and very low cache-read pricing. The closest cost comparison for bulk text work. |
Related models
GLM-5.3
The 5.3 flagship: text-only, roughly nine times the price, and the right escalation target when Flash misses the accuracy bar on hard reasoning.
View model
DeepSeek V4 Flash
DeepSeek's high-volume route with a 1M context and very low cache-read pricing. The closest cost comparison for bulk text work.
View model
GLM-5.2
The previous GLM flagship. Still relevant for clients that depend on disabling reasoning, which the 5.3 generation no longer allows.
View model
Gemini 3.7 Flash
Google's low-cost multimodal route — a useful cross-vendor baseline when image and document understanding drive the choice.
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 Is Out: What Shipped
What the August 14 release confirmed about the 5.3 generation — specs, model IDs, and staged access — and what was still open at launch.
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 across the generation.
Read guide
GLM-5.3 vs Claude
Benchmarks, API contracts, and access reality compared before you route agents to either family.
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 Flash API FAQ
Is the GLM-5.3 Flash API available through EvoLink?
Yes. GLM-5.3 Flash is available as a production model, served over both Chat Completions and Anthropic Messages.
What model ID should I use?
Use glm-5.3-flash for both Chat Completions and Anthropic Messages. The EvoLink model name matches the upstream name exactly.
What input types does Flash accept?
Text, images, video, and files. This is the main difference from GLM-5.3, which is text-only.
How do I send an image?
Put an image_url block in messages[].content[] and pass a public URL (the officially recommended form) or a Base64 data URL. For multiple images, add one image_url block per image.
How is image and video input billed?
Media is counted into prompt_tokens and charged at the input rate — there is no separate media SKU. The upstream does not publish a token conversion formula for images, so run a representative sample and read the returned usage before sizing a large batch.
Can I disable reasoning?
No. The whole 5.3 generation always reasons and rejects thinking.type: "disabled". Use thinking.type: "enabled" with reasoning_effort: "low" when you want the cheapest, fastest path.
What is clear_thinking and should I set it?
It controls whether prior reasoning is cleared between turns. Z.ai recommends clear_thinking: false for Flash. Pass it through as-is; EvoLink does not rewrite it.
How much cheaper is Flash than GLM-5.3?
Roughly a ninth of the price on input and output. That gap is usually large enough to justify running Flash first and escalating only the requests that fail your acceptance check.
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.
Is Flash a good default for high-volume work?
Yes, that is what it is priced for. Classification, extraction, routine agent steps, and document or screenshot understanding are the natural fit. Escalate to GLM-5.3 only where deeper reasoning measurably changes the outcome.
What should a production evaluation measure?
Track first-pass success, accepted deliverables, retries, reasoning-token share of output, cache-hit ratio, media tokens per request, time to accepted result, and the escalation rate to GLM-5.3.