DeepSeek V4.1 Flash 1017: The $0.15 Coding Model and Its New Search API

DeepSeek did something rare in the AI industry on October 17, 2026: it shipped a better model, lowered no prices, raised no hype, and still changed the economics of coding agents. The new DeepSeek-V4.1-Flash-1017 is a minor version bump on paper — DeepSeek's own words call it "a minor update to DeepSeek-V4.1-Flash." But under that modest framing, three things moved: benchmark scores went up, output lengths went down, and the model weights are MIT-licensed on Hugging Face. The same release also opened DeepSeek's in-house search engine to outside developers for the first time, with Search and Fetch APIs priced at $0.75 per 1,000 successful requests.

Here is what actually matters: for coding-agent workloads, DeepSeek's own charts show a finished task now costing just under $0.20, down from about $0.50 on the previous snapshot, and third-party scorer Artificial Analysis has the model at an Intelligence Index of 45, up from 39. Nothing here requires a migration — DeepSeek says every legacy Flash name now resolves to the new snapshot automatically.

The quick answer

• What it is: DeepSeek-V4.1-Flash-1017 is the October 17, 2026 snapshot of DeepSeek's V4.1 Flash model — a mixture-of-experts coding model with a 1-million-token context window and up to 384K output tokens.
• What it costs: unchanged from the September snapshot — $0.15 per million input tokens and $0.60 per million output tokens off-peak; peak-hour rates (01:00–04:00 and 06:00–10:00 UTC, weekdays) double both figures.
• The open part: the weights are MIT-licensed on Hugging Face, so you can self-host or fine-tune with no per-token bill at all.
• The new part: DeepSeek opened its Search and Fetch APIs to outside developers at $0.75 per 1,000 successful requests — free until 00:00 Beijing time on November 1, 2026.

What actually changed on October 17

The model card's introduction is admirably plain: "a minor update" delivering "substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks," with reduced task execution time and shorter outputs. Those two efficiency claims are doing more work than they appear to. For an agentic coding model, a shorter output on the same task is a double win — the job finishes faster and you pay for fewer output tokens.

The release continues a busy open month for DeepSeek — earlier in October the company open-sourced DeepInfer, its production inference engine. And it is not the only fully open coding model this month: Recursive's Flint 40B runs on a laptop, while DeepSeek's Flash plays for cloud-scale agent throughput.

The practical consequence is immediate: if you were calling deepseek-v4-flash or deepseek-v4-flash-vision-exp, you don't have to do anything. DeepSeek's API docs state those legacy names are still accepted, but the underlying models have been retired — requests are now served by DeepSeek-V4.1-Flash-1017 and billed at the Flash price. That silent upgrade is worth noting in reverse: your benchmark baselines from September may quietly have shifted underneath you, so re-run before drawing conclusions.

Two operational details from the card matter for builders. First, this release does not include a Jinja-format chat template — DeepSeek ships a self-contained Python reference implementation (encoding.py) with test cases for multi-turn conversations, tool calling, thinking mode, numeric reasoning effort, mid-conversation system messages, and interleaved image content, plus a Rust-based deepseek-recipe toolkit for production use. Second, the recommended sampling parameters are temperature 1.0 and top_p 1.0, with a 1M-token context window and at least 256K max output tokens.

The pricing, with fine print

Here is the full official rate card, from DeepSeek's own Models & Pricing page:

Read the fine print carefully, because this is where DeepSeek's pricing diverges from the flat rate cards most coverage still quotes:

• Peak hours are real and specific: 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays. Every other hour — including full weekends and Chinese holidays — bills at the off-peak half rate.
• Cache hits are nearly free: $0.003 per million tokens off-peak is one-fiftieth of the input price, which rewards agent harnesses that reuse context across steps.
• Concurrency favors Flash heavily: the 2,500 concurrency limit is 5x the 500 offered to V4 Pro users, which matters for high-throughput pipelines.
• Fees deduct from topped-up or granted balance, with granted balance used first when both exist. And the page's own caveat stands: "Product prices may vary" — recheck the docs before budgeting.

How does that compare to the flagship tier? Against DeepSeek's own V4 Pro ($0.66 input / $1.98 output off-peak), Flash is roughly 3–4x cheaper per token. Against Western flagships, independent pricing trackers put the gap at roughly 3x–25x depending on the model and the hour. Just don't read "cheapest" into that: smaller Western models have started undercutting Flash's line on pure token price, and the per-task cost — tokens times how many tokens the model actually burns — is what decides your bill.

The new Search and Fetch APIs

The bigger strategic move is hiding in the pricing page's second section. DeepSeek is opening its in-house search engine — which it says indexes hundreds of billions of web pages — to outside developers through two standalone APIs:

• Search API: $0.75 per 1,000 successful requests
• Fetch API: $0.75 per 1,000 successful requests

The price signal is deliberate. Brave's search API lists at $5 per 1,000 requests and Exa's search starts at $4 per 1,000, per their pricing pages as of October 17. DeepSeek is pricing its engine at roughly a sixth to a fifth of that — though note these are not like-for-like products, with different index sizes, latency guarantees, and retrieval features, so treat the comparison as directional.

The deadline to know: server-side tool calls from Claude Code and Codex already hit these APIs behind the scenes, and they are currently free. Per-request billing starts at 00:00 Beijing time on November 1, 2026. DeepSeek's own guidance is practical rather than promotional: rename to deepseek-flash, rerun output tests, and count your web tool calls before October 31, so the search line becomes a planned cost rather than a surprise invoice.

The benchmarks: an honest read

Every number below comes from DeepSeek's own model card — vendor-run, not independently verified, and benchmarked with the Claude Code harness for Terminal-Bench 4.0 and the Mini-SWE-V2 harness for DeepSWE v1.1. DeepSeek's chosen comparison set is telling: it measures against Moonshot's K3, Z.ai's GLM-5.3, OpenAI's GPT-5.6 Sol, and Anthropic's Opus 5, so the table is the best defense of the "roughly on par with Opus 5" chatter — and it doesn't fully back it.

The honest read: on the coding-and-agent tests, Flash 1017 beats Opus 5 on DeepSWE, Automation-Bench, Agents' Last Exam, and Apex Bench — genuine wins on the workloads this model is built for. On general reasoning (HLE) and terminal-work (Terminal-Bench), Opus 5 is clearly ahead, 56.3 to 42.6 and 51.8 to 42.3. A cheap open model scoring 72.1 on Apex Bench against Opus 5's 65.8 is a real result. "On par with Opus 5" overall is a mood — the table says Flash wins some, loses others, at roughly one-twentieth the token price. One third-party datapoint: Artificial Analysis scores it 45 on its Intelligence Index, up from 39 for the September snapshot, at $0.26 per index task against $0.49 before.

Pros and cons

Pros

• Genuine price-to-performance: coding-agent benchmark wins over Opus 5 on several tests at a fraction of the token cost, with a 1M-token context window.
• MIT-licensed weights: self-host, fine-tune, or commercialize with no API bill — a real open-weight release, not an "open" API tier.
• Zero-migration upgrade: legacy Flash model names resolve to the new snapshot automatically.
• Vision is supported on the Flash line, unlike V4 Pro — interleaved image content is part of the supported prompt format.

Cons

• Peak pricing is a trap for scheduled jobs: weekday 01:00–04:00 and 06:00–10:00 UTC doubles your token bill; batch jobs should be scheduled outside those windows.
• No published parameter count in the release material — the mixture-of-experts sizing is undisclosed, so hardware planning for self-hosting relies on community estimates, not official figures.
• No Jinja chat template shipped — production prompt formatting goes through DeepSeek's reference implementation or Rust toolkit, adding integration work.
• All headline benchmarks are DeepSeek's own runs — independent verification of the gains is still pending.

Flash vs V4 Pro: which tier?

FAQ

Is DeepSeek V4.1 Flash 1017 free to use?

The API is not free — it bills per token from a topped-up or granted balance. But the model weights are MIT-licensed on Hugging Face, so self-hosting costs only your hardware. That makes the "free" question a deployment question, not a pricing one.

When does DeepSeek start charging for Search API calls made through Claude Code and Codex?

At 00:00 Beijing time on November 1, 2026, per DeepSeek's official pricing page. Until then, those server-side tool calls are free of charge.

Do I need to migrate from deepseek-v4-flash to the new snapshot?

No. DeepSeek's docs say the legacy names are still accepted and requests are served by DeepSeek-V4.1-Flash-1017 at Flash pricing. The one action item: re-run your output tests, because shorter outputs may break parsers tuned to the old behavior.

How does V4.1 Flash 1017 compare to Claude Opus 5?

On DeepSeek's own benchmark table, Flash 1017 beats Opus 5 on DeepSWE v1.1 (77.0 vs 74.0), Automation-Bench (60.5 vs 50.3), Agents' Last Exam (35.4 vs 28.6), and Apex Bench (72.1 vs 65.8), but trails on HLE (42.6 vs 56.3) and Terminal-Bench 4.0 (42.3 vs 51.8). All are vendor-run numbers.

Can I run DeepSeek V4.1 Flash 1017 locally?

Yes — the MIT license permits it, and the model card includes an inference folder with weight-conversion and local-run instructions. Note that DeepSeek publishes no total parameter figure in the release material, so sizing hardware relies on community reporting rather than official specs.

Key takeaways

• DeepSeek-V4.1-Flash-1017 (October 17, 2026) is a genuine efficiency upgrade: better coding/agent benchmarks, shorter outputs, and a per-task cost near $0.20 per DeepSeek's charts.
• API pricing is unchanged — $0.15/$0.60 per million tokens off-peak, doubling in weekday peak hours — but watch the peak windows and recheck the docs before budgeting.
• The MIT-licensed weights on Hugging Face make this one of the strongest open-weight coding releases of the quarter, with self-hosting as a real alternative to the API.
• The new Search and Fetch APIs ($0.75 per 1,000 successful requests, free until November 1) price DeepSeek's index at roughly a fifth of Brave's and Exa's listed rates — with different feature sets, so compare before committing.
• Legacy Flash names resolve to the new snapshot automatically, but shorter outputs can break parsers: rerun your output tests and count your search calls before October 31.

The bigger story of October 17 isn't a flagship launch — it's a pricing-and-plumbing update that makes agent infrastructure meaningfully cheaper to run. DeepSeek keeps winning by making the cheap tier better instead of making a more expensive tier. As of today, October 23, 2026, the official docs still list V4.1-Flash-1017 as the current Flash snapshot — the one to pin in your agent harness.