TL;DR: The price increase Bloomberg reported in August arrived on September 10, 2026 as a model swap: V4-Flash ($0.14/$0.28 per million tokens) was retired and its API name now silently serves V4.1-Flash at $0.30/$1.20 peak — but cache hits dropped to $0.003–0.006/M and US working hours fall almost entirely in the half-price off-peak window. A typical solo Cline user’s bill moves from roughly $1.85 to $1.20–2.40 a month, which is still not a reason to buy a $1,300 GPU.

	Stay on deepseek-flash	Switch to MiniMax M3	Go local (used RTX 3090)
Best for	Solo/US-hours devs, cache-heavy Cline sessions	Output-heavy agent runs during peak UTC hours	Code that can’t leave the machine
Price / Cost	$0.15–0.30/M in, $0.60–1.20/M out	$0.30/M in, $1.20/M out, flat	~$1,300 up front + ~$10/mo power
The catch	Peak/off-peak windows; output 2–4× pricier than V4-Flash	No off-peak discount; long-context tier doubles above 512K	Never pays for itself on cost alone at these API prices

What did DeepSeek actually change on September 10, 2026?

DeepSeek retired V4-Flash and launched V4.1-Flash as the new deepseek-flash endpoint at 04:00 UTC on September 10, 2026 — with a new rate card that raises sticker prices while collapsing cache-read costs. Bloomberg’s August 6 report said DeepSeek “plans a significant API price increase” with no figure or date; this release is what it turned out to mean in practice.

The verified rate card, from DeepSeek’s official pricing page as of September 17, 2026:

Per 1M tokens	V4-Flash (until Sep 10)	V4.1-Flash peak	V4.1-Flash off-peak
Input, cache miss	$0.14	$0.30 (+114%)	$0.15 (+7%)
Input, cache hit	— (old card retired)	$0.006	$0.003
Output	$0.28	$1.20 (+329%)	$0.60 (+114%)

Peak hours are narrow: Monday–Friday, 01:00–04:00 and 06:00–10:00 UTC — Chinese daytime. Every other hour, including all weekend, is off-peak at half price. That definition matters more than the sticker numbers if you code from the US, and we’ll get to why.

Two more changes ride along, per DeepSeek’s changelog:

• The legacy model names deepseek-v4-flash and deepseek-v4-flash-vision-exp still resolve, but requests are served by V4.1-Flash and billed at the new Flash prices. There is no error, no header, no warning.
• V4 Pro survives after all. The original plan routed deepseek-v4-pro to V4.1-Flash from September 14; after user pushback, DeepSeek kept V4 Pro’s API alive with billing unchanged.

How much does the change cost a real Cline user per month?

For a typical solo developer, between a wash and about 30% more — the scary-looking +329% on output tokens mostly doesn’t survive contact with how agentic coding tools actually spend tokens. Cline and Cursor BYOK sessions resend a growing conversation context every turn, and DeepSeek’s context caching is automatic, so the bulk of your input tokens are cache hits — which now cost 46× less than V4-Flash’s old cache-miss rate.

Here’s the math for two workload shapes, 22 working days a month, assuming a 70% input cache-hit rate (conservative for Cline, where each turn resends prior context):

Monthly cost	Solo dev (500K in + 50K out/day)	Heavy agent user (5M in + 500K out/day)
V4-Flash, old flat rates	$1.85	$18.48
V4.1-Flash, all peak	$2.36 (+28%)	$23.56 (+27%)
V4.1-Flash, all off-peak	$1.18 (−36%)	$11.78 (−36%)
V4.1-Flash, peak, zero cache hits	$4.62 (+150%)	$46.20 (+150%)

Check the math against your own usage — the arithmetic is three lines:

python3 - <<'EOF'
miss, hit, out = 3.3, 7.7, 1.1   # millions of tokens/month
peak = miss*0.30 + hit*0.006 + out*1.20
off  = miss*0.15 + hit*0.003 + out*0.60
print(f"peak: ${peak:.2f}/mo   off-peak: ${off:.2f}/mo")
EOF

Expected output:

peak: $2.36/mo   off-peak: $1.18/mo

The structural change is where the money goes. Under V4-Flash, output was about 17% of the bill above; under V4.1-Flash at peak with caching, output is 56% of it. If your workload is output-heavy — big diff generation, whole-file rewrites, verbose agents — you feel the full 2–4× on the line that now dominates. If it’s context-heavy and cache-friendly, your bill barely moved, and off-peak it went down.

Boundary worth stating: the zero-cache row is real for some setups. Round-robin API gateways and some proxy layers break DeepSeek’s automatic caching by splitting a session across cache namespaces. If you route through a middleware that rewrites requests, measure your actual cache-hit rate in the DeepSeek platform usage dashboard before trusting the cached math.

Why do US developers mostly dodge the peak rates?

Because DeepSeek’s peak window is Chinese daytime, and a US workday doesn’t touch it. Peak is Mon–Fri 01:00–04:00 and 06:00–10:00 UTC. A developer coding 9am–6pm Eastern (13:00–22:00 UTC) spends zero working hours in the peak window — every session bills at $0.15 cache-miss input, $0.003 cache-hit, $0.60 output. Same for Pacific time. Weekends are off-peak everywhere.

European developers clip it: a 9am–6pm CEST day (07:00–16:00 UTC) overlaps peak for three hours each morning. If you run scheduled agents from a European server, shifting cron jobs past 10:00 UTC halves their token cost.

For this site’s mostly-US readership, the honest framing of “DeepSeek’s price hike” is: input got 7% more expensive, output doubled, and cache reads became nearly free. That’s a targeted hike on verbose agentic output, not a broad one.

Is V4.1-Flash still the cheapest usable Cursor/Cline backend?

For cache-heavy agentic work during US hours, yes. For flat-rate predictability or peak-hour output-heavy runs, MiniMax M3 now ties or wins. Prices below verified September 17, 2026:

Per 1M tokens	V4.1-Flash (off-peak / peak)	MiniMax M3	Qwen3.6-35B-A3B	Kimi K3
Input (miss)	$0.15 / $0.30	$0.30	from $0.05–0.10 (varies by provider)	$3.00
Input (cache hit)	$0.003 / $0.006	$0.06	varies	$0.30
Output	$0.60 / $1.20	$1.20	$0.15–0.70 (varies by provider)	$15.00
Context	1M	1M (rates double above 512K input)	262K	1M

Reading that table honestly:

• MiniMax M3 matches V4.1-Flash’s peak rate around the clock, with no time-of-day gymnastics. Its $0.06 cache reads are 10–20× pricier than DeepSeek’s, so DeepSeek still wins long cached sessions, but M3 is the simpler bill.
• Qwen3.6-35B-A3B is the new bargain floor — OpenRouter lists it from $0.05/M input, Alibaba Cloud from $0.10/M — but it’s a 35B-total/3B-active model with a 262K window. It’s the “cheap baseline” pick for routine edits, not a V4.1-Flash-class agent driver.
• Kimi K3 at $3/$15 is playing a different sport; it’s there to show the spread. Frontier-labeled quality still costs 10–25× more per token than the Chinese efficiency tier.
• Codestral 2 remains the Apache 2.0 pick if your endgame is self-hosting completion models; we covered its rates in the linked review.

The site’s broader ranking is in the AI code editor cost comparison; the DeepSeek rows of that article now describe a retired model, and this is the correction of record until it’s revised.

The trap: deepseek-v4-flash in your Cline config no longer means V4-Flash

We caught this while re-verifying our own V4-Flash backend guide for this article: every Cline profile pinned to deepseek-v4-flash since that June guide has, per DeepSeek’s changelog, been served V4.1-Flash at the new billing since September 10 — no error, no deprecation warning, no config change. A pinned model ID is normally how you avoid silent model swaps. DeepSeek repurposed the pin.

The fix is to make the migration explicit rather than accidental:

1. In Cline (or Cursor → Settings → Models → OpenAI-compatible), change the model ID from deepseek-v4-flash to deepseek-flash. Same model either way — but the name now says what you’re getting, and the legacy alias is marked temporary, so it’s a future breakage you’re deleting.
2. Re-run your usual smoke test on an agentic task before trusting it with a long session. V4.1-Flash is a new architecture that folds the chat/reasoning split into one model; our V4 guide’s advice about disabling thinking mode in Cline was written against V4 behavior, so re-test rather than carry the old workaround forward blindly.
3. Check the usage dashboard’s cache-hit ratio after a few sessions. Above ~60% on input, the new card is cheaper than the old one for most US-hours workloads; a near-zero ratio means something in your stack is defeating the cache, and you’re on the +150% row of the table above.

Does the hike make a local GPU pay off now?

No — the opposite. Nearly-free cache reads moved the payoff line further away, and this is coming from a site that publishes local-LLM setup guides every week.

The buy-a-GPU math, with all inputs verified this month: a used RTX 3090 24GB has a fair asking range of $1,287–$1,411 on eBay US as of September 2026 (ResalePrices; trend +7.5% over 30 days — 3090s are getting more expensive). Call it $1,300. Running it ~8 hours of coding days at 350W adds roughly 62 kWh/month — about $9–11/month at the 15–18¢/kWh most US residential customers pay.

Against the verified API bills:

• The solo-dev workload above costs $1.18–2.36/month on V4.1-Flash. A $1,300 card that saves you at most ~$2.40 a month while costing ~$10 in electricity never breaks even. Not slowly — never.
• The heavy agent workload ($11.78–23.56/month) nets maybe $14/month in savings after power at peak-hour rates. That’s a ~93-month payback — about 7.7 years, on a card that’s already two generations old.
• The two-year break-even threshold sits around $65/month of API spend. At V4.1-Flash prices that’s roughly 25× the heavy-user workload above. If you’re spending that, you’re likely running parallel agent fleets — and you should audit output verbosity before buying hardware.

What a local card actually buys post-September-10 is not savings: it’s privacy (code that contractually can’t leave the machine), offline capability, and zero marginal cost for unlimited experimentation. Those are real — they’re just a different purchase decision, covered in our hardware-tier guide for local coding models and runaihome’s local model VRAM guide.

If you’re privacy-motivated and undecided, rent before you buy: RTX 3090s start around $0.07/hr on Vast.ai (marketplace pricing, September 2026 — floating, so treat it as a floor). Twenty hours of testing your actual Cline workload on rented hardware costs less than $2 and answers the question a spec sheet can’t.

What to actually buy

Prices as of September 2026, all taken from the comparison above:

Your situation	The move	Price	Where
Solo dev, US hours, cache-friendly Cline sessions	Stay on deepseek-flash, update the model ID	~$1–3/mo	DeepSeek platform
Output-heavy agents running during peak UTC hours	MiniMax M3, flat $0.30/$1.20	~$5–25/mo	Our M3 setup guide
Code can’t leave the machine	Used RTX 3090 24GB	~$1,287–1,411	Check price
Privacy-curious, want to test the workload first	Rented RTX 3090	from $0.07/hr	Vast.ai

Honest take

DeepSeek executed the least painful “significant price increase” possible: it shipped a better model, doubled-to-quadrupled the sticker rates, then handed most of the increase back through $0.003 cache reads and an off-peak window that covers the entire American workday. Solo developers on Cline will struggle to see the difference on a bill measured in single-digit dollars. The people who genuinely pay more are running verbose agents at scale during Chinese business hours — and they have a one-line fix (MiniMax M3) or a scheduling fix (shift past 10:00 UTC).

The verdict we’d defend: stay on deepseek-flash if you’re US-based and cache-friendly; switch output-heavy peak-hour workloads to MiniMax M3; and don’t let a $1–15/month price change talk you into a $1,300 GPU — buy local hardware for privacy, not for this.

FAQ

Did DeepSeek raise prices or cut them? Both, depending on the token type. Versus V4-Flash’s $0.14/$0.28: cache-miss input rose 7% (off-peak) to 114% (peak), output rose 114–329%, and cache-hit input fell to $0.003–0.006/M — cheaper than any comparable API’s cache tier as of September 2026.

Is the old V4-Flash still available anywhere? Not from DeepSeek. V4-Flash and V4-Flash Vision Exp were retired on September 10, 2026; the legacy API names temporarily route to V4.1-Flash at V4.1 prices. V4 Pro remains available with unchanged billing after DeepSeek reversed its planned September 14 shutdown.

Do I need to change my Cline or Cursor config? It works unchanged, but update the model ID to deepseek-flash anyway: the legacy deepseek-v4-flash alias is documented as temporary, and you’re already being served and billed as V4.1-Flash either way.

What’s the cheapest API backend for Cline now? For cache-heavy sessions in US hours: V4.1-Flash off-peak (effective ~$0.05–0.10/M blended input with 70% cache hits, $0.60/M output). For flat predictable pricing: MiniMax M3 at $0.30/$1.20. For routine small edits: Qwen3.6-35B-A3B from $0.05–0.10/M input via OpenRouter or Alibaba Cloud.

Does a used RTX 3090 make sense for local coding models in late 2026? For cost savings, no — at $1,287–1,411 (September 2026, trending up) it can’t pay back against single-digit monthly API bills. For privacy, offline work, or unmetered experimentation, yes — see the situations table above, and rent on Vast.ai first (from ~$0.07/hr) to validate your workload.

Sources

• Models & Pricing — DeepSeek API Docs (official)
• Change Log — DeepSeek API Docs (official; V4.1-Flash launch, legacy-name routing, V4 Pro reversal)
• DeepSeek-V4.1-Flash debuts with $0.003/1M off-peak cached-input rate — VentureBeat
• DeepSeek launches V4.1-Flash with ultra-cheap off-peak API pricing — TechBriefly
• DeepSeek unifies three modes: V4.1 Flash launches, V4 Pro plan reversed — AIbase
• MiniMax M3 — API pricing, OpenRouter
• Kimi API Pricing (September 2026): Kimi K3 at $3/$15 — BenchLM
• Qwen3.6 35B A3B — provider pricing, OpenRouter
• RTX 3090 used price, fair asking range — ResalePrices (September 2026)
• Vast.ai GPU marketplace pricing

Last updated September 17, 2026. Pricing and peak-window definitions verified against the sources above on September 17, 2026; DeepSeek changes rates without notice — verify before committing a team budget.

Know which coding tool is worth paying for

Hands-on comparisons of AI coding assistants and what each one costs to run — including the local-model path. Sent only when something changes. Unsubscribe anytime.

Email address Subscribe