Qwen3.8-Flash-Next vs GLM-5.3-Flash: Price, Speed & Coding Compared
Both released August 26, 2026, both open-weights, both 1M-class context — but for a developer picking between Qwen3.8-Flash-Next and GLM-5.3-Flash, the decision reduces to three questions: what will it cost at your real token volumes, how fast will it actually run, and which one wins the published coding evidence. This page answers those three with the numbers that exist and flags the numbers that don't. For the full eight-dimension comparison (multimodal, reasoning, deployment trade-offs), see our GLM-5.3-Flash vs Qwen3.8-Flash-Next page.
Decision snapshot
Captured: August 30, 2026. Cheapest API this week: GLM-5.3-Flash (50% launch discount until September 9, 2026 24:00 UTC+8). Cheapest at list: Qwen3.8-Flash. Coding evidence: both published — GLM leads every shared suite. Published throughput number: Qwen only (5M TPM). Neither vendor's benchmarks are independently verified.
Price: The Calendar Problem GLM until Sept 9, Qwen after
Per million tokens, as published (accessed August 30, 2026; Qwen CNY converted at ~7.1 CNY/USD, estimates marked):
| Rate | GLM-5.3-Flash list | GLM-5.3-Flash launch (until Sept 9, 24:00 UTC+8) | Qwen3.8-Flash |
|---|---|---|---|
| Input | $0.15 | $0.075 | ¥0.8 ≈ $0.11 |
| Output | $0.50 | $0.25 | ¥2.7 ≈ $0.38 |
| Cached input | $0.03 | $0.015 | ¥0.1 ≈ $0.014 |
| Cache storage | Temporarily free | — | |
Monthly cost at three realistic developer profiles (API only):
| Profile (in / out per month) | GLM at launch | GLM at list | Qwen (converted) |
|---|---|---|---|
| Side project: 10M / 2M | $1.25 | $2.50 | ≈$1.86 |
| SaaS feature: 50M / 15M | $7.50 | $15.00 | ≈$11.20 |
| Heavy agent service: 300M / 100M | $47.50 | $95.00 | ≈$71.00 |
The pattern: GLM wins every row before September 9; Qwen wins every row after. Because output tokens cost 3.3–3.4x input on both models, agentic workloads (code, tool calls, long generations) are decided by output price — where the GLM discount is deepest. If you can commit usage before the deadline, GLM is the value pick; if your project launches in October, Qwen's list price is the baseline. Detailed cost scenarios on our GLM-5.3-Flash pricing page.
Speed: What's Published and What Isn't Unverified on both sides
- Qwen3.8-Flash (hosted): 5M tokens per minute (TPM) published on the official platform page. No per-request latency figures.
- GLM-5.3-Flash (hosted): no official TPM/TPS figure published as of August 30, 2026. Zhipu's published efficiency metrics are architectural (3.01x lower attention compute, 4.44x smaller KV cache vs GLM-5.3) — relevant to long-context throughput, not a latency number.
- Self-hosted inference: Flash-Next's ~6B active parameters point to cheaper per-token serving than GLM's ~18B, but Alibaba's own guidance warns the 51B N-gram table plus vision components inflate real VRAM — measure per engine. No independent throughput comparison exists for either model yet.
Reasoning modes affect perceived speed: GLM-5.3-Flash's thinking is always on (reasoning_effort low/high/max) — low effort is the fast path, but you cannot disable thinking entirely. Qwen3.8-Flash's thinking mode is optional with up to 262K CoT. If latency-sensitive routing is your architecture, plan for always-on thinking on GLM and compare with your own load tests.
Coding: Shared Suites vs. Extra Breadth GLM-5.3-Flash on shared suites
Both vendors published coding data in their official materials; the comparison on the suites they share is not close, but Qwen adds suites GLM doesn't cover:
| Published coding results (vendor-reported) | GLM-5.3-Flash | Qwen3.8-Flash-Next |
|---|---|---|
| Terminal-Bench 2.1 | 84.3 | Not published |
| DeepSWE v1.1 | 63.4 | 58.7 |
| NL2Repo-Bench | 56.3 | 48.1 |
| SWE-bench Pro | Not published | 62.5 |
| SWE-bench Multilingual | Not published | 81.0 |
| LiveCodeBench v6 | Not published | 91.9 |
| Internal Code Bench | "Comparable to Claude Opus 4.8" (Z.ai) | Not published |
All numbers are vendor-reported and unverified. On the suites both publish, GLM-5.3-Flash leads (DeepSWE 63.4 vs 58.7; NL2Repo 56.3 vs 48.1); Qwen publishes additional coding scores (SWE-bench Pro, SWE-bench Multilingual, LiveCodeBench v6) with no GLM equivalent, so the head-to-head is incomplete — and Flash-Next's 6B-active design could still deliver surprising efficiency per token on real workloads. For coding-agent harness context, see Best AI Coding Assistant 2026 and the DeepSeek Harness review.
Context & Memory Budget Tie on API
Both hosted APIs serve 1M context (GLM native with 128K output; Qwen default with 991K max input / 131K output). Self-hosted, GLM's 1M is native while Flash-Next's native window is 262K (extendable to 1M). For repository-scale prompts and agent memory, either works on the API; for self-hosted whole-repo work, GLM's native 1M avoids extension complexity. See Best AI Models for Long-Context Work 2026 for the wider field.
Developer Decision Matrix
| Your situation | Pick | Why |
|---|---|---|
| Indie dev, API, starting now | GLM-5.3-Flash | Cheapest until Sept 9 + leads every shared coding suite |
| Production launch after mid-September | Qwen3.8-Flash | Cheaper at list; GLM leads shared coding suites — re-benchmark before launch |
| Self-hosting on existing GPUs | Qwen3.8-Flash-Next | 6B active; check Qwen Community License terms; measure VRAM first |
| Coding agent / IDE workload | GLM-5.3-Flash | Published coding results; Coding Plan has 3x quota vs GLM-5.3 |
| Video / computer-use app | Qwen3.8-Flash | Published video/OSWorld results; built-in tools |
| Long-context RAG (1M prompts) | Either API | Both 1M; decide on cache rate (Qwen ≈$0.014 vs GLM $0.03 list) |
Bottom Line
For a developer deciding this week: GLM-5.3-Flash wins the price/coding combination — cheaper during its launch window and ahead of Qwen on every coding suite both vendors publish (DeepSWE 63.4 vs 58.7; NL2Repo 56.3 vs 48.1, vendor-reported). Qwen3.8-Flash-Next wins the efficiency/self-host angle and the post-discount price, and publishes extra coding scores (SWE-bench Pro, LiveCodeBench v6) with no GLM equivalent yet. The speed question is genuinely open — Qwen publishes a throughput number, GLM publishes architectural efficiency, and no independent latency test of either exists yet. Budget for list prices and benchmark both on your own tasks before September 9; when GLM publishes equivalents to Qwen's extra benchmarks, the head-to-head will be complete.
Related: Qwen3.8-Flash-Next review · GLM-5.3-Flash review · Full GLM vs Qwen comparison · Best Open-Weight Coding Models 2026
FAQ
Which is cheaper: Qwen3.8-Flash or GLM-5.3-Flash?
Qwen at list (≈$0.11/$0.38 vs $0.15/$0.50 per million in/out); GLM during its 50% launch discount ($0.075/$0.25 until September 9, 2026 24:00 UTC+8).
Which model is faster?
Unverified. Qwen publishes 5M TPM for the hosted Flash API; Zhipu published no throughput figure for GLM-5.3-Flash. No independent latency comparison exists as of August 30, 2026.
Which is better for coding?
By published evidence, GLM-5.3-Flash leads every suite both vendors publish (DeepSWE 63.4 vs 58.7, NL2Repo 56.3 vs 48.1, vendor-reported); Qwen adds SWE-bench Pro 62.5 and LiveCodeBench v6 91.9 with no GLM equivalent. Independent verification pending.
Should I wait for the GLM discount to end before choosing Qwen?
Only if independent verification or GLM's missing equivalents to Qwen's extra benchmarks change the picture. Use the GLM discount window to benchmark both on your own tasks.
Can I self-host both?
Yes — GLM under MIT (SGLang/vLLM/TokenSpeed/KTransformers, ~328GB FP8); Qwen3.8-Flash-Next with public weights (Qwen Community License v1.0 per its card; the Qwen3.8 repo states Apache-2.0 for the 27B/2.4T releases — verify per model).
Sources (accessed August 30, 2026): Zhipu official docs (GLM-5.3-Flash page), z.ai/subscribe, Alibaba Qwen official model card for Qwen3.8-Flash-Next (ModelScope), Qwen platform (qianwenai.com), QwenLM/Qwen3.8 GitHub, OpenRouter listings (platform pricing). All benchmarks vendor-reported; CNY→USD conversions approximate at ~7.1 CNY/USD. ToolStep claims no laboratory test results. See our editorial policy.