Notes from Việt Nam on law, technology, languages, history, culture and travel.

Twenty Cents a Task: DeepSeek's V4.1-Flash-1017 and the Shrinking Price of a Coding Agent

On Saturday, 17 October 2026, DeepSeek released DeepSeek-V4.1-Flash-1017, an update to the V4.1-Flash model it shipped in September. The weights went up on Hugging Face and ModelScope under the MIT license. The model is live in DeepSeek’s app, website and API. DeepSeek titled the announcement “Wish-Driven Programming Made Real” and summed the release up as “stronger benchmark performance, faster task completion, and shorter outputs than the previous version — a Pareto improvement in performance and cost.”

The same post opened two new products to every API user: Search and Fetch endpoints, backed by DeepSeek’s in-house search engine, at $0.75 per 1,000 successful requests each.

This is not a new frontier base model. The model card calls it “a minor update.” But the update’s practical effect on the cost of running a coding agent is large enough to deserve a close look.

What was released

From the model card and DeepSeek’s pricing page:

  • Weights: 763B parameters by Hugging Face’s count, shipped as FP8 safetensors (with FP4 experts) in 48 shards, about 510 GB in total.
  • Context: 1M tokens. The hosted API allows up to 384K output tokens, and the card recommends max_tokens of at least 256K for agentic work.
  • API name: deepseek-flash. The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are retired and now quietly served by 1017. If you call those names, the model behind your integration changed on Saturday.
  • Price, unchanged: $0.15 per million input tokens (cache miss), $0.003 (cache hit), $0.60 per million output tokens off-peak. Peak hours (weekday 01:00–04:00 and 06:00–10:00 UTC) cost double.

The parameter count needs care. The V4.1 technical report describes a 552B backbone plus a separate 196B “Engram” conditional-memory component. Hugging Face’s 763B label counts something else again, and the three numbers don’t add up neatly. Kingy AI’s analysis says the two figures use different accounting scopes and shouldn’t be treated as corrections of each other. Per that analysis, about 8B parameters are active during prefill and 16B during decode, which explains how DeepSeek can price the model this low.

DeepSeek doesn’t describe any architecture change. The only method it names is “reinforcement learning with self-feedback to improve end-to-end project development.” The demos are whole projects built from a single request: a 3D Forbidden City, a browser racing game, a Godot recreation of Portal 2. That is the “wish-driven” idea. DeepSeek never defines the phrase, but the examples suggest it means describing what you want and letting the agent build the whole thing.

The numbers

DeepSeek’s own table compares 1017 against the previous Flash, Kimi K3, GLM-5.3, GPT-5.6 Sol and Claude Opus 5:

Benchmark Flash-1017 Flash (Sept) GPT-5.6 Sol Opus 5
DeepSWE v1.1 77.0 74.2 73.0 74.0
Terminal-Bench 4.0 42.3 31.2 39.9 51.8
ProgramBench 29.1 20.3 23.0 37.0
NL2Repo-Bench 68.2 65.4 56.8 75.3
Automation-Bench 60.5 54.8 45.8 50.3
HLE (no tools) 42.6 36.8 44.5 56.3

The real jump is over its own predecessor: +11.1 points on Terminal-Bench and +8.8 on ProgramBench, which is roughly a 43% relative gain on ProgramBench. Over Opus 5 the result is mixed. Flash-1017 wins 4 of the 11 rows (DeepSWE, Automation-Bench, Agents’ Last Exam, Apex Bench) and loses the other 7, often by a wide margin. The Decoder’s headline says it “beats larger rivals,” but Kingy’s row-by-row reading says that blanket claim is hard to defend.

The more interesting figure is cost. DeepSeek charts DeepSWE accuracy against cost per task. As The Decoder reads that chart, 1017 reaches 77.0% at just under $0.20 per task, and a low-effort setting clears 72% for about $0.10. The September Flash scored 74.2% at around $0.50. Rival frontier models land somewhere between $2 and more than $10 per task. Since the per-token price didn’t change, the whole saving comes from the model using fewer tokens and finishing tasks faster.

Kingy ran its own small pilot: seven tasks, two repeats each, through Vercel’s AI Gateway. Flash-1017 passed 14 of 14 for $0.016 in total. GPT-6.1 Sol also passed 14 of 14, for $0.058, and Claude Sonnet 5.5 did the same for $0.065. That is too small to rank models, and three of them hit the ceiling. It does fit the cost story.

What to be skeptical about

  • The comparison set is a generation old. DeepSeek’s table compares against Opus 5 and GPT-5.6 Sol. It leaves out Claude Opus 5.5, GPT-6 Sol or Google’s Gemini 4 Argon, which Google reports at 77.9 on DeepSWE. “Leads on DeepSWE” holds only against the models DeepSeek chose to compare with.
  • Harness dependence. The footnotes say Terminal-Bench ran in the Claude Code harness and DeepSWE in Mini-SWE-V2. Agent scores change a lot with the scaffold, so results in your own tooling may differ.
  • No numbers for the efficiency claims. “Shorter outputs” and “faster completion” appear only as words in the text. The per-task cost data exists only as a chart image.
  • Nothing independent yet. The Decoder notes that third-party benchmarks haven’t appeared yet. Kingy says the same of its numbers.
  • Self-hosting is not trivial. Without the server and working memory, the download alone is 510 GB. Kingy rented two B300 GPUs and verified every shard, then hit a loader error (KeyError: 'F8_E8M0' in safetensors 0.8.0) before getting any inference out. The release also ships no Jinja chat template. Instead you get a Python reference encoder and the Rust-based deepseek-recipe library, so check vLLM and SGLang support before you plan a deployment.

Why it matters

Agentic coding is now priced per task, not per token. One repository-level fix costing $0.20 instead of $2–10 changes what a team can automate: run several agent attempts on every issue in parallel and keep the passing one, sweep the backlog every night, generate tests for a whole codebase. At these prices the limit is how many results you can review, not how much compute you can afford.

The Search/Fetch launch matters too. Grounding an agent in the web has meant paying a separate search vendor. DeepSeek now sells that capability alongside the model, and calls made from DeepSeek Harness, Claude Code or Codex through its server-side tools are free until 31 October 2026, 16:00 UTC. That is a clear bid to become the default backend inside agent harnesses that other companies built.

The open question is the one that follows every DeepSeek release: do the scores hold up outside DeepSeek’s own evaluation setup? For now, a model with MIT-licensed weights claims frontier-level results on one coding benchmark at one-tenth to one-fiftieth of the per-task cost. Independent evaluations over the coming weeks will show how much of that survives.

Sources