Price cut just one day after launch — $0.15/M input tokens, 89% lower training cost, coding benchmarks match DeepSeek V4 Pro
✅ Input $0.15/M tokens ✅ Output $0.47/M tokens ✅ Cache hit $0.016/M 🔥 67% cheaper than DeepSeek V4 FlashPricing via Alibaba Cloud Model Studio (international) and Bailian platform (China), per million tokens.
| Billing Item | Price (per 1M tokens) |
|---|---|
| Input | $0.15 |
| Output | $0.47 |
| Cache hit | $0.016 |
| Billing Item | Launch Price (Aug 26) | Current Price (Aug 27+) | Change |
|---|---|---|---|
| Input | ¥1.0 | ¥0.8 | ↓20% |
| Output | ¥3.0 | ¥2.7 | ↓10% |
| Cache hit | — | ¥0.1 | — |
August 2026 pricing for major LLM APIs, USD per million tokens.
| Model | Input | Output | Active Params | Context |
|---|---|---|---|---|
| Qwen3.8-Flash NEW | $0.15 | $0.47 | 6B (MoE) | 1M |
| Qwen3.5-Flash (≤128K) | $0.11 | $0.67 | — | 1M |
| Qwen3.7-Flash | $0.034 | $0.132 | — | 256K |
| Qwen3.8-Max | $2.00 | $6.00 | — | 1M |
| DeepSeek V4 Flash (peak) | $0.42 | $1.26 | 13B | 128K |
| DeepSeek V4 Flash (off-peak 50%) | $0.21 | $0.63 | 13B | 128K |
| DeepSeek V4 Pro (peak) | $1.26 | $3.78 | — | 128K |
| GLM-5.3-Flash | ~$0.15+ | ~$0.47+ | 18B (MoE) | — |
Qwen3.8-Flash (codenamed Qwen3.8-Flash-Next during development) is an early preview of the Qwen4 architecture, with four core innovations:
Qwen Sparse Attention (QSA) works alongside GDN: GDN compresses historical context while QSA selects key information. In high-cache-hit 1M-token scenarios, this delivers 7.6x faster prefill and 4.9x faster decode.
Splits the traditional single residual pathway into 4 parallel branches, dynamically gating information flow for better cross-layer communication and training stability.
Uses a refined Muon + AdamW hybrid strategy with refitted scaling laws, eliminating batch warm-up and significantly improving convergence efficiency and training throughput.
Per Qwen's official technical report, Qwen3.8-Flash excels across multiple evaluation suites:
| Benchmark | Domain | Performance |
|---|---|---|
| SWE-bench Pro | Agentic coding | 62.5, leading peers |
| CoWorkBench | Long-horizon office tasks | Beats DeepSeek V4 Flash |
| Toolathlon Verified | Real-world tool use | Matches Claude Opus 4.6 |
| MathVision | Visual math reasoning | Strong multimodal |
| AndroidWorld | Mobile agent tasks | Embodied intelligence |
| ERQA | Embodied reasoning | Multimodal understanding |
Across 14 evaluations, the base model (6B activated) achieved the best results in 8. The fine-tuned version shows even stronger performance in coding, agents, and multimodal tasks.
Access Qwen3.8-Flash via Model Studio with OpenAI-compatible API. New users get free credit. The model serves on QwenCloud with 1M context by default.
Claim $200 Free Credit →Available on Alibaba Cloud's Bailian platform with new user free token quota. First to receive the latest Qwen releases.
Bailian Console →The open-weight version, Qwen3.8-Flash-Next, is available on Hugging Face and ModelScope for download, fine-tuning, and local deployment. The production version with built-in tools and 1M default context is served via QwenCloud API.
At the 2026 Apsara Conference, Alibaba Cloud unveiled the next wave of Qwen foundation models: Qwen3.8-Omni opens a new reasoning partition, Qwen3.8-Max begins self-training, and Qwen4 is already in training (targeting 5-10 trillion parameters), with Qwen4.5 and Qwen5 on the roadmap. New releases include image model Qwen-Image-3.1, speech model Qwen-Audio-3.1 and Qwen3.8-LiveTranslate simultaneous interpretation. Notably, the end-to-end realtime voice model Qwen-Audio-3.1-Realtime is already powering Qwen Office, Qoder (Alibaba's AI coding workbench) and Qwen AI glasses; Alibaba predicts a native unified omni-modal generation model will emerge within three years. For developers, calling Qwen3.8-Flash / Qwen3.8-27B / Omni-Flash via Alibaba Cloud Model Studio remains the most cost-effective way to experience this generation of models.
Source: QbitAI coverage of the 2026 Apsara Conference (Sep 22, 2026). Original pricing and review conclusions above remain valid.
Qwen3.8-27B keeps trending in developer circles this week. In a hands-on test published by QbitAI on September 19, a single prompt led Qwen3.8-27B to produce a 52KB offline data-analysis tool in about 5 minutes — automatically cleaning tables, aggregating revenue stats and rendering bar charts — and even a realistic 12306 ticket-booking page (a convincing front-end shell that, by default, does not connect to live data or payment backends). Engineer Alok paired Qwen3.8-27B with Cerebras to build an offline "AI computer desktop" that regenerates historical versions of Google, YouTube and other sites on the fly at roughly 1,950 tokens/s (individual benchmark). On September 18, the Qwen team also launched Qwen3.8-Omni-Flash, a next-generation omnimodal model with text/image/audio/video inputs and a 1M-token context, cutting per-hour audio-input API pricing by more than 98% versus the previous generation. From local 27B deployment to cloud APIs, the Qwen3.8 family is making "one-prompt tool delivery" real — for enterprises, calling Qwen3.8-Flash, Qwen3.8-27B and Omni-Flash via Alibaba Cloud Model Studio / Bailian is the most cost-effective way to try this generation of models.
This week a Hacker News benchmark post by Quesma CTO Piotr Migdał gave a definitive answer for running Qwen3.8-27B locally: the 17GB Q4_K_M 4-bit quant matches the 55GB BF16 full model on Terminal-Bench 2.1 and fits on a 24GB consumer card (e.g. RTX 4090) with room for ~64K tokens of context. At 4-bit, GPQA Diamond, IFBench and Terminal-Bench 2.1 scores are statistically indistinguishable from the reference, at one-third the on-disk size. Compression hits a clear cliff at 1-bit: the 6.2GB UD-IQ1_S falls to near-random on GPQA Diamond, and the xhigh reasoning effort actually makes it worse (the model often runs out of token budget without committing to an answer). Unsloth's stated minimum for agentic / tool-calling work remains the 9.8GB Q2_K_XL. For teams weighing a self-hosted 27B deployment, this is the most current, third-party-validated hardware and quantization guidance available.
Source: Quesma: Benchmarking Qwen3.8 27B quantizations (202 HN points, 99 comments on Sept 8, 2026). The original pricing and architecture analysis above remain valid; this card only adds the local-deployment sizing reference.
Qwen3.8-Flash is a MoE model (125B total / 6B active) optimized for cost efficiency and high throughput. Qwen3.8-27B is a dense vision-language model (all 27B active) for deep reasoning at $0.42/$3.08 per million tokens. Choose Flash for scale, 27B for depth.
When a request hits the context cache (e.g., repeated system prompts, long document prefixes), cached input tokens are billed at $0.016/M instead of $0.15/M — a 90% discount. This dramatically reduces costs for agents, RAG, and multi-turn conversations.
Qwen3.8-Flash supports text and image input natively (input_modalities: ["text", "image"]), with video analysis in the production version. It handles chart analysis, document OCR, and visual math.
5,000 RPM (requests per minute) and 5,000,000 TPM (tokens per minute), suitable for enterprise-grade high-concurrency deployments.
$200 free credit + ECS from $4.50/month + open-weight models available
Claim $200 Free Credit → View ECS Plans