You run text work on GLM, whether long-document review, structured output, tool-calling agents or adult roleplay, and SpicyAPI lists two full-size IDs: zai/glm-5.2/chat and zai/glm-5.3/chat. Use GLM 5.3 for coding, agent and tool-calling work, and for high-volume passes where low reasoning saves output tokens. Use GLM 5.2 for adult fiction and roleplay, and when you want the lower rate. If the job is short, structured and cost-driven, try the cheaper GLM 5.3 Flash before either.
Both read 1,048,576 tokens, return up to 131,072, take text only, and share one request shape apart from the fields below.
Which one for which job
| Your job | Pick | Why |
|---|---|---|
| Coding agents, long tool loops, repository or incident analysis | 5.3 | Z.ai's post-training targets agentic coding; 5.3 also adds a lower reasoning setting and control over parallel tool calls |
| Sequential tool calls (reserve a slot, then a room) | 5.3 | Set parallel_tool_calls: false on 5.3 to request one tool at a time; enforce booking dependencies in your application |
| Bulk classification, extraction, rewriting | 5.3 with low, or 5.3 Flash | 5.2 cannot go below high |
| Hold a setting steady when comparing changed prompts | 5.3 | Only 5.3 accepts seed; it does not guarantee identical replay |
| Adult roleplay and fiction | 5.2 | Labelled Uncensored; rendered stronger scenes with plain requests more often than 5.3, which is labelled Light and softens or refuses more |
| Steady daily traffic where one 5.2 prompt already passes | 5.2 | About 7% lower token rates; no new controls needed |
The differences in the request
| Feature | GLM 5.2 | GLM 5.3 |
|---|---|---|
reasoning_effort | high, max; default max | low, high, max; default max |
seed | Not accepted | Accepted |
parallel_tool_calls | Not accepted | Accepted; default true |
max_tokens | Default 32,768; maximum 131,072 | Same |
| Temperature / top-p | 0-1, default 1 / 0.01-1, default 0.95 | Same |
| Stop sequences | One | One |
| JSON output, function tools | Supported | Supported |
Both endpoints reject fields they do not list, so a body written for 5.3 with seed fails on 5.2 instead of being ignored. Keep per-model fields apart.
Neither route accepts medium, and neither can switch reasoning off. Z.ai's own API maps none and minimal to no thinking on 5.2; that does not apply to this hosted 5.2 contract. Because the default is max on both, a call you did not tune reasons as deeply as possible, and reasoning is billed as output.
For a roleplay client with a configurable effort picker, map its lowest setting to high on 5.2 or low on 5.3, and its highest setting to max. Check the outgoing value; the client’s display label is not an API parameter.
What it costs
| USD per million tokens | GLM 5.2 | GLM 5.3 | GLM 5.3 Flash |
|---|---|---|---|
| Uncached input | $0.98 | $1.05 | $0.1275 |
| Cached input | $0.182 | $0.195 | $0.0255 |
| Output | $3.08 | $3.30 | $0.425 |
At identical token counts 5.3 costs 7.14% more than 5.2; 5.2 is 6.7% cheaper than 5.3. Count cache hits at the cached rate instead of charging them again as uncached input. Three jobs:
| Job | 5.2 | 5.3 | 5.3 Flash |
|---|---|---|---|
| 40,000 uncached input, 4,000 output | $0.05152 | $0.0552 | $0.0068 |
| 30,000 cached + 10,000 uncached input, 4,000 output | $0.02758 | $0.02955 | $0.00374 |
| Long document: 600,000 uncached input, 8,000 output | $0.61264 | $0.6564 | $0.0799 |
Ten thousand runs of the first job cost $515.20 on 5.2 and $552 on 5.3. The rate gap is small; reasoning length moves the bill more. If 5.3 at low produces an accepted answer with a third of the output tokens, it is cheaper than 5.2 at max. Compare cost per accepted answer from the returned usage, not cost per call.
Adult writing: where each model stops
Based on our September 2026 content tests (English, single requests):
- Suggestive and implied scenes. Both soften the mildest rung (a flirtatious scene) on every phrasing, and both write implied sensual scenes cleanly.
- Stronger adult content in plain terms. 5.2 writes it when asked directly or through a system prompt, and refuses when it is framed as a story. 5.3 writes it only through a system prompt.
- Written intimate scenes between adults. Neither is reliable. With the system prompt, 5.2 completed 2 of 3 requests and 5.3 completed 4 of 8. These small samples show why retries may be needed, rather than establishing a long-run success rate. Direct requests are refused on both, and story framing alone gets softened output.
- Consensual restraint or violence in fiction. Both render this in most runs.
Set the fiction context in the system prompt on either model, and expect that some requests will still be softened or refused. If this is your main workload, GLM 5.3 Flash is also labelled Uncensored at a fraction of the price. See the GLM 5.2 and GLM 5.3 reviews for the full ladder.
What the control differences mean for a creator's application
Consider a scheduling assistant with three jobs: read an approved scene brief, find a room and reserve it after confirming the time. For the first job, 5.2 has the shared text and JSON capabilities, with lower listed rates. If the extraction is already usable, the newer version alone gives no reason to move it.
For a short extraction in 5.3, low effort is a meaningful configuration option because 5.2 does not expose it. A brief with two characters, three shots and one stated hour does not require the same reasoning scope as a month-long schedule with dependencies. Set the effort explicitly instead of accepting the maximum default for both.
For reservations, parallel_tool_calls: false on 5.3 can keep calls sequential, but the order must also be enforced by your application. A reservation request needs the slot returned by the availability lookup. If the lookup returns no slot, the application should not execute a later booking with an invented time. Sequential calls cannot make missing availability appear.
A seed may help you hold a generation setting steady when comparing changed prompts. It does not make a creative answer canonical, guarantee replay or solve a missing source. For an adult companion application, store the approved character facts in your own record rather than relying on the seed to keep them stable.
A one-hour schedule that can be checked without an AI judge
The sample below reserves 14:00–14:10 for setup and 14:50–15:00 for packing, leaving 40 minutes for the three shots. One human-written allocation is 14:10–14:20 for the letter close-up, 14:20–14:35 for the shared smile and 14:35–14:50 for the walk. These are proposed timings, not a generated GLM schedule or a promise that the shoot will finish in that time.
If the brief adds a ten-minute lighting change between the second and third shots, only 30 minutes remain for shooting. Ask for that transition explicitly and recalculate the available time. A model that produces a smooth-looking table while omitting the transition has not solved the changed task. This is a more useful migration fixture than comparing how persuasive the two explanations sound.
Try it in Playground first
- Open GLM 5.3 and switch to Playground.
- Paste the three-shot teaser brief below and request a schedule at Low reasoning.
- Confirm setup, three shots and packing fit within the one-hour window. Compare 5.2 at High only if this is the route you already use.
One complete request
This turns a short romance teaser brief into a JSON schedule on 5.3 at low effort. For long inputs or deep reasoning, use stream: true: a longer client timeout cannot prevent the network edge’s roughly 100-second non-streaming cutoff.
curl --fail-with-body https://api.spicyapi.ai/v1/chat/completions \
-H "Authorization: Bearer $SPICY_API_KEY" \
-H 'Content-Type: application/json' \
-H "Idempotency-Key: $(uuidgen)" \
-d '{
"model": "zai/glm-5.3/chat",
"reasoning_effort": "low",
"max_tokens": 3000,
"temperature": 0,
"response_format": {"type": "json_object"},
"messages": [
{"role": "system", "content": "Return JSON with shots and unresolved_questions. Every shot needs start_time, duration_minutes and description. Fit within the supplied window. Do not invent bookings."},
{"role": "user", "content": "A non-graphic romance teaser with three shots: a letter close-up, a shared smile and a walk past a railway bench. Available window: 14:00 to 15:00. Allocate 10 minutes for setup before the first shot and 10 minutes for packing after the last shot."}
]
}'To send it to 5.2, change the model to zai/glm-5.2/chat and the effort to high; leave out seed and parallel_tool_calls.
Check choices[0].finish_reason and parse choices[0].message.content, then verify the schedule in your own code: shots, setup and packing must fit in 60 minutes, the first shot no earlier than 14:10, packing done by 15:00. JSON that parses can still overlap two shots or allocate 70 minutes to a one-hour window.
Let an Agent prepare the call
For repeated schedule extraction, ask your coding Agent to use the SpicyAPI Skill: “Read zai/glm-5.3/chat, then adapt the example to my input; show the request before running it.” The CLI lookup is spicyapi models get zai/glm-5.3/chat --json; the MCP lookup is spicyapi_model_get with model: "zai/glm-5.3/chat". Chat calls use a compatible client or the chat API; the media task commands do not generate chat replies.
Moving between them
For a like-for-like migration check, run your existing prompts on both IDs at high and compare accepted answers and total usage. Then try 5.3 at low separately to see whether its shorter reasoning still meets the same checks. Keep the old ID as a fallback until the new route passes.
Invalid reasoning value. medium fails on both, and low fails on 5.2.
Empty visible answer. A stop string can fire during reasoning and end the turn before any answer. Remove it unless you need it, and never use a common word such as END.
Photo or PDF rejected. Both are text only; extract the text first.
Tool call treated as done. Return the full assistant tool-call message, run an allowlisted function, and append a role: tool message with the matching tool_call_id. A proposed booking is not a booking.
Answer cut off. Raise max_tokens (maximum 131,072) or ask for less, and reject a length finish reason before importing JSON.
For a new coding or tool integration, start with GLM 5.3. For an existing adult-fiction workflow that already works, GLM 5.2 remains the less expensive, more cooperative starting point; the version change alone is not a reason to move it.
