← Back to Home

Z.AI Model Concurrency & Availability Findings

Date: 2026-02-17/18 Subscription: Coding Plan Base URL: https://api.z.ai/api/anthropic Total Keys: 20

Related:

Z.ai Knowledge Base - Complete documentation on tiers, pricing, and API configuration

Z.ai Coding Subscription - Tier limits and quotas quick reference

Chaos Mode - Cross-model concurrent throughput maximization

Troubleshooting Guide - Rate limit and error handling

Key Finding: Concurrency is PER-ACCOUNT, not per-key

Multi-key tests on glm-4.7 and glm-4.5-air both confirm:

Spreading requests across multiple API keys does NOT increase throughput
All 20 keys share the same account-level concurrency quota
The router’s concurrencyMultiplier (= number of keys) is wrong and massively inflates capacity estimates

Model Availability

Working Models (Coding Subscription)

Model	Status	Observed Max Conc (1 key)	Metadata maxConc	Notes
Model	Status	Max Clean Conc	First 429	Consistent 429	Metadata
`glm-4.7`	OK	15	10 (sporadic)	19	3
`glm-4.6`	OK	8+	not seen up to 8	-	3
`glm-4.5`	OK	8+	not seen up to 8	-	10
`glm-4.5-air`	OK	15	10 (sporadic)	17	5
`glm-5`	OK	1 (not stress-tested)	-	-	1
`glm-4.5-flash`	OK	1 (not stress-tested)	-	-	2
`glm-4.7-flash`	OK	1 (not stress-tested)	-	-	1

BLOCKED Models (Error 1113: “Insufficient balance or no resource package”)

These models return HTTP 429 with Z.AI error code 1113 at any concurrency level, including 1. They are NOT available on the Coding Plan subscription and should be removed from tier routing.

Model	Error Code	Error Message
`glm-4-plus`	1113	Insufficient balance or no resource package
`glm-4.5-airx`	1113	Insufficient balance or no resource package
`glm-4.7-flashx`	1113	Insufficient balance or no resource package
`glm-4.5-x`	1113	Insufficient balance or no resource package

INVALID Models (Error 1211: “Unknown Model”)

These model IDs don’t exist in Z.AI’s API:

Model	Error Code	Error Message
`glm-4.32b-0414-128k`	1211	Unknown Model, please check the model code
`glm-flash`	1211	Unknown Model, please check the model code

Detailed Concurrency Results

glm-4.7 (up to 20 parallel)

Concurrency  1-9:   All green (0 429s)
Concurrency 10:     Round 1: 10/10 ok, Round 2: 9/10 ok (1 429) -- SOFT LIMIT
Concurrency 11-15:  Sporadic 429s (1-2 per round), but most succeed
Concurrency 16-20:  Increasing 429s (1-4 per round), still majority succeed

Conclusion: Soft limit around 10 concurrent per account. Z.AI queues rather than hard-rejects, allowing bursts above the limit with some failures.

glm-4.5-air (up to 12+ parallel)

Concurrency  1-9:   All green (0 429s)
Concurrency 10:     Round 1: 10/10 ok, Round 2: 9/10 ok (1 429) -- SOFT LIMIT
Concurrency 11-12:  All green again (0 429s)

Conclusion: Similar soft limit around 10. Occasional 429s are transient, not a hard wall.

Impact on Proxy Configuration

Configuration: See Model Routing for tier-based routing configuration. Troubleshooting: See TROUBLESHOOTING.md for handling 429 errors.

1. Remove unavailable models from tiers

The following models must be removed from tier routing:

glm-4-plus (light tier)
glm-4.5-airx (medium tier)
glm-4.7-flashx (light tier)
glm-4.5-flash (light tier) – free model, may have very low limits
glm-4.32b-0414-128k (light tier) – model doesn’t exist

2. Fix concurrencyMultiplier

Since concurrency is per-account (not per-key), the effective concurrency multiplier should be 1, not the number of API keys (20). With 20 keys, the router currently thinks each model has 20x its actual capacity.

3. Update maxConcurrency values

Based on stress test results, recommended values:

Model	Current	Recommended	Rationale
`glm-4.7`	3	10	Clean up to 9, soft limit at 10
`glm-4.6`	3	8	Clean up to 8 (need higher testing)
`glm-4.5`	10	10	Confirmed ok up to 8, keep 10
`glm-4.5-air`	5	10	Clean up to 9, soft limit at 10
`glm-5`	1	1	Keep conservative until tested
`glm-4.7-flash`	1	1	Free model, keep conservative
`glm-4.5-flash`	2	2	Free model, keep conservative

4. Distinguish error 1113 from real 429s

The proxy currently treats error 1113 (billing/subscription) the same as rate limit 429s. Error 1113 should be classified differently:

Don’t retry (model is permanently unavailable on this subscription)
Don’t cooldown (it’s not a rate limit)
Mark model as “unavailable” and skip in routing