DeepSeek V4.1 Flash 1017 lands on Novita AI with higher coding scores, same price
Novita AI has added DeepSeek V4.1 Flash 1017, the October 17, 2026 update to DeepSeek V4.1 Flash, as a serverless API. Pricing and context limits are unchanged, while DeepSeek reports higher agentic coding scores.
- DeepSeek V4.1 Flash 1017 is live on Novita AI as deepseek/deepseek-v4.1-flash-1017, serverless and OpenAI-compatible.
- Pricing is unchanged: $0.3 per 1M input, $0.006 per 1M cache-read and $1.2 per 1M output tokens, with a 1,048,576-token context.
- DeepSeek reports DeepSWE v1.1 at 77.0, up from 74.2 for the September release on the same benchmark version.
- The earlier deepseek/deepseek-v4.1-flash ID stays available at the same price and limits.
- DeepSeek's Search and Fetch API is not hosted on Novita AI.
What changed in the 1017 update
DeepSeek V4.1 Flash 1017 is a minor update to DeepSeek V4.1 Flash, the multimodal Mixture-of-Experts model DeepSeek released on September 10, 2026. According to the Novita AI post, the DeepSeek web app and API moved to the new build, the API calling method stayed the same, and DeepSeek's official pricing did not change.
DeepSeek frames the release as a performance and cost improvement rather than a new generation. It reports higher agentic scores and says the update reduced task runtime and output length. For agent workloads, Novita AI notes, fewer decode tokens and tool-call round trips mean a lower bill for the same completed job.
The headline figure is DeepSWE v1.1, which DeepSeek reports at 77.0, up from 74.2 for the September release. Novita AI stresses that this is the only result that can be compared directly with the earlier release, because it uses the same benchmark version and metric.
- Terminal-Bench 4.0 (Pass@1): 42.3
- ProgramBench: 29.1
- NL2Repo-Bench: 68.2
- Automation-Bench v1.0.6: 60.5
- Agents' Last Exam: 35.4
- Apex Bench v1.1: 72.1
These numbers come from the DeepSeek API Change Log. DeepSWE v1.1 uses the Mini-SWE-V2 harness, and Terminal-Bench 4.0 uses the Claude Code framework. Older material lists Terminal-Bench 2.1 at 90.6 for September, but Novita AI warns that this is a different task set on a different harness, so placing the two side by side would suggest a regression that the results do not show.
Specs, pricing and access on Novita AI
The model is available under the ID deepseek/deepseek-v4.1-flash-1017 through Novita AI's OpenAI-compatible API at https://api.novita.ai/openai. Anthropic-compatible and Responses endpoint families are also listed.
- Input price: $0.3 per 1M tokens
- Cache-read price: $0.006 per 1M tokens
- Output price: $1.2 per 1M tokens
- Context window: 1,048,576 tokens
- Max output: 393,216 tokens
- Input: text and image; output: text
Novita AI lists serverless hosting, function calling, structured outputs and reasoning for the model. The specs were checked on October 19, 2026, according to the post.
The earlier ID, deepseek/deepseek-v4.1-flash, stays online at the same price, context and output limit. Novita AI describes 1017 as a snapshot to adopt, not a forced migration. Teams pinned to the older ID can stay there, and moving over is a model ID change in configuration, not an integration change. The post recommends keeping the ID in config so the next snapshot is a settings change, and keeping the previous ID as a rollback target.
DeepSeek's own platform points the alias deepseek-flash at its latest build, but Novita AI says that alias does not apply on its service, where every request names an explicit model ID.
Search API is separate
DeepSeek opened its in-house Search and Fetch API on October 17, 2026, the same day as the model update. Novita AI states that this is a DeepSeek platform API and not a Novita-hosted feature, and that its 1017 listing includes no hosted search or fetch. Developers who need grounded web results would have to plan retrieval as a separate integration.
Who should consider it
Novita AI recommends 1017 as the default for coding agents, terminal automation and tool-use pipelines. It also points to long-context agent sessions, such as large repositories and long tool-call histories, and to workflows that mix screenshots or document pages with text instructions.
The post advises against it for short-prompt classification, chat completion and other high-volume text tasks with no agentic or multimodal need. Its price matches the earlier V4.1 Flash entry, which Novita AI says is already above base DeepSeek V4 Flash, and the post sees nothing that justifies that premium for simple requests.
Novita AI also cautions that every 1017 figure comes from DeepSeek's own harnesses on a fixed setup. Tool-call schema, retry behavior and context length will shift real results, so it suggests a smoke test of one agent loop, one tool-use turn and one image request. Teams should then compare cost per completed task against their baseline, and measure tokens per completed task and wall-clock time on their own traces.
