LLM Pricing Digest: DeepSeek Quietly Drops Two New Flash Models

Weekly roundup from LLM Price Watch — tracking Claude, GPT, Gemini, DeepSeek, and Grok pricing daily.

---

This week's data is straightforward: no price changes across any of our five tracked providers. Zero. If you've got a budget locked in for GPT-4o, Claude Sonnet, or Gemini Flash, nothing moved on you.

What did happen is DeepSeek showed up with two new models on OpenRouter, and they're worth knowing about.

What's New from DeepSeek

deepseek/deepseek-v4.1-flash — spotted September 10th.

deepseek/deepseek-v4-flash-vision-exp:batch — spotted September 9th, a day earlier.

Let's talk about what these names actually tell us.

The v4.1-flash model follows DeepSeek's now-familiar naming pattern: a Flash tier is their lower-latency, lower-cost option, comparable in positioning to Gemini Flash or GPT-4o mini. If previous DeepSeek Flash releases are any guide, this is aimed squarely at high-volume, cost-sensitive workloads — think classification, summarization, structured extraction, anything where you're firing off thousands of calls and cost-per-token matters more than raw capability.

The second model is more interesting: deepseek-v4-flash-vision-exp with a :batch suffix. Two things stand out here. First, the vision tag means multimodal input — image understanding on top of text. Second, the exp label means experimental, so treat it accordingly: don't build a production dependency on it yet, but it's worth testing. The :batch variant on OpenRouter typically means asynchronous batch processing at a reduced rate — good for offline pipelines where you don't need a real-time response.

What This Means Practically

If you're currently using DeepSeek for text-only tasks and you've been waiting for a vision-capable option at Flash pricing, this is the moment to run some tests. Vision at Flash cost is a genuinely useful combination — it opens up document parsing, screenshot analysis, and image classification use cases that previously required bumping up to a pricier tier or switching providers entirely.

The batch variant specifically is worth a look if you have any async workloads. Batch endpoints typically offer meaningful discounts over synchronous calls, and if your pipeline can tolerate latency (nightly report generation, bulk document review, etc.), you can often cut costs significantly just by routing to the batch endpoint.

That said, both models are new and one is explicitly experimental. Before committing:

• Run your own evals on representative tasks from your actual workload
• Check the current pricing on OpenRouter directly — new models sometimes launch at promotional rates that adjust within weeks
• Keep an eye on the exp label; experimental models can change, be deprecated, or graduate to stable naming without much notice

The Quiet Week Is Actually Fine

No price changes across Claude, GPT, Gemini, and Grok this week. That's a stable environment for planning. If you've been putting off a cost analysis because prices keep shifting, this is a decent window to benchmark and lock in assumptions.

DeepSeek continues to move fast on model releases. Two models in two days suggests an active release cadence heading into Q4. Whether that means more aggressive pricing pressure on the other providers is something we'll be watching.

---

I track price changes and new model releases daily across all five major providers. Full live data at llmpricewatch.com.