DeepSeek V4.1 Flash-1017 is on BazaarLink: beats Opus 5 on DeepSWE at ~US$0.2 per task
DeepSeek's October 17 snapshot, V4.1 Flash-1017, is now available through an OpenAI-compatible API. Benchmarks, pricing, reasoning modes and a quick example.
DeepSeek's new V4.1 Flash snapshot, V4.1 Flash-1017, released on October 17, is now available on BazaarLink. The model ID is deepseek/deepseek-v4.1-flash-1017, and it works with your existing BazaarLink API key — no separate sign-up needed.
What changed
1017 is an updated snapshot of V4.1 Flash focused on coding and agent work. According to DeepSeek's published results, it beats flagship models that cost dozens of times more on several agent and software-engineering benchmarks:
| Benchmark | V4.1 Flash-1017 | V4.1 Flash (previous) | Claude Opus 5 | GPT 5.6-Sol |
|---|---|---|---|---|
| DeepSWE v1.1 | 77.0 | 74.2 | 74.0 | 73.0 |
| Automation-Bench v1.0.6 | 60.5 | 54.8 | 50.3 | 45.8 |
| Apex Bench v1.1 | 72.1 | 64.6 | 65.8 | 54.3 |
| Agents' Last Exam | 35.4 | 31.8 | 28.6 | 26.7 |
| Terminal-Bench 4.0 | 42.3 | 31.2 | 51.8 | 39.9 |
| NL2Repo-Bench | 68.2 | 65.4 | 75.3 | 56.8 |
| HLE (no tools) | 42.6 | 36.8 | 56.3 | 44.5 |
(Source: DeepSeek. Terminal-Bench was run with the Claude Code harness; DeepSWE with Mini-SWE-V2.)
DeepSeek's cost chart is the more interesting part: on DeepSWE, 1017 at max reasoning reaches 77% for roughly US$0.2 per task on average, while Claude Opus 5 spends around US$10 or more per task to reach 73–74%.
It doesn't win everywhere. On Terminal-Bench, NL2Repo and HLE — tasks that lean on deep reasoning or long workflows — Opus 5 is still clearly ahead.
Specs
| Item | Details |
|---|---|
| Model ID | deepseek/deepseek-v4.1-flash-1017 |
| Context length | 1,048,576 tokens |
| Input | Text, images |
| Output | Text |
| Reasoning | On by default (high); set reasoning.effort to none / low / high / max |
| Supports | Tool calling (tools / tool_choice), structured output (response_format), streaming |
| Endpoints | /v1/chat/completions, /v1/responses, /v1/messages |
Pricing (USD per million tokens)
| V4.1 Flash-1017 | V4.1 Flash (previous) | |
|---|---|---|
| Input | $0.30 | $0.21 |
| Output (incl. reasoning tokens) | $1.20 | $0.84 |
| Cache read | $0.006 | $0.03 |
1017 costs a bit more per token than the previous version, but cache reads are much cheaper. If you resend the same long context — for example an agent repeatedly reading the same codebase — your actual bill can end up lower. The previous deepseek/deepseek-v4.1-flash remains available.
Quick start
curl https://api.bazaarlink.ai/v1/chat/completions \
-H "Authorization: Bearer $BAZAARLINK_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "deepseek/deepseek-v4.1-flash-1017",
"messages": [{"role": "user", "content": "Find the bug in this function and fix it"}],
"reasoning": {"effort": "high"}
}'
With the OpenAI SDK, set base_url to https://api.bazaarlink.ai/v1 and model to deepseek/deepseek-v4.1-flash-1017. Anthropic-format tools such as Claude Code can use /v1/messages.
Tips
- For coding and agents, start with
highand switch tomaxfor hard tasks. DeepSeek's top scores were measured inmaxmode. - Use
noneorlowfor simple Q&A. Reasoning tokens are billed at the output price, so turning reasoning off saves the most when you don't need it. - Don't set
max_tokenstoo low. Reasoning uses part of the output budget first; too small a limit can cut the response off before the answer. - Use caching for repeated long context. Cache reads are only $0.006 per million tokens.
For the previous version's cost math, see DeepSeek V4.1 Flash pricing. To compare other models, search deepseek-v4.1-flash-1017 on the model list.
TWD billing · Taiwan invoices · leading AI models · OpenAI-compatible API