Quick answer. Qwen 3.7 Flash ($0.03 input / $0.13 output per 1M tokens) is the cheaper model on both axes — roughly 4.7x cheaper input and 2.2x cheaper output than DeepSeek V4-Flash ($0.14 / $0.28) — and it adds native vision. DeepSeek V4-Flash wins on raw open-weight coding depth and a deeper 98% cache discount.
By July 2026 the interesting question stopped being "which model is smartest" and became "what can I finally afford to run at scale." The budget tier is where that gets decided. On one side is DeepSeek V4-Flash, the re-post-trained 284B-parameter (~13B active) text model that anchors the low-cost end of the DeepSeek family. On the other is Qwen 3.7 Flash, Alibaba's July 27 release that undercuts almost everything and throws in native multimodality.
This is a cost-efficiency comparison, so we lead with price per token and cost per task — then we're honest about the axes where the "more expensive" option is actually the right call. If you want the full landscape of how DeepSeek V4 stacks against the Western flagships, the DeepSeek V4-Flash cost-vs-frontier breakdown is the hub post; this page zooms in on the two cheapest capable open-weight-adjacent options.
How much cheaper is Qwen 3.7 Flash than DeepSeek V4-Flash?
On the headline list price, Qwen 3.7 Flash is cheaper on both input and output. We verified the Qwen number via the live model catalogue and multiple provider listings, because it was not in our source brief — the confirmed base rate is $0.03 input / $0.13 output per 1M tokens for prompts up to 32K tokens.
| Model | Input / 1M | Output / 1M | Cached input | Context | Modality |
|---|---|---|---|---|---|
| Qwen 3.7 Flash | $0.03 | $0.13 | ~60-80% off (workload-dependent) | 1M | Text + image + video |
| DeepSeek V4-Flash | $0.14 | $0.28 | ~$0.003 (98% off) | 1M | Text only |
| Qwen 3.5 Flash (reference) | $0.07 | $0.26 | — | — | — |
The multipliers, rounded: Qwen 3.7 Flash is about 4.7x cheaper on input ($0.03 vs $0.14) and about 2.2x cheaper on output ($0.13 vs $0.28). It's also a clean generational drop from Qwen 3.5 Flash's $0.07 / $0.26 — Alibaba took input down more than half and added vision in the same step.
One important caveat on the Qwen rate: that $0.03 / $0.13 is the sub-32K-token tier. For prompts between 32K and 256K, the rate jumps to roughly $0.10 input / $0.40 output. So on long-context agentic runs where every call carries a big system prompt or repo dump, Qwen's advantage narrows or flips — its output at $0.40 above 32K is actually more expensive than DeepSeek V4-Flash's flat $0.28. Read the axis you actually run on, not just the headline.
Where does DeepSeek V4-Flash's cache discount change the math?
DeepSeek's cached-input price is the sleeper advantage. At roughly $0.003 per 1M cached tokens — a 98% discount — cache-heavy workloads (chat with a fixed system prompt, RAG over a stable corpus, repeated tool schemas) blend down to about $0.06/M. Qwen 3.7 Flash also caches, but the reported effective saving is in the 60-80% range rather than 98%, so on workloads where the same context is reused thousands of times per day, DeepSeek's cache can close and sometimes overtake Qwen's list-price lead.
The rule of thumb: Qwen 3.7 Flash wins on fresh, varied, short-context calls; DeepSeek V4-Flash wins on high-cache-hit, repeated-context pipelines. Profile your cache-hit ratio before you assume either is "the cheap one."
What does a real monthly bill look like?
Take a mid-size assistant or coding-agent workload of 50M input + 10M output tokens per month, priced at list (no cache, sub-32K context) so the comparison is apples-to-apples:
| Model | Input cost | Output cost | Monthly total |
|---|---|---|---|
| Qwen 3.7 Flash | $1.50 | $1.30 | $2.80 |
| DeepSeek V4-Flash | $7.00 | $2.80 | $9.80 |
| DeepSeek V4-Pro | $21.75 | $8.70 | $30.45 |
| GPT-5.6 Luna | $10.00 | $12.00 | $22.00 |
| Gemini 3.5 Flash | $37.50 | $45.00 | $82.50 |
| Claude Opus 5 | $250.00 | $250.00 | $500.00 |
At this profile Qwen 3.7 Flash runs about $2.80/month versus DeepSeek V4-Flash's $9.80 — roughly 3.5x cheaper end to end. Both are a rounding error next to Claude Opus 5's $500. That's the whole point of the budget tier: the cost of experimentation has collapsed. For the fuller ladder that includes GPT-5.6 Sol at $550 and Kimi K3 at $300, see the cost-vs-frontier post.
Now flip the workload to cache-heavy with a 40K-token context and the story changes: Qwen slides into its $0.10 / $0.40 tier while DeepSeek's 98% cache discount kicks in, and DeepSeek V4-Flash can end up the cheaper of the two. There is no single winner — there's a winner per workload shape.
Which one is actually better at coding?
Price only matters if the output is usable. Here the two models are built for different jobs.
DeepSeek V4-Flash is a text-only coding-and-agent workhorse. On DeepSeek's own harness (max-effort, treat as a ceiling) it posts Terminal Bench 2.1 = 82.7, Toolathlon = 70.3, and DSBench-FullStack = 68.7 — strong for the price, and it shares architecture with the V4-Flash-Preview but got a meaningful post-training jump on the July 31 re-release. It's OpenAI-API-compatible, which softens migration, and the MIT-licensed DeepSeek V4 family means you can self-host the open weights if China-hosted API residency is a compliance problem.
Qwen 3.7 Flash trades some coding depth for breadth: native text + image + video at a 1M context, which DeepSeek V4-Flash simply cannot do. Independent benchmarks for this exact Flash variant are still thin, so don't assume it matches V4-Flash on hard repo-level coding — but for multimodal subagents, screenshot-to-code, UI-diffing, and vision-in-the-loop QA, it's the obvious low-cost pick because the alternative (Opus 5, GPT-5.6, or Kimi K3 native vision) is 30-150x the price.
For a deeper head-to-head on the coding axis specifically, the Qwen 3.7 vs DeepSeek V4 coding comparison and the GLM-5.2 vs DeepSeek V4 post round out the budget-tier field.
What about open weights and what's coming next?
Open-weight availability is a real differentiator for self-hosters. DeepSeek V4's larger sibling, V4-Pro (1.6T total / 49B active), ships under MIT with open weights you can run yourself — one of the reasons the DeepSeek family keeps winning cost-conscious teams. On the Qwen side, the big story is Qwen 3.8 (2.4T parameters), which Alibaba has signalled is going open-weight soon. If it lands the way the family's prior open releases have, it will reset the "biggest open model you can self-host" bar and put real pressure on the whole budget-to-mid tier.
For most teams the practical takeaway is: neither budget model locks you in. Qwen 3.7 Flash gives you the cheapest multimodal API today; DeepSeek V4-Flash gives you a self-hostable, OpenAI-compatible text engine with the deepest cache discount. Pick by workload, keep an eye on Qwen 3.8.
When should you reach past the budget tier?
Being honest about where the pricier model earns its keep:
- Hardest long-horizon autonomous agents. Multi-hour, many-step autonomous coding still favors Claude Opus 5 and GPT-5.6 Sol. Neither budget Flash model is built for that ceiling.
- Native vision at frontier quality. Qwen 3.7 Flash does vision cheaply, but for high-stakes multimodal reasoning, Kimi K3 (native vision, 1M ctx) or Opus 5 will be more reliable.
- Data-residency and compliance. DeepSeek's API is hosted in China. If that's a blocker, either self-host the open weights or reach for a Western-hosted option — factor that cost in.
If you're weighing these models because you're building something real and want an engineering team that can ship it — and pick the right model tier for each job — hire vetted remote developers through Codersera and extend your team without the hiring risk.
The bottom line on cost per capable token
On the pure sticker price at short context, Qwen 3.7 Flash is the cheapest capable model of the two — 4.7x cheaper input, 2.2x cheaper output, plus native vision. But "cheapest" is workload-dependent: DeepSeek V4-Flash's 98% cache discount and flat $0.28 output make it cheaper on high-cache, long-context pipelines, and its text-coding depth plus MIT open weights give it an edge where self-hosting and hard coding matter. Match the model to the axis you actually run on, and either one will save you 30-150x versus a Western flagship.
FAQ
Is Qwen 3.7 Flash cheaper than DeepSeek V4-Flash?
Yes, on list price at short context. Qwen 3.7 Flash is $0.03 input / $0.13 output per 1M tokens versus DeepSeek V4-Flash's $0.14 / $0.28 — about 4.7x cheaper input and 2.2x cheaper output. Above 32K tokens Qwen jumps to $0.10 / $0.40, at which point DeepSeek's flat $0.28 output is cheaper.
Does DeepSeek V4-Flash ever cost less than Qwen 3.7 Flash?
Yes, on cache-heavy or long-context workloads. DeepSeek V4-Flash's cached input is ~$0.003/M (a 98% discount), blending to about $0.06/M on cache-heavy pipelines, while Qwen's cache saving is a shallower 60-80%. On repeated-context RAG or chat with a fixed system prompt, DeepSeek can come out cheaper overall.
Which one supports images and video?
Qwen 3.7 Flash does — it's natively multimodal (text, image, and video input) with a 1M-token context. DeepSeek V4-Flash is text-only. If your workload needs vision, Qwen 3.7 Flash is the cheapest capable option by a wide margin.
Can I self-host either model?
DeepSeek's V4 family ships MIT-licensed open weights (V4-Pro is fully self-hostable), which matters if China-hosted API residency is a compliance concern. Alibaba has signalled Qwen 3.8 (2.4T) is going open-weight soon, which will expand self-hosting options on the Qwen side.
How much can I save versus a Western flagship?
On a 50M-input + 10M-output monthly workload, Qwen 3.7 Flash runs about $2.80 and DeepSeek V4-Flash about $9.80, versus roughly $500 for Claude Opus 5 and $550 for GPT-5.6 Sol. That's a 50-180x saving — the reason the budget tier changed what teams can afford to build.