DeepSeek raised its API prices this month, and if you use it with OpenClaw, you have probably seen the panic threads: “massively increased,” “up to 1,114% for cache hits,” “four times more expensive.” The numbers are real, but the conclusion most people are jumping to is wrong.
Here is the straight version: DeepSeek got meaningfully more expensive, it is still one of the cheapest capable models you can run, and the bills that are actually exploding are caused by usage patterns, not the new rate card. This guide breaks down what changed, what it means for your OpenClaw setup, and how to keep your bill flat.
What Actually Changed on August 16
DeepSeek switched to a peak and off-peak billing model, effective August 16, 2026. Under the new model, every token has two prices: a cheaper off-peak rate and a more expensive peak rate. On top of that, DeepSeek raised the base rates, and it warned developers on August 6 that a larger increase was coming with the V4 Pro launch.
Here is the new rate card for V4 Flash, per 1 million tokens:
- Off-peak: $0.22 cache-miss input, $0.66 output
- Peak: $0.44 cache-miss input, $1.32 output
- Cache-hit input: dramatically cheaper, but the cache-hit price itself went up sharply
And for V4 Pro, per 1 million tokens:
- Off-peak: $0.66 cache-miss input, $1.98 output
- Peak: $1.32 cache-miss input, $3.96 output
Yes, those are real increases. Before the change, V4 Flash was around $0.14 input and $0.28 output. The Reddit thread about cache-hit prices going up as much as 1,114% blew up for good reason: if your workload leans on prompt caching, that specific line item got a lot more expensive.
Why “Expensive” Is Still the Wrong Word
Here is the part the panic threads leave out. Compare DeepSeek’s new peak rates to the frontier competitors people actually use with OpenClaw:
- DeepSeek V4 Flash peak: $0.44 input / $1.32 output per 1M
- DeepSeek V4 Pro peak: $1.32 input / $3.96 output per 1M
- GPT-5.5 class models: around $5 input / $30 output per 1M
- Claude frontier models: comparable to or above that
Even at DeepSeek’s most expensive, most penalized rate, you are paying roughly 10 to 30 times less than the equivalent OpenAI or Anthropic call. V4 Pro output at peak is still about a quarter of what it cost a year ago. DeepSeek went from absurdly cheap to merely very cheap. Those are different problems, and the fix for the second one is not to panic and migrate your whole stack. If you are weighing whether a frontier model is worth the premium, our ChatGPT vs Claude vs Gemini comparison breaks down the trade-offs.
Why Some OpenClaw Bills Are Actually Exploding
If your DeepSeek bill jumped way more than the rate card explains, the culprit is not the new pricing. It is how OpenClaw uses the API. OpenClaw resends the full system prompt, your injected memory files, tool schemas, and conversation context with every single API call. On a chatty session, that adds up to hundreds of thousands of input tokens per day.
That matters now because of how the new pricing treats those repeated tokens:
- Cache hits are the cheap path. Repeated prefixes are supposed to be billed at the cache-hit rate. If your setup actually hits the cache, the repeated system prompt costs almost nothing.
- Cache misses are the expensive path. If your prompt changes between calls (memory file updated, session context shifted, tool list changed), you pay the full cache-miss input rate for the entire prompt every time. At peak, that is $0.44 per 1M for Flash.
- Long sessions compound it. The bigger your injected context, the more tokens you re-send, and the more each cache miss costs.
So the people reporting 3x or 4x bill jumps are usually not victims of the price hike. They are victims of cache misses on bloated contexts, running at peak hours, with retry loops on top. The fix is not switching models. The fix is fixing the usage pattern.
How to Keep Your DeepSeek Bill Flat
1. Compact what you inject on every call
Your memory files and workspace context get re-sent constantly. Keep MEMORY.md, SOUL.md, USER.md, and TOOLS.md tight. A 1,000-token file you trim to 300 tokens cuts every single cache-miss input bill by 70%. This is the highest-leverage fix you can make, and it was already the right move before August 16.
2. Make cache hits actually happen
DeepSeek caches stable prefixes. If you change your system prompt or memory on every turn, you defeat the cache. Keep the static parts of your prompt truly static, put volatile content at the end of the context rather than the front, and avoid rewriting injected files unless something actually changed.
3. Run heavy jobs off-peak
Peak rates roughly double the input and output price. Move your cron jobs, heartbeat checks, and batch work to off-peak hours. For a nightly reporting job, that is a free 50% discount on the same tokens. OpenClaw cron scheduling makes this a one-line change.
4. Route cheap models to background work
V4 Flash at peak is still far cheaper than Pro, and Pro is far cheaper than frontier models. Use Flash for summaries, classification, drafts, and anything where slightly sloppier output does not break the pipeline. Save Pro and frontier models for the hardest reasoning and coding. OpenClaw supports per-task model routing, so this does not require manual switching.
5. Cap retries and loops
A tool call that returns an empty result can trigger retry loops that burn thousands of tokens in minutes. Set a max retry count and a cooldown on your cron and agent tasks. One user reported cutting usage roughly 40% with a 30-second cooldown and a 3-retry cap, before any of this pricing drama.
One More Thing: Your Model ID May Have Died
While you are checking costs, verify your model reference. DeepSeek retired the old deepseek-chat and deepseek-reasoner model IDs on July 24, 2026. If your OpenClaw config still points at those, requests are failing or falling back. Update your config to deepseek/deepseek-v4-flash or deepseek/deepseek-v4-pro, then check OpenClaw’s cost estimates against the new published rates.
Bottom Line
DeepSeek raised prices, and the new peak and off-peak model changes how you should schedule work. But DeepSeek is still one of the cheapest capable models available, and the horror-story bills are mostly caused by cache misses, bloated context, and retry loops, not the rate card.
Compact your context, stabilize your prompt prefix, run batch jobs off-peak, route simple work to Flash, and cap your retries. Do those five things and your DeepSeek bill will barely move, no matter what the next rate change brings.
Running OpenClaw on a budget?
Token costs and model routing are only half the battle. The OpenClaw setup guide covers config hygiene that keeps every provider cheaper, including prompt caching and context management from day one.
About The Author
