DeepSeek V4.1 Flash: Faster Open-Source Model, Cheaper API

DeepSeek V4.1-Flash-1017 open-source AI model — benchmarks, pricing and how to try it
DeepSeek V4.1-Flash-1017 open-source AI model — benchmarks, pricing and how to try it


DeepSeek just dropped another model that has the AI world doing a double-take. On October 17, 2026, the Chinese AI company released DeepSeek V4.1-Flash-1017 — a faster, cheaper, and fully open-weight update to its Flash model (PYMNTS). The headline numbers sound almost absurd: coding benchmark scores that rival models costing over 40 times more per token, and API pricing so low it costs about twenty cents to complete a full coding task. Here's what's real, what's marketing, and what it means for you.

⚡ Quick answer: DeepSeek V4.1-Flash-1017 is a free-to-download, MIT-licensed AI model released October 17, 2026. It scores 77.0 on the DeepSWE coding benchmark (ahead of GPT-5.6 Sol and Claude Opus 5 on that specific test, per DeepSeek's own numbers), completes coding tasks for about $0.20 each, and costs $0.15/$0.60 per million input/output tokens — roughly one-fortieth of Claude's price. The catch: all benchmark scores come from DeepSeek itself and haven't been independently verified yet.

In this article:

What DeepSeek V4.1 Flash Actually Is

DeepSeek V4.1 Flash is the company's fast, affordable model — the workhorse tier, roughly analogous to what GPT-5.6 Sol or Claude Haiku are in rival lineups. The 1017 build is a checkpoint update to the Flash model that first launched on September 10, 2026. Same model name, same API identifier, new weights.

That's actually one of the smartest parts of the release: if you were already calling deepseek-flash through the API, you woke up on October 17 with a better model and no code changes required. DeepSeek also quietly routed its web and mobile apps to the new checkpoint. There was even a period where the upgrade shipped under the existing API identifier before the changelog documentation appeared — a "it just got better" rollout that spared every developer a migration headache.

The company describes the update as delivering "stronger benchmark performance, faster task completion, and shorter outputs than the previous version." Behind that is a training technique DeepSeek calls "reinforcement learning with self-feedback," aimed specifically at end-to-end project development — the kind of long, messy, multi-step work involved in things like game creation or 3D modeling (TECHi's technical breakdown). Shorter outputs matter here more than you'd think: fewer wasted tokens per task means every job literally costs less, even before you look at the price list.

Benchmarks and Performance: The Numbers

DeepSeek published scores across 11 benchmarks. The headline: 77.0 on DeepSWE v1.1, a coding benchmark, up from 74.2 on the September checkpoint. For context, DeepSeek's own comparison table puts Claude Opus 5 at 74.0 and GPT-5.6 Sol at 73.0 on that same test. The new model reportedly beats GPT-5.6 Sol on 9 of 11 tests and Kimi K3 on 10 of 11 — while trailing Claude Opus 5 on 7 of 11, including raw reasoning (Humanity's Last Exam: 42.6 vs. Opus 5's 56.3).

The biggest jumps came on Terminal-Bench 4.0, where the score leapt from 31.2 to 42.3 — an 11.1-point gain that made even Thomas Wolf, Hugging Face's chief science officer, quip on X that it's "truly funny that DeepSeek calls it a 'small update'." AI researcher Nathan Lambert asked the question on everyone's mind: "Has the king of open weights returned?"

Benchmark Flash-1017 (new) Flash (Sept 10) Rival comparison
DeepSWE v1.1 (coding) 77.0 74.2 Claude Opus 5: 74.0, GPT-5.6 Sol: 73.0
Terminal-Bench 4.0 (agent coding) 42.3 31.2 Opus 5 still leads on terminal execution
HLE without tools (reasoning) 42.6 36.8 Opus 5: 56.3 (still ahead)
ProgramBench 29.1 20.3 —

But here's the honesty check, and it's a big one: every one of these scores is DeepSeek's own, produced on test harnesses DeepSeek chose, and none has been independently reproduced as of publication. That's the standard vendor-scorecard problem — the numbers most favorable to the story come from the party telling the story. As one developer put it: the price part checks out; the performance part is their word for now. The good news is that open weights mean independent verification is possible, and we expect it soon. We'll update this article when third-party boards score the 1017 build.

The Pricing: Why It's a Bigger Deal Than the Scores

Even if every benchmark turns out 10–20% inflated, the price-performance trajectory is the real signal. Look at the numbers side by side:

Model Input / 1M tokens Output / 1M tokens Cost per coding task
DeepSeek V4.1-Flash-1017 $0.15 (off-peak) $0.60 (off-peak) ~$0.20
GPT-5.6 Sol (OpenAI) — $20.00 (off-peak) $2–$10+
Claude Opus 5 (Anthropic) $25.00 (off-peak) $2–$10+

That's roughly one-fortieth the output price of Claude Opus 5. On DeepSeek's internal evaluation, the cost per coding task dropped 60% versus the September build — from about $0.50 to about $0.20. These rates are public on DeepSeek's pricing page and checkable, so unlike the benchmarks, this part needs no trust (MoneyToday's price breakdown).

Note the peak/off-peak detail: peak hours cost double ($0.30 input, $1.20 output) on weekday windows. Cached input is nearly free at $0.003 per million tokens — a meaningful discount if your app repeats prompts. For comparison, Google's Gemini 4 Argon launched at $2/$10 per million tokens, and Meta's Muse Spark Open targets the opposite end of the open-weight spectrum. Flash-1017 isn't trying to be the smartest model on earth — it's trying to be the cheapest good-enough one, and that might be the more disruptive move.

Open Weights (MIT) and What They Mean for You

Alongside the API launch, DeepSeek published the model weights on Hugging Face as deepseek-ai/DeepSeek-V4.1-Flash-1017 under the MIT license — one of the most permissive licenses in software. You can download, modify, fine-tune, and commercialize it with essentially no strings attached.

What does this mean in practice?

  • Run it yourself: host the model on your own GPUs, on-premise or in your cloud, with no per-token bill at all. This is huge for companies with data-residency requirements or predictable heavy workloads.
  • Fine-tune it: the September Flash weights were already fine-tunable; the 1017 checkpoint inherits that. Train it on your domain's data for specialized coding assistants.
  • Verify the claims: anyone with enough compute can reproduce the benchmarks — which is exactly why independent verification should arrive quickly.
  • Distill it: smaller, cheaper derivatives can be trained from it, legally, thanks to the MIT license.

Architecturally, the repo's model config is identical to the September release — the 1017 checkpoint changes the weights, not the shape of the model (ArtificialWatch's changelog analysis). Think of it as the same engine with a much better tune. It keeps the 1-million-token context window and native image input, so you can feed it whole codebases plus screenshots or diagrams in one prompt.

New Search and Fetch APIs Explained

Buried in the same announcement was a second product: DeepSeek opened its in-house web search engine — the same one powering search inside its own app — to all developers as two new APIs: Search and Fetch.

  • Price: $0.75 per 1,000 successful requests, per endpoint. Billing only triggers on a request that returns an HTTP 200 with an actual result — failed calls, empty results, and timeouts are free.
  • Rate limit: capped at 5 queries per second per account.
  • Free window: server-side tool calls routed through DeepSeek Harness, Claude Code, or Codex stay free until 00:00 Beijing time on November 1, 2026.

This is a pricing model worth studying: DeepSeek is charging for the tool an agent calls, not just the tokens it burns. If you're building agents that need live web data, a successful search at $0.00075 is essentially free infrastructure. For context on how fast this market is moving, check our AI statistics roundup — adoption numbers keep climbing even as prices collapse, a classic case of Jevons' Paradox: cheaper AI doesn't shrink the market, it explodes usage.

How to Try It in 3 Steps

Option A — use it in the chat app (easiest):

  1. Open DeepSeek's web app or mobile app — both were switched to the 1017 checkpoint automatically.
  2. Ask it to code, debug, or plan a multi-step project. Watch how concise the outputs are; shorter answers are one of the advertised improvements.
  3. Try an end-to-end task (e.g., "plan and scaffold a simple 2D game") to test the self-feedback agentic behavior.

Option B — use the API:

  1. Get an API key from DeepSeek's developer platform.
  2. Call the model with the existing deepseek-flash identifier — it now serves the 1017 build. OpenAI-compatible endpoints work.
  3. Monitor your first few tasks: check output length and cost per task against the ~$0.20 claim.

Option C — self-host the weights:

  1. Download deepseek-ai/DeepSeek-V4.1-Flash-1017 from Hugging Face (MIT license).
  2. Deploy with your preferred inference engine. Note the 1M-token context needs serious GPU memory — DeepSeek quotes about 0.9 GB of KV cache per million tokens, with the weights themselves the bigger bill.
  3. Run your own benchmark suite. Seriously — don't take vendor scores at face value, ours included.

Students learning to code on a budget should find this especially interesting — at these prices, you can run hundreds of practice tasks for the cost of a coffee. Our guide to the best free AI tools for students covers more budget-friendly options in the same spirit.

The Honest Caveats

No balanced coverage without the downsides:

  • Unverified benchmarks. Worth repeating: every performance claim is DeepSeek's own. Until independent labs score the 1017 build, treat the scores as a claim, not a fact.
  • Reliability questions. Bloomberg reported a five-hour DeepSeek outage shortly before this launch, with some experts attributing it to insufficient computing resources. Slowdowns and service disruptions have reportedly risen as the low-cost models strain DeepSeek's servers. Cheap is great until it's down.
  • It's a speed-tier model. On raw reasoning, Claude Opus 5 still leads by a wide margin (56.3 vs. 42.6 on HLE without tools). If you need the deepest reasoning, this isn't the pick — see our GPT-6 coverage and Gemini 4 Argon breakdown for the frontier tier.
  • Silent rollout. Shipping new weights under the same API name before documentation appeared is convenient — but it also means behavior changed under your app without a changelog. Pin checkpoints when stability matters.

Should You Switch? Our Verdict

✅ Switch or test if: you run high-volume coding agents, batch tasks, or student/hobby projects where cost dominates; you want open weights for self-hosting or fine-tuning; you build agents that need cheap web search.

❌ Don't switch if: you need maximum reasoning depth (frontier models still win there); your workload demands five-nines reliability with an SLA; you're mid-project on another provider and migration risk outweighs savings.

Pros: MIT-licensed open weights; ~1/40th of Claude's output price; strong coding/agentic scores (claimed); 1M-token context with vision; zero-migration upgrade for existing users; new search/fetch APIs at near-zero cost.

Cons: benchmarks unverified; recent outages; trails frontier models on deep reasoning; off-peak pricing has peak-hour doubles; search API capped at 5 QPS.

The bigger picture: whether or not Flash-1017 dethrones anyone, it keeps pushing the industry's cost floor down. Last year DeepSeek's R1 wiped roughly $589 billion off Nvidia's market cap in a single day by questioning Big Tech's AI spending. The "second DeepSeek shock" narrative is already circulating — and with NASDAQ down 2.1% on October 16 amid AI-profitability fears, markets will be watching Monday's open for the reaction to this weekend's announcement. One way or another, the trajectory is clear: capable AI agents are getting dramatically cheaper every few months.

Frequently Asked Questions

Is DeepSeek V4.1 Flash free?

The model weights are free to download under the MIT license — you can run, modify, and commercialize them. The hosted API isn't free, but it's extremely cheap: $0.15/$0.60 per million input/output tokens off-peak, with cached input at $0.003.

Is DeepSeek V4.1-Flash-1017 better than GPT-5.6 or Claude Opus 5?

On DeepSeek's own DeepSWE v1.1 coding benchmark, yes — 77.0 vs. 74.0 (Opus 5) and 73.0 (GPT-5.6 Sol). But these are vendor-reported scores, not independently verified, and Opus 5 still leads on raw reasoning benchmarks. Treat it as "remarkably strong for the price" rather than "best overall."

How do I access the new model?

Three ways: use DeepSeek's web/mobile app (auto-upgraded), call the deepseek-flash API identifier (now serving the 1017 build), or download the MIT-licensed weights from Hugging Face as deepseek-ai/DeepSeek-V4.1-Flash-1017 and self-host.

What are the Search and Fetch APIs?

DeepSeek's in-house web search engine, now available to developers. Search finds web results; Fetch retrieves page content. Each costs $0.75 per 1,000 successful requests, and calls through DeepSeek Harness, Claude Code, or Codex are free until November 1, 2026.

Why is it so much cheaper than competitors?

A combination of aggressive pricing strategy, efficient architecture (KV-cache compression is the subject of its companion paper), shorter outputs that waste fewer tokens, and the open-weights model that shifts compute costs to self-hosters. Whether the unit economics are sustainable long-term is an open question.

Should developers switch from OpenAI or Anthropic?

For high-volume, cost-sensitive coding and agent workloads, it's worth testing — many teams route simpler tasks to cheap models and reserve frontier models for hard reasoning. The seamless API-id upgrade means existing DeepSeek users get the improvement automatically.

Last updated: October 19, 2026.

Methodology: This article was researched from DeepSeek's official release documentation and API docs, plus reporting from PYMNTS, MoneyToday, ArtificialWatch, and TECHi. All benchmark figures are DeepSeek's self-reported numbers; no third-party verification of the 1017 build existed at publication time, and we state that explicitly rather than presenting vendor scores as established fact.

Disclosure: The AI Tools is an independent publication. This article contains no sponsored content, and we received no compensation from DeepSeek or any company mentioned. Learn more about us on our About page or reach out via Contact. More AI coverage: theaitools.tech.




 

Comments

Post a Comment