DeepSeek V4.1-Flash-1017: a $0.75 search API, a silent model swap, and a V4-Pro you shouldn't build on
On October 17, 2026, DeepSeek published DeepSeek-V4.1-Flash-1017. It's a "minor update" to the September V4.1-Flash, released under MIT on Hugging Face and now serving every deepseek-flash API call. The same release opened DeepSeek's in-house search engine as a first-party Search API and Fetch API at $0.75 per 1,000 successful calls each.
Two things matter if you run agents on DeepSeek. The model behind your model ID changed, with no pinning option. And web grounding, which most teams build or buy separately, now comes from the model vendor, with its own price, rate limits and a free window ending October 31.
What actually shipped
The model is V4.1-Flash-1017, the same 552B causal encoder-decoder MoE (8B active for input, 16B for output) launched September 10. DeepSeek's release note promises "stronger benchmark performance, faster task completion, and shorter outputs". API usage is unchanged: deepseek-flash now resolves to DeepSeek-V4.1-Flash-1017. Context is 1M and max output 384K. Vision is supported.
The benchmarks are self-reported (model card). DeepSWE v1.1 rises from 74.2 to 77.0, Terminal-Bench 4.0 from 31.2 to 42.3 (Claude Code harness), and ProgramBench from 20.3 to 29.1. It still trails Opus 5 on HLE (42.6 vs 56.3) and ProgramBench (29.1 vs 37.0).
Search and Fetch live at POST /v1/tools/search and POST /v1/tools/fetch on api.deepseek.com. Search returns 1 to 50 results, each with a ~600 to 800-character summary meant for models, and full content if need_content: true. Fetch returns a page as Markdown. Billing applies only to HTTP 200 responses with non-empty results. Errors, empty results and timeouts are free, and count doesn't change the price.
There is a free window: server-side Search and Fetch calls made through DeepSeek Harness, Claude Code or Codex are free until October 31, 2026, 16:00 UTC (00:00 November 1, Beijing time). As for legacy names, deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, but those models are retired. Requests are served by 1017 and billed at Flash rates.
Numbers that matter
Here is how Flash pricing has moved since the August hike (per 1M tokens, cache-miss input / output):
| Period | Off-peak | Peak |
|---|---|---|
| Before Aug 16 (flat) | $0.14 / $0.28 | same |
| Aug 16 hike | $0.22 / $0.66 | $0.44 / $1.32 |
| Since Sept 10 (unchanged by 1017) | $0.15 / $0.60 | $0.30 / $1.20 |
Cache hits cost $0.003 off-peak and $0.006 peak. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays. Weekends and Chinese public holidays are off-peak.
The 1017 release itself didn't change per-token prices. The "price cut" in some coverage refers to the September 10 change, plus DeepSeek's claim that shorter outputs lower the cost per task. DeepSeek's own DeepSWE chart, as reported by The Decoder, puts 1017 at just under $0.20 per task, against about $0.50 for the September checkpoint. The input price is still ~7% above where it was before August. Output is still more than double. For background, see the August price hike.
Illustrative cost of first-party web grounding per agent task
| Item | Calculation | Cost |
|---|---|---|
| Tool fees, research-style task (15 searches, 10 fetches) | 25 × $0.00075 | $0.019 |
| Reading results: 10 fetched pages at ~5K tokens (50K) plus 15 searches × 10 summaries × ~200 tokens (30K) | 80K cache-miss input tokens × $0.15/M | $0.012 off-peak, $0.024 at peak |
| At scale: 1,000 such tasks a day | 25,000 calls × $0.00075 | ≈ $19 a day, ~$560 a month in tool fees alone |
A useful rule of thumb: one Fetch call costs the same as reading 5,000 uncached tokens off-peak ($0.00075). Tool fees are on the same order as the tokens they produce. Next to a ~$0.20 coding task, web grounding adds roughly 10 to 15%.
The V4-Pro situation
The lead story going around is "V4-Pro retired". The record is messier than that. The September 10 launch post said all deepseek-v4-pro requests would route to V4.1-Flash, billed at Flash rates, from 04:00 UTC on September 14 "until V4.1-Pro launches". PricingSaaS then documented a notice on September 12 that kept V4-Pro billing unchanged, and that notice was quietly removed on September 20. As of today, DeepSeek's pricing page still lists deepseek-v4-pro as DeepSeek-V4-Pro-0813 at $0.66 / $1.98 off-peak (peak $1.32 / $3.96), with a concurrency limit of 500 against Flash's 2,500.
Whatever DeepSeek's routing is today, V4-Pro is an alias on borrowed time.
My read: whatever DeepSeek's routing is today, V4-Pro is an alias on borrowed time. It also costs 4.4× Flash's input price and 3.3× its output price, for a model that DeepSeek's own tables show losing to 1017 on agentic coding. I'd treat the September notice as the migration deadline that already passed.
Build or buy the search layer
The first-party tool wins on integration. It works server-side in the harnesses you probably already use, and it returns model-sized summaries. The trade-offs start with vendor coupling: your grounding quality now depends on DeepSeek's index, and you can't A/B the model and the search engine independently. Rate limits are unclear. The docs list Search at 100 QPS and Fetch at 25 QPS per account, with per-domain limits on Fetch, while some coverage reported 5 QPS; test this before you fan out parallel sub-agents. On the data path, your queries and fetched URLs go to DeepSeek, and retention isn't documented. Finally, Fetch is the easy part. A self-hosted fetcher costs almost nothing. The search index is what you're paying for.
What I'd do
Pin with weights, not IDs. deepseek-flash changed checkpoints without a version suffix, so if eval stability matters, re-run your regression suite now, and where you need reproducibility, self-host the MIT weights. Move off deepseek-v4-pro to deepseek-flash, and off the legacy deepseek-v4-flash* names, before DeepSeek moves you. Measure 1017's output length on your own tasks; the per-task savings depend on the "shorter outputs" claim holding for your workload.
Use Search and Fetch for free until October 31, log the call counts per task, then multiply by $0.00075 before deciding whether to keep them. And keep your own Fetch: pay DeepSeek for search only if its results beat your current provider on your queries.
What I couldn't verify
Benchmarks and per-task cost are DeepSeek-reported; The Decoder notes that no independent benchmarks were out at publication. I also couldn't pin down V4-Pro's actual routing today. The pricing page lists it with its own version and prices, while the September notice said it would route to Flash, and I couldn't resolve this from primary sources. Rate limits conflict between the official guide (100/25 QPS) and press reports (5 QPS); I've used the guide. On the "price cut" on October 17, no per-token change appears in DeepSeek's release note. The cut I can document is September 10's. I dropped claims of a "two-week billing grace period" and a market-moving "hit to chip stocks". Neither appears in DeepSeek's own materials.
I'll revisit when V4.1-Pro ships or DeepSeek publishes a dated deprecation schedule. See also the V4.1-Flash architecture and DeepSeek in 2026.
Sources
Primary: DeepSeek API News: DeepSeek-V4.1-Flash-1017 Release (2026/10/17); DeepSeek API: Models & Pricing; DeepSeek API: Search & Fetch guide; DeepSeek-V4.1-Flash-1017 model card (Hugging Face); DeepSeek: V4.1-Flash launch post (Sept 10, 2026). Coverage: InfoWorld: DeepSeek raises some V4 prices by more than 10x; PricingSaaS: DeepSeek quietly retires V4 Pro again; The Decoder: V4.1 Flash 1017 cuts per-task costs; MarkTechPost: DeepSeek releases V4.1-Flash-1017, opens Search API.