Headroom · for DeepSeek people
Your DeepSeek API balance,
one click from the clock.
No tiers, no windows, no reset day — a DeepSeek key draws on a prepaid wallet that returns 402 the moment it's empty. Headroom polls the balance endpoint and keeps what's left, per currency, in the taskbar flyout.
Windows 10 & 11 · signed auto-updates · no account needed
How your DeepSeek API balance actually works
One wallet holds two kinds of money, and only one of them is yours.
DeepSeek sells the API one way: prepaid. An account holds money, requests are billed at tokens × price, and your DeepSeek API balance is the entire limit system — nothing resets because there's no window to reset. When it hits zero, every billable request returns HTTP 402 Insufficient Balance until you top up. The key never stops being valid; 402 is an empty account, not a broken credential.
Checking it costs one GET: api.deepseek.com/user/balance with your ordinary API key as Bearer auth — no admin key, no separate management token. The response splits the wallet into granted and topped-up money per currency, and the split has teeth: DeepSeek spends granted money first, and grants expire.
The prepaid balance
The only limit DeepSeek enforces on your spend. Nothing resets — the number falls as tokens are billed and rises only when you top up on platform.deepseek.com. At zero, every billable request returns HTTP 402; after a top-up, requests resume on the same key with no re-auth. Every platform account with an API key works this way; there are no tiers to compare.
Granted money and topped-up money
Fees come out of granted, promotional balance first, then out of money you paid in. DeepSeek's FAQ says topped-up balance never expires and that unused paid balance is self-service refundable from the Billing page. Grants are different: per DeepSeek's FAQ they carry expiry dates shown on the Billing page, and the balance API counts only the unexpired portion toward your total.
Per-token pricing, cache-aware
deepseek-v4-flash: $0.14 per million input tokens on a cache miss, $0.0028 on a hit, $0.28 per million out. deepseek-v4-pro: $0.435, $0.003625 and $0.87 for the same three. A cache hit costs 50–120× less than a miss, so token counts alone can't predict spend — one more reason to read the balance instead of estimating it. Checked against the official pricing page on July 27, 2026; DeepSeek says prices can change at its discretion.
Concurrency caps, not RPM tiers
DeepSeek publishes no requests-per-minute or tokens-per-minute quotas. The limits it does publish are per-model concurrency: 500 simultaneous requests on deepseek-v4-pro, 2,500 on deepseek-v4-flash. Exceed one and you get HTTP 429 — a different failure than 402, and worth keeping separate: 429 means slow down, 402 means pay up. The caps are defaults, and capacity expansion is free on request.
The slowdown under load
When the platform is saturated, DeepSeek stretches responses rather than rejecting them. Non-streaming requests receive continuous empty lines while they wait; streaming requests get ': keep-alive' SSE comments. If inference hasn't started after 10 minutes the connection closes. A request that looks hung may simply be queued behind everyone else's.
What Headroom shows for DeepSeek
A wallet needs a different gauge than a clock — Headroom shows DeepSeek as money, and says so plainly when it can't read it.
Money left, per currency
Headroom polls the balance endpoint with your key and shows one row per currency — 12.50 remaining, formatted the way DeepSeek reports it. USD and CNY stay separate; converting them would mean inventing an exchange rate. No percent, either: a wallet has no denominator, and a bar without one would be theater.
A “!” beats a stale number
If DeepSeek rejects the key, the taskbar shows a “!” and the flyout says what to do: remove and reconnect the account. A rate-limited balance check, an empty response and a network failure each get their own message, and a reading Headroom couldn't refresh turns hatched instead of posing as fresh.
Clocks for your other plans
DeepSeek is the only account where the wallet is the whole story, so it gets the one number that exists. Subscription plans report quota windows, and those get fuller meters — percent left, pace tick, reset countdown — on Claude and Codex, with Gemini, Grok and Cursor on the same pattern.
Asked right after a 402
How do I check my DeepSeek API balance?
Two official readouts: the Billing page on platform.deepseek.com, or GET https://api.deepseek.com/user/balance with your normal API key as Bearer auth — it returns total, granted and topped-up balance per currency. Headroom polls that same endpoint and keeps the total one click from the Windows taskbar clock.
Why am I getting 402 Insufficient Balance from the DeepSeek API?
Your prepaid balance reached zero. DeepSeek deducts granted balance first, then topped-up balance; once both are gone, every billable request returns 402 until you top up on platform.deepseek.com. The key itself is fine, and the free chat at deepseek.com keeps working — it's billed separately, so it proves nothing about your API account.
Does the DeepSeek API have rate limits?
There are no published RPM or TPM tiers. The documented limits are per-model concurrency caps — 500 simultaneous requests for deepseek-v4-pro, 2,500 for deepseek-v4-flash — with HTTP 429 past them and free capacity expansion on request. Under heavy load DeepSeek slows instead of rejecting: non-streaming calls receive empty lines, streams receive ': keep-alive' comments, and the connection closes if inference hasn't started within 10 minutes.
Does DeepSeek API balance expire?
Paid balance doesn't: DeepSeek's FAQ states topped-up balance never expires, and unused paid balance is self-service refundable from the Billing page. Granted promotional balance does carry expiry dates — the balance API counts only the unexpired portion — and since grants are spent before your own money, an expiring grant can lower your total without any usage.
How much does the DeepSeek API cost in 2026?
Per-token and prepaid, with no subscription tier. deepseek-v4-flash: $0.14 per million input tokens on a cache miss, $0.0028 on a cache hit, $0.28 per million output. deepseek-v4-pro: $0.435, $0.003625 and $0.87 for the same three. Both take 1M tokens of context and emit up to 384K. The 2025 off-peak discount is gone — the current pricing page lists no time-based discounts — and these numbers were checked against it on July 27, 2026.
Why does deepseek-chat return an error now?
The legacy model names deepseek-chat and deepseek-reasoner stopped resolving on July 24, 2026, three months after DeepSeek's changelog announced their retirement. Point your code at deepseek-v4-flash or deepseek-v4-pro; reasoning is now a thinking-mode toggle, not a separate model name. It's a naming error, not a billing one — your balance is untouched.
For DeepSeek the wallet is the whole limit. The clocks have pages of their own: