DeepSeek released DeepSeek-V4.1-Flash-1017 on 17 October 2026, a minor update to Flash, on its website, app and API, with weights on Hugging Face under the MIT license. The vendor’s own benchmarks and one small independent pilot show a mixed picture: strong value, but not a clear win over the top models.
The short answer
- Vendor benchmarks: DeepSeek’s table shows it ahead of Opus 5 on DeepSWE v1.1 (77.0 versus 74.0). It is behind on Terminal-Bench 4.0 (42.3 versus 51.8) and ProgramBench.
- Independent pilot: Kingy AI had it pass 14 of 14 first attempts at a fraction of the cost of Sol and Sonnet 5.5. The sample cannot separate the top three models.
- API price: output costs $0.60 per million tokens off-peak and $1.20 at peak.
- Release: DeepSeek-V4.1-Flash-1017, 17 October 2026, “a Pareto improvement in performance and cost” in DeepSeek’s words.
- Recommended harness: DeepSeek Harness v0.2.1-rc.1 or later for the best experience.
What are the model basics?
- Weights: 763B parameters per the Hugging Face card (active parameters not stated), MIT license, 1M-token recommended context.
- API name: deepseek-flash, served by this version per the pricing page.
- Old names: deepseek-v4-flash and deepseek-v4-flash-vision-exp are retired but still accepted. Flash-1017 serves them at Flash prices.
What does DeepSeek’s own benchmark table show?
The announcement’s table is an image, so these figures come from the model card. The full table has columns for Flash-1017, the previous Flash, K3, GLM-5.3, GPT-5.6 Sol and Opus 5. Some are reproduced below. The comparison column is named “Opus 5,” not Opus 5.5.
| Task | Flash-1017 | Previous Flash | GPT-5.6 Sol | Opus 5 |
|---|---|---|---|---|
| DeepSWE v1.1 (Mini-SWE-V2 harness) | 77.0 | 74.2 | 73.0 | 74.0 |
| Terminal-Bench 4.0 (Claude Code harness) | 42.3 | 31.2 | 39.9 | 51.8 |
| ProgramBench | 29.1 | 20.3 | 23.0 | 37.0 |
| NL2Repo | 68.2 | 65.4 | 56.8 | 75.3 |
| Apex Bench v1.1 | 72.1 | 64.6 | 54.3 | 65.8 |
| Automation-Bench v1.0.6 | 60.5 | 54.8 | 45.8 | 50.3 |
| HLE without tools | 42.6 | 36.8 | 44.5 | 56.3 |
What it means: Flash-1017 improves on the previous Flash in every row shown. It beats Opus 5 on DeepSWE, Apex and Automation-Bench, and trails it on Terminal-Bench, ProgramBench, NL2Repo and HLE. By our arithmetic the DeepSWE lead is 3.0 points (77.0 minus 74.0), while the Terminal-Bench gap is 9.5 points (51.8 minus 42.3). It wins narrowly on repository bug fixing and loses more on terminal-style agent work.
These are DeepSeek’s figures and should not be merged with other benchmark versions. This table’s Sol is “GPT-5.6 Sol,” while the Kingy pilot below used “GPT-6.1 Sol,” so they are different rows.
What did the independent Kingy AI pilot find?
On 17 October Curtis Pyke ran a small test through Vercel AI Gateway. It had seven fixed tasks, each twice per model, plus one revision for two single-file repair attempts (56 first submissions and 8 revisions, 64 calls).
The tasks were interval subtraction (184 cases), exact cent allocation (184 cases), revenue SQL on in-memory SQLite, source-bound JSON extraction, the latest versions of three jobs among 164 records, chart reading, and a ledger.py repair against 40 scenarios.
| Model | First submissions passed | Median seconds, ordinary tasks | Gateway credit for first submissions (USD) |
|---|---|---|---|
| DeepSeek Flash 1017 (high) | 14/14 | 2.47 | 0.015731 |
| GPT-6.1 Sol (high) | 14/14 | 4.77 | 0.057710 |
| Claude Sonnet 5.5 (high) | 14/14 | 1.59 | 0.065318 |
| Qwen3.8-27B (medium) | 12/14 | 15.06 | 0.084562 |
Our arithmetic from the cost column: Sol cost about 3.7 times Flash-1017 (0.057710 divided by 0.015731), and Sonnet 5.5 about 4.2 times. All three passed everything, so on this sample Flash-1017 was the cheapest way to the same score.
Settings: 8,192 max output tokens and a 180-second timeout. Reasoning budgets were not equal. Qwen failed on a 180-second timeout (still charged $0.0263945) and one wrong SQL answer. DeepSeek’s interval-subtraction calls took 27.1 to 27.9 seconds against 2.8 to 3.4 for Sonnet. Both passed.
What are the author’s caveats?
- It is a small pilot, not a ranking.
- Three models hit the ceiling, so the test cannot separate them.
- Checks were deterministic, with no human or model judge.
- The grader initially omitted divmod and object. It was corrected and all responses were rescored.
- Timings depend on routing, network and provider.
- Costs are Gateway credits, not provider invoices.
- It does not measure long shell or browser agents.
What are the direct API prices?
| Per million tokens | Off-peak | Peak |
|---|---|---|
| Cache-hit input | $0.003 | $0.006 |
| Cache-miss input | $0.15 | $0.30 |
| Output | $0.60 | $1.20 |
Peak is Monday to Friday, 01:00-04:00 and 06:00-10:00 UTC. Chinese public holidays are off-peak all day, and all other hours including weekends are off-peak.
The page lists a 1M context, up to 384K output, concurrency of 2,500, vision support, and Responses and Anthropic API formats. These are API prices, not a chat subscription. For why token price is not task cost, see this article.
How should you decide?
- Test Flash-1017 on 10 to 20 of your own tasks with a written acceptance rule, rather than relying on a vendor table. This is our suggestion.
- Price peak and off-peak separately. The price doubles at peak.
- Do not generalize the Kingy pilot to long-running agents. Terminal-style work is where DeepSeek’s own table shows the biggest gap.
Questions and answers
Is Flash-1017 better than Opus 5?
Partly. In DeepSeek’s own table it leads on DeepSWE, Apex and Automation-Bench and trails on Terminal-Bench, ProgramBench, NL2Repo and HLE.
What license does it use?
MIT, per the Hugging Face card.
How much does output cost?
$0.60 per million tokens off-peak and $1.20 at peak.
Did it pass the independent coding test?
Yes. In the Kingy AI pilot it passed 14 of 14 first submissions, tying GPT-6.1 Sol and Claude Sonnet 5.5, at the lowest gateway cost of the four models.
Which API model name should I use?
Use deepseek-flash. The retired names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work but are served by Flash-1017.
Terms in plain language
- Benchmark: a fixed set of tasks used to score models.
- Weights: the trained numbers that make up a model.
- Harness: the tool that runs a model on coding tasks.
- Token: a small piece of text that a model reads or writes.
- Cache hit: input the service has already processed, which costs less.
Sources
- DeepSeek: Flash-1017 announcement, 17 October 2026
- Model card on Hugging Face
- DeepSeek: models and pricing
- Kingy AI: Flash-1017 pilot
Facts last checked against the linked pages on 18 October 2026. Benchmark figures are the developer’s; prices can change.
Part of the guide: AI Model Pricing and API Cost Guides
