DeepSeek Updates V4.1-Flash Model With Faster Task Completion and No Price Increase

Quick Facts

• DeepSeek-V4.1-Flash-1017 scored 77% on the DeepSWE v1.1 software engineering benchmark, up from 74.2% for the September release and above Claude Opus 5's 74%.
• API pricing is unchanged at $0.30 per million cache-miss input tokens and $1.20 per million output tokens, with off-peak rates at half those figures.
• DeepSeek also launched Search and Fetch APIs alongside the model, priced at $0.75 per 1,000 requests, with free usage through Nov. 1 for calls made via DeepSeek Harness, Claude Code, or OpenAI Codex.

DeepSeek released an updated AI model on Oct. 18, 2026, continuing its push against rivals Anthropic and OpenAI. The company announced DeepSeek-V4.1-Flash-1017 via posts on X late Saturday, saying the update delivers stronger benchmark performance, faster task completion, and shorter outputs than the prior version.

The release comes roughly five weeks after the original V4.1-Flash launched in September as the smallest model in a new architecture family. DeepSeek's model card on Hugging Face describes it as a minor update. The company did not specify how much faster the model runs.

Architecture Unchanged

The 1017 update preserves the architecture introduced in September without modification. According to the Hugging Face model card, V4.1-Flash-1017 keeps the same 552-billion-parameter backbone, activating roughly 8 billion parameters per token during prefill and 16 billion during decoding, plus 196 billion parameters of Engram conditional memory.

The model uses a Causal Encoder-Decoder design organized as a 20-layer encoder followed by a 20-layer decoder. Its global key-value cache compresses to 890 bytes per token, four times smaller than the prior V4-Flash generation. Analysts believe the performance gains came from additional reinforcement learning post-training. Configuration files, tokenizer, and prompt encoding are unchanged, so existing integrations require no code modifications.

Benchmark Results

DeepSeek published results across 11 benchmarks. The model card shows V4.1-Flash-1017 scored 77.0 on DeepSWE v1.1, a software engineering test, compared to 74.2 for the September version and 74.0 for Claude Opus 5.

The new model also outperformed Claude Opus 5 on Automation-Bench v1.0.6 (60.5 vs. 50.3), Agents' Last Exam (35.4 vs. 28.6), and Apex Bench v1.1 (72.1 vs. 65.8). Claude Opus 5 leads on seven of the 11 benchmarks DeepSeek published, including Humanity's Last Exam, where Anthropic's model scores 56.3 against V4.1-Flash-1017's 42.6.

Built for Coding Agents

DeepSeek tuned the update for agentic workflows. The prefill architecture makes the model particularly efficient for coding agents that issue long context inputs followed by short outputs, a pattern common in automated software development pipelines.

Named integrations include DeepSeek's own Harness tool and two Tencent products: CodeBuddy, a coding assistant, and WorkBuddy, a desktop agent for office tasks. DeepSeek recommends upgrading to Harness version 0.2.1-rc.1 or later. The company also noted it uses reinforcement learning with self-feedback to support end-to-end project development spanning game creation and 3D modeling.

Pricing and API Access

The updated model is live in DeepSeek's web and mobile apps. API customers reach it through the existing deepseek-flash model ID at the same price: $0.30 per million cache-miss input tokens and $1.20 per million output tokens during peak hours, with off-peak rates cut in half. Model weights are available on Hugging Face under an MIT license.

DeepSeek also launched a Search API and a Fetch API alongside the model update. The Search API returns up to 50 ranked web results per query. The Fetch API converts a single web page to Markdown. Both cost $0.75 per 1,000 successful requests, though calls made through DeepSeek Harness, Anthropic's Claude Code, or OpenAI's Codex are free through Nov. 1, 2026. Both APIs are capped at five queries per second per account.