On October 17, 2026, DeepSeek posted a new model to Hugging Face called DeepSeek-V4.1-Flash-1017. The model card calls it "a minor update" to the V4.1-Flash model released on September 10. The benchmark table, though, does not look minor: DeepSeek reports gains on all 11 benchmarks it published, ranging from 2.0 to 11.1 points, with the terminal-coding score jumping from 31.2 to 42.3. (Source: AI Weekly; Ecosistema Startup)
The release also bundled something easy to miss: new Search and Fetch APIs priced at $0.75 per 1,000 successful requests each. For anyone building an AI agent that browses the web, that second number may matter as much as the model.
Quick summary: DeepSeek-V4.1-Flash-1017 was released October 17, 2026 under an MIT license, with a 1M-token context window and up to 384K output tokens. API pricing is $0.30 per million input tokens and $1.20 per million output tokens at peak, and half that off-peak. DeepSeek reports gains on all 11 published benchmarks, but most figures are vendor-reported and Claude Opus 5 still leads on 7 of the 11. The new Search and Fetch APIs cost $0.75 per 1,000 successful requests each, with a free promotion for certain tools that was set to end November 1.
What shipped
The model accepts text and image input, produces text, and offers a 1M-token context window with up to 384K tokens of output. The weights are open under the MIT license, which permits commercial use, and are hosted on Hugging Face and ModelScope. (Source: SmartScope; Product Hunt)
It is served under the existing API identifier, deepseek-flash. The Product Hunt listing says existing apps need no code changes and that the older deepseek-v4-flash names now route to this model. That is convenient, but it also means your production traffic may already be running on a different model than last week. (Source: Product Hunt)
A note on size: The Hugging Face page for the 1017 checkpoint lists 763 billion total parameters and does not state an active-parameter count. Some coverage cites 552 billion, which matches reporting on the earlier September V4.1-Flash release. (Source: Ecosistema Startup; SmartScope)
The benchmark jump, in numbers
DeepSeek's own before-and-after numbers, as listed on the Product Hunt page, show gains across coding, agent and reasoning tests. These are vendor-reported results, not independent evaluations. (Source: Product Hunt)
| Benchmark | Previous Flash | 1017 | Change |
|---|---|---|---|
| DeepSWE v1.1 | 74.2 | 77.0 | +2.8 |
| Terminal-Bench 4.0 | 31.2 | 42.3 | +11.1 |
| Automation-Bench | 54.8 | 60.5 | +5.7 |
| Apex Bench | 64.6 | 72.1 | +7.5 |
| ProgramBench | 20.3 | 29.1 | +8.8 |
| HLE (no tools) | 36.8 | 42.6 | +5.8 |
Context matters. The DeepSWE score of 77.0 sits above the 74.0 reported for Claude Opus 5 and 73.0 for GPT-5.6 Sol in the AI Weekly summary. But the same coverage says Flash-1017 trails Opus 5 on 7 of the 11 published benchmarks, and on Humanity's Last Exam it scores 42.6 against 56.3 for Opus 5. (Source: AI Weekly)
There is one independent data point. Ecosistema Startup reports an Artificial Analysis Intelligence Index score of 45 for the model, versus a median of 19 for comparable open-weight models above 150 billion parameters, with output speed of 230.4 tokens per second versus a 77.9 median. It also notes the model is more verbose than average, generating 210 million tokens while running the index versus a 140 million median. (Source: Ecosistema Startup)
What it costs
At peak, the API charges $0.30 per million input tokens (cache miss) and $1.20 per million output tokens, with cache hits at $0.006. Off-peak, the prices are half: $0.15 input, $0.60 output and $0.003 for cache hits. Peak hours are weekdays 01:00 to 04:00 and 06:00 to 10:00 UTC. (Source: SmartScope)
SmartScope suggests budgeting at the peak output rate of $1.20 to avoid surprises. Ecosistema Startup estimates a blended rate of roughly $0.18 per million tokens under a 7:2:1 mix of cache hits, input and output, and an average cost of $0.26 per task on the index. Those are third-party estimates, so check your own traffic mix. (Source: SmartScope; Ecosistema Startup)
The part most people will miss: the $0.75 search API
DeepSeek launched Search and Fetch APIs alongside the model. Each costs $0.75 per 1,000 successful requests, failed, empty or timed-out requests are not billed, and the Search API is capped at 5 queries per second per account. Search returns up to 50 ranked web results per query, and Fetch converts a web page to Markdown. (Source: Ecosistema Startup; AI Weekly)
Calls made as server-side tools through DeepSeek Harness, Claude Code or Codex were free until October 31, 2026, 16:00 UTC (November 1 in Japan), after which standard rates apply. (Source: SmartScope)
| Search API (per 1,000 requests) | Price | Note |
|---|---|---|
| DeepSeek Search / Fetch | $0.75 | Each endpoint; failed calls not billed |
| Brave | $5 | As listed by Ecosistema Startup |
| Exa | $7 standard / $12 deep | As listed by Ecosistema Startup |
| Tavily | $5 to $8 | As listed by Ecosistema Startup |
| Perplexity Sonar | From $5 plus token costs | As listed by Ecosistema Startup |
On those competitor prices, which come from one secondary source and should be rechecked on each vendor's page, DeepSeek would be about 6.7 times cheaper than the next-cheapest option. (Source: Ecosistema Startup)
Caveats worth knowing
- Most benchmark figures, including DeepSWE and Terminal-Bench, are DeepSeek's own measurements, and independent evaluations were thin when the model launched.
- Ecosistema Startup reports an API outage of nearly five hours on October 15, after which DeepSeek apologized and issued credits. Single-vendor dependence is a real risk.
- Claude Opus 5 still leads on most hard-reasoning benchmarks, including a gap of almost 14 points on Humanity's Last Exam.
- The model card lists 763B total parameters but no active count, so check hardware needs before planning to self-host.
Before switching a production workload, run 20 to 30 of your own real prompts through the new model and your current one, then compare quality and total cost per finished task, not just price per million tokens. A verbose model can erase part of a cheap per-token price.
01When was DeepSeek-V4.1-Flash-1017 released?
October 17, 2026, on Hugging Face under the MIT license. It updates the V4.1-Flash model released September 10.
02How much does the API cost?
At peak, $0.30 per million input tokens and $1.20 per million output tokens; off-peak is half that. Cache hits are $0.006 at peak and $0.003 off-peak.
03What does the search API cost?
$0.75 per 1,000 successful requests for each of the Search and Fetch endpoints. A free period for calls through DeepSeek Harness, Claude Code and Codex ended around November 1.
04Does it beat Claude Opus 5?
On DeepSWE it scores 77.0 versus 74.0 for Opus 5, per vendor-reported coverage. Opus 5 still leads on 7 of 11 published benchmarks, including Humanity's Last Exam (56.3 versus 42.6).
05Is it safe to self-host?
The MIT license permits commercial use, but the model card lists 763B total parameters with no active count, so check hardware needs before committing.
Cheap open models are most useful when you can test them against the premium ones on your own tasks. LumiChats offers 40+ premium AI models side by side in one subscription, so you can see whether a budget model like this one is good enough for your work before you rebuild anything around it.
