TL;DR: DeepSeek retired the deepseek-chat and deepseek-reasoner model names on July 24, 2026 — requests using them now error out instead of silently routing to V4. The fix is a two-minute rename to deepseek-v4-flash or deepseek-v4-pro, plus one config flag if you don’t want thinking mode burning your output budget. A 2× peak-hour surcharge is announced but not yet active.

	Update names, stay on DeepSeek API	Route via OpenRouter	Self-host the MIT weights
Best for	Almost everyone already using the API	Multi-provider fallback, no vendor lock	Hard privacy or compliance requirements
Price / Cost	V4-Flash $0.14/M in, $0.28/M out; V4-Pro $0.435/M in, $0.87/M out	Same base rates plus provider spread	~160GB+ memory for V4-Flash; V4-Pro is out of reach
The catch	Announced 2× peak-hour pricing, date TBD	Another middleman sees your prompts	Hardware costs more than years of API usage

Honest take: rename your models today, turn thinking mode off where you don’t need it, and stay on the API — self-hosting V4 only makes sense if you can name the privacy requirement that justifies it.

DeepSeek’s V4 rollout was unusually staged. The weights and preview API landed on April 24, 2026 under a plain MIT license — same terms as V3, no usage caps, no attribution clause (see our DeepSeek V4 Pro review for the model itself). The disruptive part came in July: the official V4 release retired the legacy model names that every Open WebUI, LibreChat, and shell-script integration on the planet had been using since 2024, and DeepSeek announced the first time-of-day pricing scheme from a major LLM provider.

If your self-hosted stack calls the DeepSeek API anywhere — as a fallback behind Ollama, as the brain for a RAG pipeline, as a cheap batch summarizer — here is exactly what broke, how to fix it, and how to schedule around the surge pricing before it switches on.

What actually changed on July 24

Three things, in order of how likely they are to have broken your setup:

1. Legacy model names stopped resolving. As of July 24, 2026 at 15:59 UTC, deepseek-chat and deepseek-reasoner return errors. DeepSeek had been silently forwarding both aliases to V4 since the April preview, so plenty of people didn’t notice they were “on V4” until the alias died. The replacements are deepseek-v4-flash (the everyday model: fast, cheap, 284B MoE with 13B active) and deepseek-v4-pro (the flagship: 1.6T MoE, 49B active, DeepSeek-reported 80.6% on SWE-bench Verified).

2. Thinking mode is now on by default — on both models. The old split was one non-thinking model (deepseek-chat) and one reasoning model (deepseek-reasoner). The new names unify that: both V4 endpoints reason by default. If you migrated a high-volume extraction or classification job from deepseek-chat to deepseek-v4-flash without touching anything else, your latency and output-token bill went up. More on the fix below.

3. Context and output limits were unified. Both models now default to a 1M-token context window with output up to 384K tokens. That’s good news for RAG pipelines that were chunking around the old 128K limit, and quietly dangerous for anything that assumed responses would stay short.

There’s one genuinely new integration path worth knowing about: DeepSeek now runs an Anthropic-format endpoint at https://api.deepseek.com/anthropic, so tools that speak the Anthropic Messages API natively can point at V4 without an OpenAI-compatibility shim.

The ten-minute migration

First, find every place the old names live. From your config or project root:

$ grep -rln "deepseek-chat\|deepseek-reasoner" ~/stacks/
/home/you/stacks/librechat/librechat.yaml
/home/you/stacks/n8n/workflows/email-triage.json
/home/you/stacks/scripts/nightly-digest.sh

The mechanical rename, once you’ve reviewed the hits:

grep -rl "deepseek-chat" ~/stacks/ | xargs sed -i 's/deepseek-chat/deepseek-v4-flash/g'
grep -rl "deepseek-reasoner" ~/stacks/ | xargs sed -i 's/deepseek-reasoner/deepseek-v4-pro/g'

Then verify against the live API before restarting anything:

$ curl -s https://api.deepseek.com/v1/chat/completions \
  -H "Authorization: Bearer $DEEPSEEK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"deepseek-v4-flash","messages":[{"role":"user","content":"reply with ok"}],"max_tokens":10}'

A healthy response comes back with "model":"deepseek-v4-flash" in the body. If you still get an invalid-model error, you’re hitting a cached config — restart the container.

Per-app notes for the common self-hosted stack:

• Open WebUI: Settings → Connections → your DeepSeek endpoint. The model list is fetched from /v1/models, so after July 24 the old names simply vanish from the dropdown — re-select the V4 names in any saved workspace or model preset. Everything else (base URL https://api.deepseek.com/v1, existing API key) is unchanged.
• LibreChat: update the models list under your DeepSeek custom endpoint in librechat.yaml and restart the container. If you followed our LibreChat setup guide, that’s the only file to touch.
• AnythingLLM: the generic OpenAI-compatible provider stores the model name per workspace — update it in workspace settings, not just the global connection.
• Coding agents (aider, Cline, Continue.dev): model strings live in each tool’s own config (.aider.conf.yml, provider profiles, etc.). Same rename, and see the sister-site guide to DeepSeek as a coding backend for per-tool details.

No endpoint URLs change, no API keys rotate, and the OpenAI-compatible request schema is untouched. This really is a rename — with one trap.

The trap: thinking mode doubled my batch job’s bill

Real problem from migrating a nightly digest script: a job that summarized ~200 documents with deepseek-chat used to produce roughly 150–300 output tokens per document. After the blind rename to deepseek-v4-flash, the same job produced several times that, because the model now emits reasoning tokens before every answer — and output tokens are the expensive kind ($0.28/M on Flash, $0.87/M on Pro). Latency per document roughly tripled too.

The fix is to disable thinking for workloads that don’t need it. DeepSeek exposes a request-level toggle for this in the current API reference (the exact parameter has changed between releases, so check the docs for the version you’re on rather than copying a snippet from a March-era blog post). For chat UIs like Open WebUI and LibreChat, put the toggle in the endpoint’s default request parameters so every workspace inherits it. Keep thinking on for the workloads where V4 actually earns its benchmark numbers: multi-step reasoning, agentic coding, long-context analysis.

Rule of thumb after two weeks on the new names: deepseek-v4-flash with thinking off behaves like the old deepseek-chat at nearly the same cost; deepseek-v4-pro with thinking on replaces deepseek-reasoner and is noticeably stronger.

Peak pricing: announced, not yet live — schedule around it now

The headline change from the July release, first reported by TechNode on June 30, is time-of-use pricing: a 2× multiplier on all billing items during two daily windows, 09:00–12:00 and 14:00–18:00 Beijing time. In DeepSeek’s own announcement that means flagship output going from 6 to 12 yuan per million tokens during peak, and Flash from 2 to 4.

The part most coverage missed: as of mid-August 2026 the surcharge is not actually being billed yet. DeepSeek announced the policy alongside the V4 release but has not published an effective date — and on August 6 it separately said it plans a general API price increase soon, also without rates or dates. Treat the current bill as a grace period, not the steady state.

Here’s what the announced windows look like from a self-hoster’s timezone:

Billing window	Beijing (UTC+8)	UTC	US Eastern (EDT)
Peak 1 (2×)	09:00–12:00	01:00–04:00	9:00 PM–12:00 AM
Peak 2 (2×)	14:00–18:00	06:00–10:00	2:00 AM–6:00 AM
Off-peak (1×)	everything else	10:00–01:00	6:00 AM–9:00 PM

That table is unusually kind to the ~65% of you in US timezones: the entire US working day is off-peak. The people who get hit are US evening hobby projects (9 PM–midnight Eastern lands in Peak 1) and — the big one — unscheduled cron jobs. The classic 0 3 * * * “run it at 3 AM local” habit drops US East Coast servers (07:00 UTC) and most EU servers (01:00–03:00 UTC) straight into a 2× window.

The fix costs nothing: pin batch work to UTC and pick an hour in the 10:00–01:00 UTC band.

# Pin batch inference outside both announced peak windows.
# 11:00 UTC = off-peak, and daylight-saving shifts can't drag it into one.
0 11 * * * /usr/local/bin/nightly-digest.sh

For interactive use, don’t bother rescheduling your life around a surcharge that isn’t live. Set a calendar note to re-check your DeepSeek invoice when the effective date is announced, and remember cache-hit input is nearly free ($0.0028/M on Flash) — structuring repeated prompts to share a stable prefix saves more than peak-dodging ever will.

Should you just self-host V4 instead?

The weights are MIT — genuinely, both models, weights and code, confirmed on the official Hugging Face repos (see our LLM license shootout for how rare that clean licensing still is). License is not the obstacle. Physics is.

V4-Pro is a 1.6T-parameter MoE. At 4-bit that’s on the order of 800GB before KV cache — datacenter territory, full stop. V4-Flash at 284B total parameters works out to roughly 160GB at ~4.5 bits per parameter, so even the smaller model wants a multi-GPU rig or an aggressive RAM-offload setup that will crawl. For context: when we priced self-hosting Qwen3-Coder 480B, a 4×A100 community pod on RunPod ran about $5.60/hour — around $90/month for evenings-only use, and that’s the cheap cloud path.

Now the API math. A heavy personal workload — say 5M input and 1M output tokens per day on V4-Flash — costs about $0.98/day, roughly $30/month, less with cache hits. Even if the announced increases double that, the API undercuts any hardware that can hold the model. A single RTX 4090 (24GB) can’t hold V4-Flash at any useful quant, and a used RTX 3090 stack gets you to ~96GB for about the price of three years of this API bill — before electricity. If you want the full hardware-tier breakdown, runaihome.com’s GPU guides cover what V4-class MoE models actually need.

When NOT to stay on the API — because there are real cases:

• Prompts contain data that contractually or legally cannot leave your infrastructure. That’s the one unbeatable argument, and it’s the same logic as our FOSS vs SaaS cost breakdown: self-hosting is a privacy play, not a savings play.
• You’re already past ~10M tokens/day, every day, on Pro-class work — at that volume a rented multi-GPU node starts to pencil out.
• You need deterministic availability. DeepSeek’s API has historically throttled under load during Chinese business hours — which, note, is exactly what the peak windows monetize.

For everyone else: the honest verdict is unchanged from the V4 preview. Run a small model locally for private and latency-sensitive work, and treat the DeepSeek API as the cheap overflow brain — now with new model names and a calendar.

FAQ

Do I need a new API key or endpoint URL for V4? No. Existing keys and https://api.deepseek.com/v1 (OpenAI-compatible) keep working. Only the model names changed — plus there’s a new optional Anthropic-format endpoint at /anthropic for tools that speak that API natively.

Is the 2× peak pricing being charged right now? Not as of mid-August 2026. The windows (09:00–12:00 and 14:00–18:00 Beijing time) and the 2× multiplier are announced, but DeepSeek hasn’t set an effective date. A separate general price increase was also flagged on August 6, 2026 — watch your invoice.

Which V4 model replaces the one I was using? deepseek-chat → deepseek-v4-flash, deepseek-reasoner → deepseek-v4-pro. Both new names enable thinking by default, so disable it on Flash for high-volume extraction or classification jobs to keep costs at old-deepseek-chat levels.

Sources

• TechNode: DeepSeek plans mid-July V4 release with peak-hour pricing (June 30, 2026)
• TheRouter.ai: deepseek-chat → deepseek-v4-flash deprecation guide, July 24 deadline
• Developers Digest: DeepSeek retires deepseek-chat and deepseek-reasoner on July 24
• DeepSeek-V4 open weights (MIT), official Hugging Face org
• tokenkarma: DeepSeek off-peak pricing math for V4-Flash

Recommended Gear

• RTX 4090 — 24GB flagship for the local half of a hybrid local+API stack
• RTX 3090 — the used-market VRAM-per-dollar pick if you’re stacking cards