What the DeepSeek Roadmap Looks Like After the V4 Preview

If you maintain an integration with the DeepSeek API, the DeepSeek roadmap has moved quickly since the V4 Preview of 24 April 2026: an official V4-Pro, a new V4.1-Flash architecture with vision, peak and off-peak pricing, and the end of the old deepseek-chat / deepseek-reasoner names. This guide is for engineers and technical buyers who need to know what changed, what DeepSeek has actually announced next, and which dates belong on the planning board. It is built from DeepSeek’s own Change Log, news pages, pricing page and model cards.

The short version of the DeepSeek roadmap

As of 24 October 2026 the API lists two models: deepseek-flash, served by DeepSeek-V4.1-Flash-1017 (released 17 October 2026, with vision), and deepseek-v4-pro, served by DeepSeek-V4-Pro-0813 (released 13 August 2026, no vision). Both have a 1M-token context window, up to 384K output tokens and thinking mode on by default, and both publish their weights under MIT on Hugging Face.

The roadmap items, in date order:

• 24 April 2026 — V4 Preview (deepseek-v4-pro, deepseek-v4-flash); 1M context becomes the default across DeepSeek’s official services.
• 24 July 2026, 15:59 UTC — the scheduled retirement of deepseek-chat and deepseek-reasoner; neither is listed in the official docs any more.
• 31 July 2026 — DeepSeek-V4-Flash-0731, the official V4-Flash, in API public beta.
• 13 August 2026 — DeepSeek-V4-Pro GA (V4-Pro-0813) on app, web and API; low / high / max thinking effort; native Responses API.
• 16 August 2026, 16:00 UTC — peak and off-peak pricing takes effect, with off-peak at half the peak rate.
• 10 September 2026 — DeepSeek-V4.1-Flash as deepseek-flash, with native image understanding; V4-Flash and V4-Flash-Vision-Exp retired; prices reduced.
• 17 October 2026 — DeepSeek-V4.1-Flash-1017 (pricing unchanged); Search and Fetch APIs open to everyone.
• 31 October 2026, 16:00 UTC — announced start of billing for Search and Fetch calls made through server-side tools in DeepSeek Harness, Claude Code and Codex.

What happened to the “V4 Preview”

The April 2026 preview models — V4-Pro (Preview) with 1.6T total / 49B active parameters and V4-Flash (Preview) with 284B total / 13B active, figures that belong to the preview only — have both been superseded. The 31 July 2026 Change Log said “The official release of DeepSeek-V4-Pro will follow soon”; on 13 August 2026 the GA release arrived as DeepSeek-V4-Pro-0813, which its model card describes as built on the V4-Pro (Preview) structure with a DSpark speculative-decoding module attached.

The Flash line went further. DeepSeek-V4-Flash-0731 kept the preview’s architecture and size and was only re-post-trained, but on 10 September 2026 DeepSeek replaced it with DeepSeek-V4.1-Flash, which the Change Log calls “the smallest model in our new architecture family”, designed for “a higher capability ceiling, faster inference, higher throughput, and scaling to larger models”. Its news page describes a 552B-parameter MoE with 8B active parameters for input and 16B for output. The 17 October update, DeepSeek-V4.1-Flash-1017, is what deepseek-flash serves today.

Migration: legacy IDs to current model IDs

If any of your code still sends an older model name, treat the switch as a deliberate one-line change. Chat requests hit POST /chat/completions at https://api.deepseek.com (OpenAI format) or https://api.deepseek.com/anthropic (Anthropic format), and only the model field needs to change. DeepSeek’s own guidance from the 10 September 2026 Change Log: “Change the model name to deepseek-flash to call the latest V4.1 Flash model.”

Mapping table

Two things to remember on every request: the API is stateless, so your client must resend the conversation history every call (the web chat keeps state for you; the API does not), and thinking mode returns reasoning_content alongside the final content. For a step-by-step walkthrough, the DeepSeek API getting started guide covers SDK setup end to end, and the DeepSeek API documentation guide covers the request shape.

Pricing on the roadmap: where the rates go from here

Two pricing changes have landed since the preview. Peak and off-peak pricing took effect at 16:00 UTC on 16 August 2026, and the V4.1-Flash release cut API prices from 04:00 UTC on 10 September 2026; the 17 October update left them unchanged. The rate cards as of October 2026, from the official pricing page (USD per 1M tokens):

Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays; every other hour, including weekends, is off-peak at half the peak rate. deepseek-v4-pro costs 3.3× deepseek-flash on output ($3.96 ÷ $1.20), 4.4× on cache-miss input ($1.32 ÷ $0.30) and about 7.3× on cache-hit input ($0.044 ÷ $0.006), in both periods. The one announced change ahead: the Search and Fetch APIs cost $0.75 per 1,000 successful requests each, and calls made through server-side tools in DeepSeek Harness, Claude Code and Codex, free until now, are billed from 16:00 UTC on 31 October 2026 (00:00 Beijing time on 1 November). Beyond that, DeepSeek has not announced future price changes, and its pricing page says prices may vary. Worked cost examples are in our DeepSeek API pricing guide, and the DeepSeek pricing calculator handles your own shapes.

What V4 is good at — and where it isn’t

DeepSeek’s headline claims for the current models, all DeepSeek-reported:

• DeepSeek-V4-Pro-0813 (13 August 2026 Change Log): HLE 42.7 without tools / 60.0 with tools, Terminal Bench 2.1 87.9, DeepSWE 62.7, Toolathlon-Verified 74.1.
• DeepSeek-V4.1-Flash-1017 (17 October 2026 Change Log): HLE without tools 42.6 (44.0 on the text-only subset), Terminal-Bench 4.0 42.3, DeepSWE v1.1 77.0, NL2Repo-Bench 68.2.

The two sets use different benchmark versions (Terminal Bench 2.1 vs Terminal-Bench 4.0, DeepSWE vs DeepSWE v1.1), so they are not a head-to-head. DeepSeek’s 10 September 2026 news page says tests by multiple parties put V4.1-Flash ahead of V4-Pro on performance, cost, speed and total runtime.

Where the line has changed:

• Multimodality — DeepSeek’s 21 August 2026 Change Log still called V4-Flash “the text model”. Image input arrived with the experimental V4-Flash-Vision-Exp that day and became native in V4.1-Flash on 10 September 2026; today deepseek-flash accepts images, while deepseek-v4-pro does not.
• Long-context efficiency — the April preview card reported that V4-Pro (Preview) needs 27% of the single-token inference FLOPs and 10% of the KV cache of V3.2 at 1M tokens; the V4.1-Flash news page says its KV cache needs a quarter of the previous generation’s HBM and an eighth of its SSD storage.

Hardware and deployment track of the roadmap

DeepSeek’s official pages cited here do not say which chips its current models are trained or served on. What DeepSeek has said about deployment: every current model publishes open weights under MIT on Hugging Face; the V4.1-Flash news page (10 September 2026) promises to “work closely with the open-source community on V4.1-Flash inference support and explore more deployment options”, and invites teams planning “a large-scale deployment with 2,000 GPUs + a storage cluster” to get in touch. For running the weights yourself, the V4-Pro-0813 model card recommends a maximum output length of 384K tokens for the high and max effort levels.

Reasoning-effort modes: a parameter, not a model ID

Since V4, thinking mode is a request parameter rather than a separate model ID. Both deepseek-flash and deepseek-v4-pro accept the same settings, per DeepSeek’s Thinking Mode guide:

1. Thinking (default) — enabled unless you turn it off, with default effort high; returns reasoning_content before the final content.
2. Effort levels — reasoning_effort="low", "high" or "max"; DeepSeek’s 13 August 2026 guidance is low for simple tasks, high for daily agent tasks and max for more complex scenarios.
3. Non-thinking — opt out with extra_body={"thinking": {"type": "disabled"}}; FIM completion (Beta) works only in this mode.

Other parameters worth pinning in your client wrappers: temperature (in non-thinking mode only — thinking mode ignores it — DeepSeek recommends 0.0 for code/math, 1.0 for data analysis, 1.3 for general chat and translation, 1.5 for creative writing), top_p, max_tokens, JSON mode (response_format={"type": "json_object"}, designed to return valid JSON but not guaranteed — include the word “json” plus a small example schema and set max_tokens high), tool calling, streaming, and context caching.

What’s next on the roadmap (and what isn’t confirmed)

DeepSeek has published few forward-looking dates. What it has said, and where:

• Search and Fetch billing — server-side tool calls through DeepSeek Harness, Claude Code and Codex are billed from 16:00 UTC on 31 October 2026 (17 October 2026 news page). This is the only dated future item.
• V4.1-Pro — the 10 September 2026 news page said deepseek-v4-pro requests would route to V4.1-Flash “until V4.1-Pro launches”. DeepSeek gave no date, and its Change Log entry for 10 September 2026 drops the routing plan: DeepSeek decided to keep providing V4 Pro after 14 September 2026 with billing unchanged and said it “will provide further notice should there be any changes”.
• Larger models in the V4.1 architecture — the 10 September 2026 Change Log calls V4.1-Flash “the smallest model in our new architecture family” and says the architecture is designed for “scaling to larger models”; DeepSeek has named no larger model and given no date.
• Legacy Flash names — deepseek-v4-flash and deepseek-v4-flash-vision-exp are routed to V4.1-Flash “temporarily”; DeepSeek has not announced an end date.
• AGI framing — DeepSeek said in its April 2026 preview announcement that it remains committed to longtermism, “advancing steadily toward our ultimate goal of AGI”. Treat that as direction, not a deliverable.

As of 24 October 2026, DeepSeek’s Change Log and news pages announce no V5, no R2 and no “Speciale” variant of V4 or V4.1, and give no date for V4.1-Pro. Track confirmed changes on DeepSeek’s Change Log or our DeepSeek V4 release date page.

Practical next steps for teams

1. Audit your model strings. Grep your codebase for deepseek-chat, deepseek-reasoner, deepseek-v4-flash and deepseek-v4-flash-vision-exp, and move them to deepseek-flash.
2. Pick a tier per workload. Default to deepseek-flash, the only current model with vision; use deepseek-v4-pro only where a measured quality gain on your own tasks justifies 3.3× the output price.
3. Re-check prompts and parameters. Thinking is now on by default; if a path relied on the old non-thinking deepseek-chat, disable thinking explicitly, and note that temperature has no effect in thinking mode.
4. Reset your cost model. Off-peak hours cost half the peak rate, so move batch work outside 01:00–04:00 and 06:00–10:00 UTC on weekdays; the cache-hit/miss split still matters, and the DeepSeek context caching guide explains how to structure prompts to hit it. Budget for Search and Fetch tool calls from 31 October 2026.
5. Compare honestly. If you are benchmarking against Anthropic or OpenAI, the DeepSeek vs Claude and DeepSeek vs ChatGPT comparisons cover the trade-offs.

For background reading on prior generations and how the lab got here, see DeepSeek history and the broader DeepSeek beginner guides.