DeepSeek V4.1 Flash 1017: Pricing, Setup, Benchmarks (2026)

On 17 October 2026, DeepSeek announced DeepSeek V4.1 Flash 1017 — an updated checkpoint of its DeepSeek V4.1 Flash coding model. It scores 77.0 on DeepSWE v1.1, ahead of Claude Opus 5 and GPT-5.6 Sol in DeepSeek’s own table, and it finishes a typical coding task for about 20 cents — roughly one-tenth to one-fiftieth of what US frontier models charge for the same job. This guide covers what shipped, the pricing maths, how to call it from your agent, and the caveats you should know. Facts verified 18 October 2026.

Table of Contents

What actually shipped on October 17

DeepSeek V4.1 Flash 1017 is an updated post-training checkpoint of the V4.1 Flash that launched on 10 September 2026 — not a new architecture. DeepSeek calls the release “a Pareto improvement in performance and cost”: stronger benchmark scores, faster task completion, and shorter outputs. That last point is the one that matters most for your bill, and we’ll come back to it.

Three practical details worth knowing:

It is served under the existing deepseek-flash API name — same name, same price. You get the new checkpoint by default.
The old model names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work, but they’re now served by the new checkpoint.
It is live on DeepSeek’s website, app, and API right now.

The release also bundles DeepSeek’s training method, which the company calls reinforcement learning with self-feedback: the model grades its own work using internal confidence and self-written critiques instead of human raters. DeepSeek’s demo examples range from game creation to 3D modelling — including a 3D model of Beijing’s Forbidden City and a recreation of the puzzle game Portal 2 in the Godot engine. The model is explicitly tuned for agentic coding setups: DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy.

Why it’s cheap: the architecture in plain English

The headline specs of DeepSeek V4.1 Flash:

552B-parameter backbone (plus ~196B “Engram” parameters), sparse mixture-of-experts
8B parameters active per token while reading (prefill), 16B while writing (decode)
1M-token context window, up to 384K output tokens
Native image input (multimodal), text output
890 bytes of KV cache per token — about a quarter of the previous Flash generation
Configurable reasoning effort from 1 to 100 (API tiers low/high/max map to 50/75/100)

The architecture is called Causal Encoder-Decoder (CED): a 40-layer transformer split into a 20-layer encoder and 20-layer decoder, so the decoder projects its cache from the encoder output instead of building one cache per layer. Combined with FP4-quantised keys/values and Compressed Sparse Attention 2, the memory cost of holding a long conversation drops dramatically. That matters because agent runs are input-heavy: every loop step re-reads the whole transcript, the repo context, and the tool history. Shrinking the per-token cache is what lets a 1M-context agent stay affordable.

The deeper trick is the shorter outputs. Every word a model writes is a word you pay for, and agent loops burn through hundreds of generations per task as the model tests, fails, and retries. Training the model to say less is the cheapest performance gain there is.

The benchmark picture (and the grain of salt)

All scores below are DeepSeek’s own figures. Independent evaluator Vals AI ranks V4.1 Flash the number-one open-weight model on its index — but also found DeepSeek’s headline coding score was inflated by its own harness. External verification of the 1017 checkpoint’s numbers hasn’t landed yet. Keep that in mind while reading this table.

Benchmark	V4.1-Flash-1017	Prev V4.1-Flash	Claude Opus 5	GPT-5.6 Sol
DeepSWE v1.1	77.0	74.2	74.0	73.0
Terminal-Bench 4.0	42.3	31.2	51.8	39.9
Automation-Bench	60.5	54.8	50.3	45.8
Agents’ Last Exam	35.4	31.8	28.6	26.7
Apex Bench	72.1	64.6	65.8	54.3
ProgramBench	29.1	20.3	37.0	23.0
HLE (no tools)	42.6	36.8	56.3	44.5

The pattern is consistent: 1017 leads on agentic, end-to-end software engineering tasks (DeepSWE, Automation-Bench, Apex Bench, Agents’ Last Exam) while Claude Opus 5 still leads on raw reasoning and terminal execution (HLE, Terminal-Bench 4.0, ProgramBench). That matches the training method — RL with self-feedback for whole-project work, not for puzzle benchmarks. If your workload is “build the feature, run the tests, fix the failures,” this is the model optimised for exactly that loop. If it’s careful reasoning on a small problem, the frontier models still win.

What a coding task costs in real money

Per-token pricing on DeepSeek’s API:

Input: $0.15 per 1M tokens off-peak, $0.30 peak (cache-miss)
Output: $0.60 per 1M tokens off-peak, $1.20 peak
Cache hits: $0.003/$0.006 per 1M tokens — nearly free, and this is what makes long agent sessions viable

Per-token prices don’t tell you what a job costs, so DeepSeek published a cost chart: a typical coding task on V4.1-Flash-1017 lands at roughly $0.10–$0.20. The previous checkpoint cost about $0.50 a task. On the same chart, Claude Opus 5, GPT-6 Astra, and Gemini 3.8 Flash land between $2 and $10+ per task. Even allowing for generous vendor cherry-picking, that’s a 10–50x cost gap on agent workloads.

Why the gap is structural, not promotional:

Shorter outputs — fewer generated tokens per task, directly lowering the most expensive line item.
Faster task completion — fewer wasted loop iterations.
Cache economics — with 890 bytes per token of KV cache and cache-hit pricing at $0.003/1M, re-reading a huge transcript each turn barely moves the needle.

If you run agents in production, this changes the default calculus: the cheap model is no longer the one you use for trivial tasks and abandon when things get hard. It’s now competitive on the agentic benchmarks where it counts.

Calling it: API, Claude Code, and Harness setup

You call the model as deepseek-flash through DeepSeek’s OpenAI-compatible or Anthropic-compatible endpoints:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.deepseek.com",
    api_key="sk-your-deepseek-key",
)
response = client.chat.completions.create(
    model="deepseek-flash",
    messages=[{"role": "user", "content": "Refactor this module to async/await"}],
)

To point Claude Code at DeepSeek’s API, set the Anthropic-compatible base URL and your key, then map every model slot explicitly to deepseek-flash:

export ANTHROPIC_BASE_URL="https://api.deepseek.com/anthropic"
export ANTHROPIC_AUTH_TOKEN="sk-your-deepseek-key"
export ANTHROPIC_MODEL="deepseek-flash"
export ANTHROPIC_DEFAULT_OPUS_MODEL="deepseek-flash"
export ANTHROPIC_DEFAULT_SONNET_MODEL="deepseek-flash"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="deepseek-flash"
export CLAUDE_CODE_SUBAGENT_MODEL="deepseek-flash"

Set all of them explicitly — an unset Opus slot maps to a different model. Then run /status inside Claude Code to confirm the base URL and model before doing real work. One known limitation: PDFs sent as document blocks and MCP tool-use blocks are not supported on DeepSeek’s Anthropic endpoint, so test MCP-heavy sessions before migrating a complex setup.

If you use DeepSeek’s own agentic setup, DeepSeek recommends DeepSeek Harness v0.2.1-rc.1 or later for the best experience with the 1017 checkpoint, alongside Tencent CodeBuddy and WorkBuddy. And if you manage context costs in a multi-model pipeline, pair the cheap flash model with the two-model pattern — a strong model for review, a cheap one for generation — which we covered in our AI code review prompts guide.

The new Search and Fetch APIs

Alongside the model, DeepSeek opened its Search and Fetch web APIs to every developer — a first-party web layer behind the same API key, which is exactly what agent builders need for research and grounding tasks.

Endpoints: POST /v1/tools/search and POST /v1/tools/fetch on api.deepseek.com
Search: 1–50 results per query, optional page content, operators like site:, lang:, after:, before:; each result carries a rerank score and an authority tier
Fetch: returns a single page as Markdown
Price: $0.75 per 1,000 successful requests per endpoint — empty results and failures are free
Rate limit: 5 QPS per account per endpoint
Free window: server-side calls made through DeepSeek Harness, Claude Code, or Codex stay free until 1 November 2026

DeepSeek says its in-house engine already serves hundreds of millions of requests daily. For agent builders, the practical play is obvious: let the cheap model do the retrieval and drafting, and use the Fetch API to ground citations instead of paying a frontier model to browse.

The open-weights story

The September V4.1 Flash weights are on Hugging Face under the MIT licence (deepseek-ai/DeepSeek-V4.1-Flash), free for commercial use. On the 1017 checkpoint specifically, DeepSeek has not disclosed whether or when its weights will be released — so treat the open-weights story as “the base model is MIT; the 1017 checkpoint is API-only for now.” It’s also live as a managed API on Google’s Model Garden, and third-party inference providers (Baseten, among others) serve it.

If self-hosting is your lane, remember what “open weights” really costs you here: a 552B-parameter MoE needs serious VRAM even sparsely activated, which is why the hosted API at $0.15/1M off-peak is the realistic default for most teams. The MIT licence matters most for adaptation and distillation research, not for running it on a home server.

What to watch out for

No model release is all upside. The honest caveats:

Vendor-reported benchmarks. Every score in the table above came from DeepSeek. Vals AI found the headline coding score was inflated by DeepSeek’s own harness, and CHOSUNBIZ notes external verification of the 1017 numbers hasn’t happened yet. Wait for independent agentic evals before betting production on the 77.0 DeepSWE figure.
Reasoning ceiling. 42.6 on HLE without tools versus 56.3 for Opus 5 is a real gap. For tasks where correctness on a small, hard problem matters more than speed across a big one, the frontier models still win.
Harness-tuning bias. RL with self-feedback optimised for DeepSeek’s own Harness loop. Performance in your agent’s different loop may vary — measure on your own evals, not theirs.
Endpoint limitations. The Anthropic-compatible API lacks MCP tool-use blocks and PDF document blocks. If your workflow depends on them, you’ll hit walls.
1017 weights undisclosed. Don’t plan a self-hosted 1017 deployment until DeepSeek confirms the weights release.

None of that changes the bottom line: for agentic coding loops — where cost per task was the blocker — a 20-cent task at near-frontier agentic quality is the most consequential price move in AI coding this year. DeepSeek’s post-training lead Wu Yu called it “the era of democratized wish-driven coding,” and for once the marketing line undersells the maths.

DeepSeek V4.1 Flash 1017: What Developers Should Know About the 20-Cent Coding Model (2026)