Under the model name deepseek-flash, DeepSeek has pushed DeepSeek-V4.1-Flash-1017 live across its website, app and API — a fresh post-training checkpoint of the Flash line.
DeepSeek describes the step as a Pareto improvement, with every one of the 11 benchmarks it reported showing gains over the September V4.1-Flash, by margins between 2.0 and 11.1 points. Search and Fetch APIs opened to all developers in the same release.
The architecture behind the cost story
Total capacity is 552B parameters plus 196B Engram parameters. Active per token: 8B during prefill and 16B during decode. Context reaches 1M tokens, and output can run to 384K tokens.
A causal encoder-decoder keeps prefill compute nearly halved by projecting global key-value state out of the final encoder layer. Main and indexer keys are shared across layers through a compressed sparse attention scheme, and an FP4 main cache squeezes global KV to roughly 890 bytes per token — a quarter of what V4-Flash needs.
The benchmarks are DeepSeek’s own numbers. On DeepSWE v1.1 the new build scores 77.0, against 74.2 for its predecessor and 74.0 for Claude Opus 5 in the same table. Terminal-Bench 4.0 rises to 42.3 from 31.2, though Opus 5 still leads at 51.8. Apex Bench lands at 72.1 and Agents’ Last Exam at 35.4.
Ranked results arrive from the Search API with rerank scores and authority tiers attached; Fetch hands back Markdown. Successful requests cost $0.75 per 1,000, and failures or empty results are free.
Usage through DeepSeek Harness, Claude Code or Codex carries no charge through November 1. Standard API pricing runs $0.30 per 1M cache-miss input tokens and $1.20 per 1M output tokens during peak hours, cut in half off-peak.
