DeepSeek upgraded V4.1 Flash to DeepSeek-V4.1-Flash-1017 on October 17, 2026. The official API still uses deepseek-flash, and token pricing is unchanged. DeepSeek reports stronger coding and agent results, shorter outputs and faster task completion. Those gains are vendor-reported; they do not establish a universal improvement for every coding workflow.
For an existing application, the immediate issue is version identity: keeping the same model string does not keep the September model behind it.
Existing Flash API calls now reach the new model
DeepSeek’s October 17 changelog confirms that its website, app and API have moved to 1017. It recommends DeepSeek Harness 0.2.1-rc.1 or later for the new release.
The API integration documentation says deepseek-v4-flash and deepseek-v4-flash-vision-exp remain accepted as legacy names, but their original models are retired. Those requests now reach 1017 and are billed at the Flash rate. Changing back to a legacy alias therefore does not provide an old-model control group.
Use deepseek-flash for a new configuration. Record the request date, provider, endpoint and response metadata when evaluating it. A third-party provider’s similarly named model should be checked against that provider’s own revision information; DeepSeek’s routing statement describes its official API.
The benchmark gains have different sizes and conditions
The released model card places 1017 beside the previous V4.1 Flash. Selected scores are below. The change column is EyesTech arithmetic from those published endpoints, expressed in score points rather than relative percentage gains.
| Benchmark | Previous Flash | 1017 | Change |
|---|---|---|---|
| Terminal-Bench 4.0 | 31.2 | 42.3 | +11.1 |
| DeepSWE v1.1 | 74.2 | 77.0 | +2.8 |
| ProgramBench | 20.3 | 29.1 | +8.8 |
| NL2Repo-Bench | 65.4 | 68.2 | +2.8 |
DeepSeek identifies Claude Code as the harness for Terminal-Bench 4.0 and Mini-SWE-V2 for DeepSWE v1.1. The changelog also specifies Pass@5 for ZeroBench-main. A score from five attempts should not be read as a one-attempt success rate, and Terminal-Bench releases are separate evaluations rather than one interchangeable scale.
The official launch announcement attributes project-development improvements to reinforcement learning with self-feedback and shows games and 3D scenes. These examples document what the company chose to demonstrate. They do not establish the rate at which an unfamiliar repository can be completed without human repairs.
Unchanged pricing still has peak and off-peak rates
The current pricing table lists Flash input cache misses at $0.15 per million tokens off-peak and $0.30 peak. Output costs $0.60 off-peak and $1.20 peak; cache-hit input is $0.003 and $0.006 respectively.
Peak periods are 01:00–04:00 and 06:00–10:00 UTC on weekdays, excluding Chinese public holidays. Other hours, weekends and those holidays are off-peak. A single price without its time band is incomplete.
For a transparent example, 100,000 uncached input tokens plus 20,000 output tokens costs $0.027 off-peak: (0.1 × $0.15) + (0.02 × $0.60). At peak rates it is $0.054. This is an illustrative token calculation, not a measured task cost; retries, changing context, cache hits and tools change the bill.
Search and Fetch are separate request-priced services. The launch announcement says each costs $0.75 per 1,000 calls, while qualifying server-side tool calls through the named agent integrations remain free until October 31, 2026, 16:00 UTC. Keep those charges separate from model tokens when estimating a research agent’s cost.
Self-hosting needs the right prompt format
The weight release is MIT-licensed, but the model card says it does not include a Jinja chat template. It supplies an encoding reference and points production integrations to deepseek-recipe. That toolkit handles protocol encoding and parsing; inference and tool execution remain the caller’s responsibility.
The encoding guide documents V4.1’s tool markup, numeric reasoning budgets and mid-conversation system messages. These are V4.1 format rules, not necessarily new inventions in the October refresh. Reusing a V4-era formatter without checking those rules can turn a deployment mismatch into an apparent model-quality problem.
Evaluate the task, model and harness together
For a useful comparison, preserve a small set of real bug reports or feature requests with fixed repository revisions and acceptance checks. Keep tool access, time limits, reasoning settings and network permissions consistent. Record the final result, human corrections, total elapsed time, token usage and tool charges for each attempt.
Separate model and harness changes where the provider allows it. If you upgrade both at once, describe the outcome as a workflow comparison. If an old API snapshot is unavailable, retain the old traces and acknowledge that the comparison cannot be rerun against the same hosted model.
The release provides a clear reason to re-evaluate Flash: higher reported coding scores at the same token rates. Evidence of fewer failed tasks and less repair work on the reader’s own workload would establish the practical gain. A shorter answer or a successful showcase alone would not.
Publisher information: Prepared with AI assistance from DeepSeek’s official release, model and API documents, checked October 18, 2026. Benchmark scores are vendor-reported. Score differences and token-cost examples are EyesTech calculations; no independent model benchmark was performed. The featured image is an original AI-generated conceptual illustration.
