DeepSeek: DeepSeek v4.1 Flash 1017

What it claims

This listing is the October 17, 2026 snapshot of DeepSeek V4.1 Flash. Beyond the catalog summary, it adds four things an adopter can use:

• Cached input is cheap. Cache reads cost $0.008/M tokens, a tenth of the $0.08/M input rate.
• Image input, text output only. The model accepts images but cannot generate them.
• Vision is native. Visual and text embeddings were trained jointly from the start of pre-training. The earlier experimental V4 Flash Vision Exp added vision after pre-training instead.
• Failover matters. OpenRouter's multi-provider failover materially changes availability, shown under Evidence.

Evidence

The page gives reliability figures. It gives no capability scores.

• Uptime, past 3 days: 100.00%. This means at least one provider was reachable.
• Availability, past 3 days: 99.93%. This means inference was actually returned, with errors and empty responses counted as failures.
• Availability without routing: 95.89%, measured over the last 24 hours. OpenRouter's 99.93% figure relies on rerouting failed requests to another healthy provider.

The page says DeepSeek reports V4.1 Flash beats V4 Pro on performance, speed and task completion time. It gives no benchmarks, numbers or settings for that claim. The Benchmarks, Performance (throughput, latency, time to first token) and Pricing-paid sections are described, but their values do not appear in the fetched text.

The top public apps sending traffic are Hermes Agent (Nous Research) and Claude Code. A figure of 13.7B tokens appears near them, without a stated time window. This traffic suggests the model is mainly used for agentic and coding work.

How to use it

• API: OpenAI-compatible. Swap the base URL and set the model slug.
• Model slug: the page refers to a slug but the fetched text does not print it. The URL path is deepseek/deepseek-v4.1-flash-1017; confirm the slug on the page.
• Tool calling: pass tools and tool_choice.
• Structured output: pass a JSON schema in response_format.
• Routing modes:
• Balanced: price plus speed.
• Nitro: fastest.
• Floor: cheapest.
• Providers: DeepInfra, inference.net, Morph, Cloudflare, DeepSeek, Relace, Baseten and Parasail. You can pin or exclude providers through provider routing.
• Per-provider uptime: available programmatically through the Endpoints API.
• Filters and failover: restrictive provider filters reduce failover. The 95.89% unrouted figure shows what a single-provider setup can expect.

Limits

• Performance claims are unverified here. The claim that it beats V4 Pro is DeepSeek's own, and no numbers support it on this page.
• Short measurement windows. Availability covers only the last 72 hours, and the unrouted comparison covers 24 hours.
• Prices may vary by provider. The listed prices are not broken down per provider. The page notes that the price customers actually pay is often below the listed one because of caching and discounts.
• Missing details. The page states no licence, weights availability or architecture specifics for CED beyond the parameter split.