What Is DeepSeek V4.1 Flash? The Ultra-Cheap AI Model Beating GPT and Claude

DeepSeek released a new AI model called V4.1 Flash on September 10, 2026, and the headline isn’t just what it can do — it’s what it costs. At roughly $0.75 per million tokens off-peak, DeepSeek V4.1 Flash is one of the cheapest frontier-level AI models on the market, yet DeepSeek says it matches or beats OpenAI’s GPT-5.6 Sol and Anthropic’s Claude Opus 5 on several coding and reasoning benchmarks. Here’s what the model actually is, what it costs in plain terms, and how it stacks up against the competition.

Key Takeaways

• DeepSeek launched V4.1 Flash on September 10, 2026 — a 552-billion-parameter mixture-of-experts (MoE) model with native vision support and a 1-million-token context window.
• Off-peak, it costs about $0.15 per million input tokens and $0.60 per million output tokens, undercutting most other frontier-class models while scoring competitively on benchmarks like DeepSWE v1.1.
• The low price comes from new memory-saving techniques — Compressed Sparse Attention 2 and FP4 KV caching — that shrink how much data the model has to store and move per token.
• Claude Opus 5 still leads on some benchmarks (like Terminal-Bench), and GPT-5.6 Sol leads on others (like GPQA Diamond), so V4.1 Flash isn’t “better” across the board — it’s dramatically cheaper for comparable work.
• DeepSeek has already begun retiring its older V4-Flash and V4-Pro models in favor of V4.1 Flash, signaling this is now the company’s default option going forward.

What Is DeepSeek V4.1 Flash?

DeepSeek V4.1 Flash is the Chinese AI lab’s latest large language model, and it’s built as a mixture-of-experts (MoE) model — a design where the system has many specialized “expert” sub-networks but only activates a small fraction of them for any given task, rather than running the entire model every time. The full model has 552 billion parameters (the internal settings a model uses to process information), plus another 196 billion in rarely-used specialized modules, but it only actively uses about 8 billion parameters to read your input and 16 billion to generate a response. That’s what keeps it fast and cheap despite its enormous total size.

V4.1 Flash also supports native vision (it can process images, not just text) and has a 1-million-token context window — meaning it can consider roughly 750,000 words of text in a single conversation, enough to hold an entire codebase or a long document collection at once.

How Much Does DeepSeek V4.1 Flash Cost?

DeepSeek prices the model differently depending on the time of day. During off-peak hours, cached input tokens cost $0.003 per million, uncached input costs $0.15 per million, and output costs $0.60 per million. During peak hours, all of those rates roughly double. Added together, a typical off-peak request (mostly uncached input plus output) works out to about $0.75 per million tokens total — among the lowest prices of any production-grade AI API, behind only a handful of budget models from Meta and Xiaomi.

For context, that pricing structure is part of a much bigger story: 2026 has seen AI labs slash prices repeatedly as they compete for developers, a trend we broke down in our look at the AI price war of 2026. V4.1 Flash is the latest, and one of the most extreme, entries in that race.

How Does It Compare to GPT-5.6 Sol and Claude Opus 5?

On DeepSeek’s own benchmark testing, V4.1 Flash scored 74.2 on DeepSWE v1.1 (a coding benchmark), just ahead of Claude Opus 5’s 74.0 and GPT-5.6 Sol’s 73.0. That’s a near-tie at the top — but a handful of points on one benchmark doesn’t tell the whole story, and each model still has areas where it clearly leads.

Where Each Model Wins

Claude Opus 5 pulled ahead on Terminal-Bench 3.0, a test of how well a model can operate a command-line environment, scoring 43.3 versus V4.1 Flash’s 30.0. GPT-5.6 Sol led on GPQA Diamond (a graduate-level science reasoning test) and SEC-Bench Pro (financial and regulatory document analysis). Meanwhile, on design-related tasks, the analysis firm OpenDesign reported that V4.1 Flash reached “98% of GPT-6 Astra’s quality score at 1.4% of its cost” — a huge efficiency edge, even where it isn’t the single best performer.

The practical takeaway: V4.1 Flash isn’t trying to be the smartest model in every category. It’s optimized for what DeepSeek calls “completed work per cached token, per second, per dollar” — a metric aimed squarely at developers running AI agents and high-volume workloads, where cost efficiency matters as much as raw intelligence.

Why Is It So Cheap?

The savings come down to how the model manages memory. Every AI model has to store a running record of the conversation it’s processing, called a KV cache — think of it as short-term memory the model keeps handy so it doesn’t have to reprocess everything from scratch with each new word it generates. The bigger that cache gets, the more expensive (and slower) the model becomes to run.

DeepSeek built V4.1 Flash around two techniques that shrink this cache dramatically: Compressed Sparse Attention 2, which lets the model focus computing power only on the most relevant parts of a conversation instead of everything at once, and FP4 KV caching, a more compact way of storing that short-term memory using smaller numeric values. Together, DeepSeek claims these cut the cache size to about 890 bytes per token — roughly a quarter the size of its predecessor, V4-Flash, and an eighth the persistent storage of earlier designs. Less data to store and move means lower server costs, which DeepSeek is passing on directly through its pricing.

What This Means for You

If you’re a developer or a business building on AI, V4.1 Flash adds real pressure on OpenAI and Anthropic to keep cutting prices, especially for high-volume use cases like customer support bots, coding assistants, and AI agents that make thousands of API calls a day. DeepSeek has already deprecated its older V4-Flash model and started routing V4-Pro traffic to V4.1 Flash as of September 14, 2026, so anyone already using DeepSeek’s API is likely using this model already, whether they realize it or not.

It’s also part of a broader shift where the most cost-efficient models increasingly come from outside the “big three” US labs — we’ve seen the same pattern with other open-weight models climbing the leaderboard over the past year. If you’d rather sidestep cloud pricing altogether, it’s also worth knowing that smaller open-source models can now run directly on a laptop or phone, though a model the size of V4.1 Flash still requires DeepSeek’s own cloud infrastructure. For everyday users rather than developers, the direct impact is smaller — but every price drop like this one tends to filter down into cheaper subscriptions and free tiers across the apps you already use.

Frequently Asked Questions

Sources: VentureBeat, OpenRouter.