On 10 September 2026, DeepSeek released V4.1 Flash, a new open-weight model under the deepseek-flash API name. The release itself would be routine — DeepSeek ships often — except for what happened four days later: from 04:00 UTC on 14 September, every API request sent to deepseek-v4-pro, DeepSeek's previous flagship, is now silently served by V4.1 Flash instead, billed at Flash's much lower prices. If your product calls DeepSeek's API by model name rather than pinning a specific version, the model actually answering your requests changed underneath you without you doing anything.
What actually shipped
V4.1 Flash is a 552-billion-parameter model built on a causal encoder-decoder design: a 20-layer encoder reads the input using roughly 8 billion active parameters, and a 20-layer decoder writes the response using roughly 16 billion active parameters. It supports a 1-million-token context window with up to 384,000 tokens of output, ships with native vision support, and compresses its key-value cache to around 890 bytes per token — a quarter of the previous Flash model's footprint, which is what makes long-context requests meaningfully cheaper to serve. The weights are released under an MIT license, meaning you can download and self-host the model with no usage restrictions, not just call it through DeepSeek's API.
On DeepSeek's own benchmark set, V4.1 Flash beats the older, larger V4 Pro model it replaces: 90.6 on Terminal-Bench 2.1, 74.2 on DeepSWE, and a 3,471 Codeforces rating. On raw reasoning it is more modest — 36.8 on Humanity's Last Exam, against 56.3 for Anthropic's Claude Opus 5 — so this is not a claim of frontier-beating intelligence. It is a claim that, for coding and agentic tasks specifically, a smaller and dramatically cheaper model now outperforms DeepSeek's own previous flagship.
What it actually costs
Pricing runs on two schedules. During off-peak hours (outside 01:00–04:00 and 06:00–10:00 UTC on weekdays), input costs $0.15 per million tokens on a cache miss and $0.003 per million on a cache hit, with output at $0.60 per million. During peak hours, those roughly double: $0.30 per million input, $0.006 per million on a cache hit, and $1.20 per million output. For comparison, that peak-rate output price is roughly a fortieth of GPT-6 Astra's $50-per-million-output rate. If your workload can tolerate scheduling around DeepSeek's peak window — which lines up with Chinese business hours — the off-peak rate is close to the cheapest frontier-adjacent inference available anywhere right now.
The part worth actually paying attention to
The pricing is the headline, but the auto-routing decision is the more consequential detail for anyone with production traffic. DeepSeek did not deprecate deepseek-v4-pro and force a migration; it kept the model name live and pointed it at a different model, at different pricing, with different latency and output characteristics, on a fixed date and time it published in advance. That is a defensible engineering decision — DeepSeek's position is that V4.1 Flash "comprehensively surpassed" V4 Pro, so routing old requests to the better, cheaper model is a straightforward win for most callers. But it is also a decision made on your behalf, inside a dependency you do not control, that changed your product's actual behaviour without a version bump you would necessarily notice.
If any part of your product calls the DeepSeek API by the deepseek-v4-pro model name, check three things this week. First, confirm your billing reflects the new Flash pricing rather than the old V4 Pro rate — if a rate you were budgeting against just dropped, that is worth knowing rather than discovering at your next invoice. Second, spot-check output quality and format on a handful of your actual production prompts; a model swap under a familiar name is exactly the kind of change that silently degrades a specific use case your automated evals were not built to catch, even when it improves average benchmark performance. Third, if latency or output length matters to your product, verify both are still within the bounds you designed around, since Flash-tier models are generally tuned for throughput over the kind of long, deliberate output a "Pro" name might have implied.
More broadly, this is a reminder to pin model versions explicitly wherever a routing change would actually hurt you, and to treat "same model name, unannounced-to-you swap" as a real operational risk with any API provider that reserves the right to do it — not just DeepSeek.
The bottom line
DeepSeek's V4.1 Flash is a genuinely cheap, MIT-licensed, million-token-context model that beats the flagship it quietly replaced on coding and agentic benchmarks. But the more important fact for anyone building on DeepSeek's API is the auto-routing: since 04:00 UTC on 14 September, deepseek-v4-pro requests are served by an entirely different model at different prices. If that model name appears anywhere in your stack, check your bill and spot-check your outputs this week rather than assuming nothing changed.