This shortlist covers five model listings that pair top-tier quality scores with very low token prices. Four of them belong to the DeepSeek Flash family and share a quality score of 100.0, which is the top of the scale in our data. The fifth, GLM 5.2 from Z.ai, scores 96.6 and has a very different cost profile. The choice between these models mostly comes down to how much output your workload generates, whether you need a pinned version, and whether you need a 1M-token context window.
The Pricing at a Glance
| Model | Input $/M tokens | Output $/M tokens | Quality score |
|---|---|---|---|
| DeepSeek V4.1 Flash 1017 | 0.006 | 0.50 | 100.0 |
| DeepSeek Flash Latest | 0.006 | 0.50 | 100.0 |
| DeepSeek V4.1 Flash 1017 (batch) | 0.011 | 0.20 | 100.0 |
| DeepSeek V4.1 Flash (batch) | 0.011 | 0.20 | 100.0 |
| Z.ai GLM 5.2 | 0.006 | 6.80 | 96.6 |
Input pricing is almost negligible across the board. Every model here charges either $0.006 or $0.011 per million input tokens, so feeding large prompts or documents costs very little. Output pricing is where the real differences show up, and it should drive your decision.
DeepSeek V4.1 Flash 1017: The Default Pick
DeepSeek V4.1 Flash 1017 is the October 17 snapshot of V4.1 Flash, served under the `deepseek-flash` API model name. It is a sparse mixture-of-experts model, and its listing notes that it activates 8B parameters on input. It is the first DeepSeek model built on the company's Causal Encoder-Decoder (CED) architecture.
With a quality score of 100.0, $0.006 per million input tokens, and $0.50 per million output tokens, this is the strongest quality-per-dollar option for interactive work. A request with one million input tokens and one million output tokens costs about $0.51 in total. That makes it a sensible default for chat applications, coding assistants, summarization, extraction, and most general-purpose production traffic.
Use the dated 1017 snapshot when you need reproducible behavior. Pinning to a specific snapshot means your prompts, evaluations, and output parsers will not shift underneath you when a newer version is released.
DeepSeek Flash Latest: Same Price, Moving Target
DeepSeek Flash Latest always redirects to the newest model in the DeepSeek Flash family. Right now it carries the same prices and the same 100.0 quality score as the 1017 snapshot.
This alias suits prototypes, internal tools, and teams that want automatic upgrades without changing code. It is a poor fit for production systems with strict output formats or regression test suites, because the underlying model can change without notice. A practical pattern is to build against Flash Latest during development and switch to a dated snapshot before you ship.
The Batch Listings: Cheapest Output in the Group
Both batch listings, DeepSeek V4.1 Flash 1017 (batch) and DeepSeek V4.1 Flash (batch), are priced at $0.011 per million input tokens and $0.20 per million output tokens, with a quality score of 100.0.
The pricing structure here is unusual. Input costs nearly twice as much as the standard listing, while output costs 60% less. The breakeven depends on your input-to-output ratio. For every million output tokens you generate, batch saves $0.30, and for every million input tokens you send, it costs an extra $0.005. That means batch is cheaper unless you send more than 60 input tokens for every output token. Very few workloads exceed that ratio, so batch wins on price for nearly all jobs that can run through it.
Good candidates include bulk content generation, dataset labeling, translation runs, and offline report generation, where output volume is high and you can submit work as a batch rather than waiting on individual responses. The 1017 batch variant is the pinned choice. The unversioned batch listing is described as V4.1 Flash generally, so prefer the dated one if consistency across runs matters.
Z.ai GLM 5.2: When You Need Long Context and Reasoning
GLM 5.2 is a large-scale reasoning model from Z.ai with text input and output and a 1M-token context window. Its listing positions it for long-horizon agent workflows and project-level software engineering.
Its input price matches DeepSeek Flash at $0.006 per million tokens, but output costs $6.80 per million, which is 13.6 times the standard Flash rate and 34 times the batch rate. Its quality score of 96.6 is also slightly lower than the Flash family's 100.0. On pure quality per dollar, it is not competitive with DeepSeek Flash.
GLM 5.2 earns its place in a narrower set of situations. The 1M-token context window and cheap input pricing make it workable for loading an entire codebase or a large document set into a single prompt, as long as the responses stay short. An agent that reads a lot and writes concise patches or answers can keep output costs under control. If your workload generates long outputs, the price gap grows quickly and Flash becomes the better choice.
How to Choose
Start with DeepSeek V4.1 Flash 1017 for any interactive or real-time workload. Move high-volume, output-heavy jobs that can be submitted as batches to the 1017 batch listing to cut output costs by 60%. Use Flash Latest only where automatic model changes are acceptable. Reserve GLM 5.2 for long-context, read-heavy agent and software engineering tasks, and set output token limits on those calls so a single verbose response does not cost more than an entire batch run on Flash.