Is DeepSeek Still Worth It After the 3x Price Rise? V4 Flash vs GPT-5.6 Luna vs Gemini 3.8 Flash (September 2026)

TL;DR: DeepSeek is still the cheapest frontier-class API, but only off-peak. Since August 16, 2026, V4 Flash costs $0.22 in / $0.66 out off-peak and $0.44 / $1.32 during seven UTC peak hours on weekdays. An 80K-in, 20K-out agent session runs $0.031 off-peak, $0.062 at peak, $0.040 on GPT-5.6 Luna, $0.135 on Gemini 3.8 Flash and $0.18 on Claude Haiku 4.5. US teams never touch peak during work hours; Singapore teams sit in it six hours a day. With a repeated prefix, DeepSeek’s $0.007 cache-hit rate still beats everyone by a wide margin.

What actually changed on August 16

DeepSeek did not just raise prices. It split the day into peak and off-peak and doubled the rate during peak. The DeepSeek pricing page (checked September 4, 2026) lists peak as 01:00-04:00 and 06:00-10:00 UTC, Monday through Friday. Every other hour, including all weekend, is off-peak, and off-peak is exactly half of peak.

Here is the before and after. Previous list prices are as recorded in the September 2026 API pricing roundup.

So “3x” is the peak story. Off-peak, the blended rise on an 80/20 input/output workload is closer to 1.8x. Output took the biggest hit, which matters for agents and code generation far more than for classification.

Also on that page: 1M-token context on both models, up to 384K output tokens, and both OpenAI-format and Anthropic-format endpoints. The rate limit page caps concurrency at 2,500 simultaneous requests for V4 Flash and 500 for V4 Pro per account, with 429 errors beyond that and a no-cost expansion form.

Does your usage land in peak? Five time zones mapped

Peak is defined in UTC, so whether you pay double depends on where your users are awake. September offsets (US Pacific UTC-7, US Eastern UTC-4, London UTC+1, Bangalore UTC+5:30, Singapore UTC+8), 09:00-17:00 local working day:

A US-facing product with daytime users almost never pays the peak rate. One caveat for Pacific: peak runs Monday to Friday in UTC, so Sunday 18:00-21:00 Pacific is Monday peak in UTC and costs double. Evening US consumer usage (7 pm to 2 am Pacific) does land in peak on weeknights.

A Singapore or Bangalore B2B workload is the opposite. Most of the working day is at the doubled rate, and the “DeepSeek peak” rows below are the ones that apply.

Peak covers 7 hours a day times 5 days, or 35 of 168 weekly hours. If usage is flat around the clock, 20.8% of tokens are billed at peak, and the blended rate is 1.21x off-peak.

Per-task cost: chatbot reply

Assumptions: 1,500 input tokens (system prompt plus short history), 300 output tokens. Prices per 1M tokens: DeepSeek V4 Flash $0.22 / $0.66 off-peak and $0.44 / $1.32 peak; GPT-5.6 Luna $0.20 / $1.20; Gemini 3.8 Flash $0.75 / $3.75 (promotional through December 31, 2026); Claude Haiku 4.5 $1 / $5. All checked September 4, 2026.

Arithmetic for DeepSeek off-peak: 1,500 x $0.22 / 1M = $0.00033 in, 300 x $0.66 / 1M = $0.000198 out, total $0.000528. Multiply by 1,000 for the table.

Off-peak DeepSeek beats Luna by 20%. Peak DeepSeek loses to Luna by 60%. Gemini 3.8 Flash and Haiku 4.5 are not in the same bracket for short chat at all: 4.3x and 5.7x more than DeepSeek off-peak.

Per-task cost: 10K-token document

Assumptions: 10,000 input tokens, 1,000 output tokens (summary or extraction). Input-heavy, which favours anyone with a cheap input rate.

Luna’s $0.20 input is cheaper than DeepSeek’s $0.22 off-peak input. DeepSeek wins the document job only on output, $0.66 against $1.20. At peak, Luna wins by 44%.

Per-task cost: 80K/20K agent session, the worked example

This is the workload that decides most budgets in 2026. A coding or research agent that accumulates 80,000 input tokens across tool calls and emits 20,000 output tokens.

DeepSeek V4 Flash off-peak: 80,000 x $0.22 / 1M = $0.0176 in. 20,000 x $0.66 / 1M = $0.0132 out. Total $0.0308.

DeepSeek V4 Flash peak: $0.0352 + $0.0264 = $0.0616.

GPT-5.6 Luna: 80,000 x $0.20 / 1M = $0.016. 20,000 x $1.20 / 1M = $0.024. Total $0.040.

Gemini 3.8 Flash: $0.060 + $0.075 = $0.135.

Claude Haiku 4.5: $0.080 + $0.100 = $0.180.

The same session at the old V4 Flash price was 80,000 x $0.14 / 1M + 20,000 x $0.28 / 1M = $0.0168. Off-peak is 1.83x that, peak is 3.67x. That is the honest version of “DeepSeek tripled.”

If agent runs are spread evenly across the week, the blended cost is 0.792 x $0.0308 + 0.208 x $0.0616 = $0.0372 per session. That is 7% under Luna. DeepSeek Flash and Luna are effectively the same price for round-the-clock agent work now, and Luna has no time-of-day surprise.

V4 Pro at peak costs more than Haiku 4.5 on this workload. Pro only makes sense off-peak, and only if Flash cannot do the job.

The cache-hit angle: where DeepSeek still runs away with it

Every vendor discounts cached input. Nobody discounts it like DeepSeek. The context caching page says caching is on by default with no code changes, and any request whose prefix matches a persisted cache unit is billed at the cache-hit rate: $0.007 per 1M off-peak and $0.014 at peak on V4 Flash. There is no separate cache-write price on the pricing page.

Compare: Luna cached input is $0.02 per 1M with a $0.25 cache write, per the OpenAI pricing page. Haiku 4.5 cached input is $0.10 with a 1.25x write on the five-minute cache. Gemini 3.8 Flash’s cached rate could not be verified, so it is shown at full input price below; treat that row as a ceiling.

Workload: a support or product-docs bot with a fixed 30,000-token prefix (system prompt plus knowledge base), 500 new input tokens per turn, 300 output tokens per turn. Assume the prefix hits cache on every turn.

DeepSeek off-peak per turn: 30,000 x $0.007 / 1M = $0.00021 cached, 500 x $0.22 / 1M = $0.00011 new, 300 x $0.66 / 1M = $0.000198 out. Total $0.000518.

Luna per turn: 30,000 x $0.02 / 1M = $0.0006 cached, 500 x $0.20 / 1M = $0.0001 new, 300 x $1.20 / 1M = $0.00036 out. Total $0.00106.

Even at peak, DeepSeek ties Luna on a cached workload. Off-peak it is half the price, and roughly 10x cheaper than Haiku. The bigger the fixed prefix, the wider the gap: $0.007 is 65% below Luna’s $0.02 and 93% below Haiku’s $0.10. This is the workload where the August rise barely registered; a 30K cached prefix costs $0.21 per 1,000 turns off-peak, and the bill is dominated by the 300 output tokens.

From the caching docs: The cache matches complete prefix units, so keep the static part of the prompt at the front and byte-identical across calls. And DeepSeek persists common prefixes it detects across requests, so slightly different prompts with the same opening still benefit after a couple of calls. If you route through an aggregator, check that it preserves the prefix; the OpenRouter pricing guide covers where caching survives.

BetOnAI Verdict

Stay on DeepSeek V4 Flash if either of these is true: your users are US-daytime (you are structurally off-peak, and it is the cheapest frontier-class model you can buy), or your workload has a large repeated prefix (RAG, support bots, document Q&A over the same corpus), where the $0.007 cache hit makes the time of day almost irrelevant. Either way, ignore the headline “3x.”

Move to GPT-5.6 Luna if your usage is Asia-daytime and not cacheable. Singapore and Bangalore teams running agents during business hours are paying $61.60 per 1,000 sessions on DeepSeek against $40.00 on Luna, and Luna’s $0.20 input is already cheaper than DeepSeek’s off-peak $0.22. Luna is also the default for anyone who cannot tolerate a bill that doubles on a schedule. The OpenAI vs Anthropic vs Google comparison has the full Luna tier breakdown.

Move to Gemini 3.8 Flash only if you need something DeepSeek and Luna do not offer, and you accept paying 3.9x to 4.4x more per task than DeepSeek off-peak. At $0.75 / $3.75 until December 31, it is a quality play, not a price play, and the listed rate doubles to $1.50 / $7.50 on January 1, 2027, so any Gemini budget set today is wrong by 2x in four months.

Three things to do now. One: log the UTC hour on every DeepSeek call for a week and compute your peak share; under 25%, do nothing. Two: over 50%, A/B Luna on the same prompts for a day and move the non-cached calls. Three: anything with a fixed prefix over 10K tokens goes on DeepSeek regardless of time zone; check the cache-hit token count in the usage response to confirm it is hitting. The AI Pricing Watch hub tracks every vendor change month by month.

Frequently Asked Questions

How much does DeepSeek V4 Flash cost after the August 2026 price change?

$0.22 per million input tokens and $0.66 per million output off-peak, doubling to $0.44 and $1.32 during peak hours, which are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays. Cache hits are $0.007 per million off-peak. Before August 16 the list price was $0.14 / $0.28.

Is DeepSeek still cheaper than GPT-5.6 Luna?

Off-peak, yes: an 80K-in, 20K-out agent session costs $0.031 on DeepSeek V4 Flash against $0.040 on Luna. At peak DeepSeek costs $0.062, 54% more than Luna. Luna’s $0.20 input rate is already below DeepSeek’s off-peak $0.22, so input-heavy work is a near tie even off-peak.

When does DeepSeek’s peak pricing apply in my time zone?

US Pacific and Eastern working hours never touch peak. London overlaps for about two working hours, Bangalore for about 4.5 and Singapore for six of eight. If your usage is flat around the clock, 20.8% of tokens land in peak and the blended rate is 1.21x off-peak.

Sources

• DeepSeek, Models & Pricing: https://api-docs.deepseek.com/quick_start/pricing
• DeepSeek, Rate Limit & Isolation: https://api-docs.deepseek.com/quick_start/rate_limit
• DeepSeek, Context Caching: https://api-docs.deepseek.com/guides/kv_cache
• OpenAI, API Pricing: https://developers.openai.com/api/docs/pricing
• Google, Gemini API Pricing: https://ai.google.dev/gemini-api/docs/pricing
• Anthropic, Claude Pricing: https://platform.claude.com/docs/en/about-claude/pricing
• BetOnAI, AI API Pricing September 2026: https://betonai.net/ai-api-pricing-september-2026/
• BetOnAI, OpenRouter Pricing Guide: https://betonai.net/openrouter-pricing-2026-complete-guide-to-every-model-tier-and-hidden-cost/
• BetOnAI, OpenAI vs Anthropic vs Google API Pricing: https://betonai.net/openai-vs-anthropic-vs-google-api-pricing-2026/
• BetOnAI, AI Pricing Watch hub: https://betonai.net/ai-pricing-watch-2026-the-hub/