DeepSeek replaced the model behind its web app, mobile app, and API with V4.1-Flash-1017 on October 17, 2026. The official table puts DeepSWE v1.1 at 77.0, ahead of Claude Opus 5 at 74.0, though Terminal-Bench 4.0 still trails Opus 5 at 42.3 versus 51.8. World of AI's review lists the price at $0.14 per million input tokens and $0.28 per million output tokens, about one percent of Opus 5.5's input price, and its five-scenario hands-on found the strengths in code generation, refactoring, and root-cause work, with terminal discipline, very long log retrieval, and UI taste still favoring a flagship. DeepSeek calls the new behavior wish-style programming: you describe the goal, and the model plans, codes, runs, checks, and fixes by itself, a capability it ties to reinforcement learning with self-feedback. Launch demos include a Forbidden City 3D scene, the NEON VELOCITY browser racing game, a playable Godot remake of Portal 2, and a 3D volcano eruption simulation. The Search and Fetch API endpoints are listed at $0.75 per 1,000 successful requests (≈¥5) and stay free inside DeepSeek Harness until November 1, 2026.
On October 17, 2026, DeepSeek swapped the model behind its web app, mobile app, and API to V4.1-Flash-1017. The name carries the release date, and the model list briefly leaked before the announcement, which is how regular users spotted it first. This review compiles the third-party coverage that followed: the official table as republished by two Chinese outlets, a five-scenario hands-on from World of AI, and demo tests from an independent developer in China.
The toolchain needs a matching update. DeepSeek Harness 0.2.1-rc.1 or newer picks up a profile tuned for this model, with the same tuning covering the Tencent CodeBuddy and WorkBuddy agent frameworks. QbitAI's launch coverage adds that eligible first-time Harness logins receive a 6 CNY credit, expiring after 7 days while supplies last.
How do the benchmarks look?
The official table is strongest where the work repeats: code generation and agent runs. Against the previous V4.1-Flash, DeepSWE v1.1 climbs from 74.2 to 77.0, Terminal-Bench 4.0 jumps from 31.2 to 42.3, Automation-Bench V1.0.6 goes from 54.8 to 60.5, and Apex Bench v1.1 rises from 64.6 to 72.1. Against other vendors the picture splits: the 77.0 on DeepSWE v1.1 lands above Claude Opus 5 at 74.0 and GPT-5.6-Sol at 73.0, while Terminal-Bench 4.0 stays well behind Opus 5 at 51.8, HLE sits at 42.6/44.0 against 56.3, and NL2Repo-Bench trails at 68.2 versus 75.3.
The table's own footnotes name the evaluation settings: Terminal-Bench 4.0 ran on the Claude Code harness, DeepSWE on Mini-SWE-V2, and ZeroBench-main on Pass@5. The table compares Claude Opus 5, not Opus 5.5, so the two are kept separate throughout this review.
DeepSeek also published a cost curve for DeepSWE v1.1 that plots pass rate against average cost per task. The 1017 model sits in the top left corner at 77.0 percent and roughly 1 to 2 CNY per task, with Claude Opus 5.5, GPT-6 Astra, Gemini 3.8 Flash, GLM-5.3, and Kimi K3 spread to the right. The company sums up the release as a new performance-cost Pareto frontier across several agent harnesses, ahead of Fable 5, Opus 5, and GPT-6 Astra. That is the vendor's own claim as relayed in Chinese coverage, and SaveMyToken has not rerun any of it.
What did third-party testing find?
World of AI ran the model through five scenarios. The short version: it wins the high-volume work and stumbles on the finishing touches.
What went well
- A 3D voxel sandbox game generated in one shot. It took 18 seconds to build and ran at a stable 60 FPS.
- A refactor of 1,500 messy lines of FastAPI. It finished in 24 seconds, about half of what the previous generation needed, produced 39 percent less code, and broke no imports.
- An async deadlock hunt that pinned the root cause in an 8-line diff.
- A full-stack kanban board built against Opus 5.5. Every requested feature made it in.
What went badly
- Needle-in-a-haystack retrieval: asked to find the one relevant line inside 200,000 tokens of logs, it lost to Gemini 4 Argon.
- Terminal discipline: a Docker database rollback got done, but it typed a few redundant commands that GPT-6.1 Sol did not.
- Zero-shot UI taste: on the same kanban brief, the interface looked plain next to Opus 5.5's, and the micro-animations were missing.
Put together: code generation, refactoring, and root-cause work are the strengths. Terminal discipline, long-log retrieval, and UI taste are where it pays to keep a flagship.
What is wish-style programming?
It is DeepSeek's name for the new workflow: you describe the thing you want, and the model handles planning, coding, running, checking, and fixing on its own.
Before, a full project meant breaking the work down yourself, feature by feature, and feeding the pieces to the model one at a time. DeepSeek credits the change to reinforcement learning with self-feedback. The model reviews its first version, catches what is wrong, and goes back to fix it instead of handing over a half-built artifact.
What can it actually build?
The launch demos lean toward two genres: worlds you can move through, and games that actually run. One renders the Forbidden City in 3D with a free camera. NEON VELOCITY is a racing game that runs in a browser. There is a playable Godot remake of Portal 2, and a dynamic 3D world that walks from the Stone Age into modern times.
An independent developer from China tested the release build on Harness v0.2.1-rc.2 at Max reasoning. His staple test is a 3D volcano eruption, 14 seconds long, with lava particles, a multi-stage blast, smoke, lighting, and heat haze all present. His verdict, quoted from the writeup: "No exaggeration, it scared me."
How cheap is it?
World of AI's review lists the model at $0.14 per million input tokens and $0.28 per million output tokens. Claude Opus 5.5 sits at $15 and $75 on the same units, so the input price works out to about one percent of Opus 5.5's. The official change log announced no price change with the model swap, and SaveMyToken has not confirmed the list price against DeepSeek's official price page.
DeepSeek also opened its own search API with two endpoints, Search and Fetch. The official guide lists both at $0.75 per 1,000 successful requests (≈¥5), billed per successful request; failed calls and timeouts are not billed. QbitAI and the independent writeup quote the same terms as 5 CNY per 1,000 calls. Server-side tool calls inside DeepSeek Harness stay free until November 1, 2026.
If you want to try this model inside your own coding setup, have a look at our coding plan.
Should you switch or stay?
Move high-frequency agent loops, everyday backend work, and quick prototypes to V4.1-Flash-1017. These jobs run in volume, they tolerate the occasional sloppy step, and the price gap compounds. The reviewer's math has a heavy monthly bill dropping from thousands of dollars to tens.
Keep Opus 5.5 or an equivalent flagship for production terminal automation, design-sensitive frontend work, and root-cause hunts across very large codebases or log dumps. Those are exactly the three places where hands-on testing found the new model trailing.
They are not ranked against each other. They optimize for different points on the same cost and capability curve.
Frequently asked questions
How much does DeepSeek V4.1-Flash-1017 cost?
$0.14 per million input tokens and $0.28 per million output tokens, according to World of AI's review. Claude Opus 5.5 lists at $15 and $75 on the same units.
Can it replace Claude Opus 5.5?
For high-volume work such as agent loops, backend refactors, and prototypes, hands-on testing says yes. Three areas still favor Opus 5.5: production terminal automation, design-sensitive frontend work, and retrieval across very long logs.
What is wish-style programming?
You state the goal and the model handles planning, coding, running, checking, and fixing. DeepSeek ties the behavior to reinforcement learning with self-feedback.
Is the DeepSeek Search API free?
Server-side tool calls inside DeepSeek Harness are free until November 1, 2026. The listed price is $0.75 per 1,000 successful requests (≈¥5), and failed or timed-out calls are not billed.
Which DeepSeek Harness version do I need?
0.2.1-rc.1 or newer. Older builds do not pick up the tuned profile for this model.
Sources and references
- QbitAI: DeepSeek launches the new model ↗ — wish-style programming framing, the four official demo clips, and the Search API and credit terms
- An independent developer from China: V4.1 Flash hands-on ↗ — official benchmark table and Pareto chart reposts, plus the volcano and other demo recordings
- World of AI: Flash v2 hands-on review ↗ — five-scenario testing, the $0.14/$0.28 list price, and the Opus 5.5 comparison