DeepSeek V4.1-Flash-1017 Full Switch: 11 Benchmarks Up, How to Cut Agent Costs?

TL;DR

DeepSeek completed the V4.1-Flash-1017 full switch on October 17, 2026, with web, app, and API all updated simultaneously. This is not a major architectural upgrade but a "Pareto improvement of performance-cost": all 11 benchmarks up, pricing unchanged, and a self-built search API opened alongside. For developers relying on agents that make multi-round tool calls, the real cost isn't the per-response price — it's total task runtime and output token volume. This update targets exactly those two dimensions.

What Changed

Version & Architecture

V4.1-Flash-1017 is a new checkpoint of V4.1 Flash, not a new architecture. Core specs:

11 Benchmarks All Improved

Compared to the September 10 version, all 11 benchmarks rose by 2.0-11.1 points. The two largest gains:

Terminal-Bench tests terminal execution — AI actually completing tasks in the command line rather than generating code snippets. ProgramBench tests program construction/building. These two large gains signal a shift from "code snippet generator" toward "project-level build tool."

Self-feedback RL is especially effective at improving "can the task run to completion" metrics, which is why Terminal-Bench and ProgramBench saw the biggest gains. However, the official statement is candid: RL can teach a model to do the job right, but can't teach a model to know more — knowledge-depth benchmarks like HLE improved marginally.

Comparison with Claude Opus 5

Official claim: "Surpasses Claude Opus 5 in 4 agentic tests." But the inverse is clearer: lost 7 out of 11.

• Won (4 items): concentrated in bug fixing and workflow execution
• Lost (7 items): concentrated in knowledge depth and complex reasoning

Largest deficit — HLE (Humanity's Last Exam): DeepSeek 42.6 vs Claude Opus 5's 56.3, a 13.7-point gap. HLE tests cross-disciplinary hard knowledge where tools cannot help.

Model profile summary: not the smartest one; the one that can finish the work and is especially cheap.

Pricing

Search API: The Real Strategic Move

Endpoints & Pricing

Technical Specs

• Previously served as the underlying search for DeepSeek's web and API
• Processes hundreds of millions of daily requests globally
• Indexes hundreds of billions of web pages
• Multi-language support

Quality Filtering

Official statement emphasizes strict filtering against: SEO spam/gaming, GEO manipulation, low-quality AIGC content.

Free Period

• Currently free in DeepSeek Harness, Claude Code, and Codex
• Billing starts November 1, 2026, 00:00

Strategic value: a cheap, capable code model plus a clean self-owned web layer — one API key is all you need. This eliminates the entire hassle of integrating a third-party search service.

Harness 0.2.1 & "Wishful Programming

DeepSeek Harness must be upgraded to 0.2.1-rc.1 or above to benefit from the 1017 model's specialized optimizations.

"Wishful programming" is the new concept: describe what you want in natural language, and the model builds the entire project itself — moving from "code snippet generation" to "project-level construction." Official demonstrations included Forbidden City 3D modeling, a browser-based racing game, and a Godot engine recreation of Portal 2 — but note these are hand-picked cases by the official team.

Core workflow: model writes → self-checks → finds issues → reworks → delivers again. When agents make multi-round tool calls, the bill gets amplified, and the per-response price no longer represents real cost — which is why this update specifically targets Harness optimization.

Open Source

How to Use: Developer Action Guide

1. If you're using deepseek-flash

You automatically get the new version — no action needed. Price unchanged, capability increased. No reason not to switch.

2. If you're using Harness

Upgrade to 0.2.1-rc.1 or above. The new version includes specialized optimizations for the 1017 model. Search and fetch are already available within Harness.

3. If you're running multi-round agent calls

Focus on the Terminal-Bench and ProgramBench improvements — these directly measure your agent's task completion rate in terminal environments. The +35.6% and +43.3% relative gains mean your agent's real-world task completion rate will see noticeable improvement.

4. If you're integrating search

The search API is currently free in Harness, Claude Code, and Codex. Billing starts November 1 at CNY 5/1K calls. If your agent needs web search capability, this is one of the cheapest self-owned search APIs available.

5. Don't be fooled by "all up

The 11-item improvement is vs. its own September version, not vs. competitors. It competes with Claude Opus 5 on "doing work" but still lags on "deep thinking." Align model choice to the task, not to the benchmark score.

6. All numbers are self-reported

As of this article's publication, no third-party leaderboard has independently scored the 1017 version. DeepSeek's own table notes the data is self-reported. Self-reported data isn't necessarily fake, but it is cherry-picked. Real-world feel requires running it yourself.

Industry Context

The same week, the AI industry was shifting from "who's smarter" to "who can make money": Anthropic Q3 revenue $16.8B, adjusted profitable, but IPO valuation expectation cut from $2T to $1.5T; Zhipu and MiniMax secondary market valuations shrunk significantly; Broadcom CEO Hock Tan stated "AI investors are essentially betting in a casino"; Nvidia lost approximately 4 trillion RMB in market value over 9 days.

DeepSeek's positioning: same-price-better model + free search layer until November 1 — not a technology flex, but strategic positioning. Competition is shifting from parameter curves to cost curves.