TokenMix Research Lab · 2026-04-13

DeepSeek API Free Credits 2026: 5M Grant Is Not Guaranteed
Last Updated: 2026-07-20
Author: TokenMix Research Lab
Data verified: 2026-07-20 against DeepSeek's official pricing, billing, model, error-code, token-usage, caching, and changelog pages
DeepSeek does not publicly guarantee every new API account a 5M-token grant in 2026. Check the Billing dashboard before planning free usage.
DeepSeek's official pricing page confirms that an account can have a granted balance and that this balance is deducted before topped-up funds. It does not state that every signup receives 5 million tokens, that every promotion lasts 30 days, or that a permanent API free tier exists. Those details may be account-, region-, or campaign-specific. The same page currently lists deepseek-v4-flash at $0.14 per million uncached input tokens and $0.28 per million output tokens, while deepseek-v4-pro costs $0.435 and $0.87 respectively. DeepSeek also says the legacy deepseek-chat and deepseek-reasoner aliases will be deprecated on July 24, 2026 at 15:59 UTC. See the official pricing table and official changelog.
This is a surgical July 2026 correction to our earlier guide. The 5M-token calculations remain useful only as a scenario for accounts whose dashboard actually shows that allowance.
Table of Contents
- Quick Verdict
- What DeepSeek Officially Confirms
- How to Check Your Actual Credit
- Current DeepSeek API Pricing
- How Far 5M Tokens Would Go
- Cost per Workload
- How to Make a Granted Balance Last
- What Happens When the Balance Runs Out
- Use Case Matrix
- Risks and Caveats
- Final Recommendation
- FAQ
- Sources
- Related Articles
Quick Verdict
The verified answer is narrower than the old headline: granted balances exist, but a universal 5M-token signup offer is not documented.
| Claim | Status | Source |
|---|---|---|
| DeepSeek accounts can contain a granted balance | Confirmed | DeepSeek pricing docs |
| Granted balance is consumed before topped-up balance | Confirmed | DeepSeek deduction rules |
| Every new account receives exactly 5M API tokens | Not confirmed | No universal 5M promise appears in current official docs |
| A 5M promotion may appear on some accounts | Likely | Account- or campaign-specific reports exist; verify in your dashboard |
| Every promotional balance expires after 30 days | Not confirmed | Current official pricing docs do not publish a universal expiry |
| DeepSeek offers a permanent public API free tier | False as a general claim | Official docs describe token billing and balances, not an ongoing free quota |
deepseek-v4-flash costs $0.14 input and $0.28 output per 1M uncached tokens |
Confirmed | Official model pricing |
deepseek-chat and deepseek-reasoner remain stable long-term aliases |
False after the stated cutoff | DeepSeek says both will be deprecated on 2026-07-24 15:59 UTC |
What DeepSeek Officially Confirms
DeepSeek confirms the balance mechanism and current token rates, not a universal signup quantity.
The deduction rules use two terms: topped-up balance and granted balance. If both exist, DeepSeek says the granted balance is used first. That is direct evidence that promotional or granted funds can exist. It is not evidence that every account receives the same amount.
| Item | Officially documented? | Practical meaning |
|---|---|---|
| Granted balance | Confirmed | Some accounts can receive non-top-up funds |
| Granted balance priority | Confirmed | It is spent before deposited funds |
| Universal 5M signup amount | Not confirmed | Do not budget around 5M until the dashboard shows it |
| Universal 30-day expiry | Not confirmed | Use the date shown in your own account, if any |
| Permanent API free tier | Not confirmed | The free chatbot and paid API are separate products |
| Prices can change | Confirmed | DeepSeek explicitly reserves the right to adjust rates |
This distinction matters because search results often collapse three different things into one phrase: free access to the DeepSeek chat product, a temporary account promotion, and a permanent API free tier. They are not interchangeable.
How to Check Your Actual Credit
Your account dashboard is the source of truth; if it shows no granted balance, assume the API is pay-as-you-go.
| Step | What to inspect | Decision |
|---|---|---|
| 1 | Sign in at platform.deepseek.com | Use your actual API account, not the public chat app |
| 2 | Open Billing or Balance | Look for a separate granted or promotional balance |
| 3 | Record the unit | Determine whether the dashboard displays currency or tokens |
| 4 | Record any expiry | Treat only the dashboard date or account terms as authoritative |
| 5 | Create an API key | Follow DeepSeek's first API call guide |
| 6 | Run one small request | Confirm the balance changes and inspect returned token usage |
Do not add a card solely because an article says a grant should appear. First confirm the account balance. If the dashboard shows zero, that is the operative result for your account even when another user reports receiving a promotion.
Current DeepSeek API Pricing
As of July 20, 2026, V4 Flash is the lower-cost route and V4 Pro charges roughly 3.1 times the uncached input and output rate.
| Model | Cache-hit input / 1M | Cache-miss input / 1M | Output / 1M | Context | Max output | Status |
|---|---|---|---|---|---|---|
deepseek-v4-flash |
$0.0028 | $0.14 | $0.28 | 1M | 384K | Confirmed |
deepseek-v4-pro |
$0.003625 | $0.435 | $0.87 | 1M | 384K | Confirmed |
DeepSeek publishes both prices on its Models & Pricing page. Context caching is enabled by default, and the API exposes prompt_cache_hit_tokens and prompt_cache_miss_tokens in the usage object. Cache matching is best effort, not guaranteed; read the official caching rules before projecting savings.
The old model names require immediate attention:
| Alias | Current compatibility mapping | Official cutoff | Action |
|---|---|---|---|
deepseek-chat |
Non-thinking mode of V4 Flash | 2026-07-24 15:59 UTC | Migrate to deepseek-v4-flash |
deepseek-reasoner |
Thinking mode of V4 Flash | 2026-07-24 15:59 UTC | Migrate and set thinking explicitly |
For a full rate comparison, use our updated DeepSeek API pricing guide and DeepSeek V4 review.
How Far 5M Tokens Would Go
If your dashboard explicitly grants 5,000,000 tokens, simple token division supports about 1,428 to 10,000 calls across the workloads below.
This section is scenario math, not a claim that DeepSeek gives every account 5M tokens. DeepSeek says the response usage fields are the authoritative measure because tokenization varies by model and language. See Token & Token Usage.
| Workload | Input per call | Output per call | Total per call | Calls from 5M tokens | Status |
|---|---|---|---|---|---|
| Short chat | 300 | 200 | 500 | 10,000 | Calculated scenario |
| Typical API request | 500 | 300 | 800 | 6,250 | Calculated scenario |
| Code generation | 500 | 400 | 900 | 5,555 | Calculated scenario |
| Document summary | 2,000 | 500 | 2,500 | 2,000 | Calculated scenario |
| Long-form generation | 200 | 1,000 | 1,200 | 4,166 | Calculated scenario |
| RAG request | 3,000 | 500 | 3,500 | 1,428 | Calculated scenario |
The formula is:
calls = granted_tokens / (average_input_tokens + average_output_tokens)
For 500 input and 300 output tokens, 5,000,000 / 800 = 6,250 calls. Production results will differ because conversation history, tools, retrieved context, and thinking output change token usage.
Cost per Workload
At current uncached rates, five million combined tokens cost roughly $0.80 to $1.28 on V4 Flash or $2.48 to $3.99 on V4 Pro in these four mixes.
| Workload mix totaling about 5M tokens | V4 Flash | V4 Pro | Calculation status |
|---|---|---|---|
| 6,250 calls at 500 in + 300 out | $0.96 | $2.99 | Confirmed rates, calculated workload |
| 2,000 calls at 2,000 in + 500 out | $0.84 | $2.61 | Confirmed rates, calculated workload |
| 4,166 calls at 200 in + 1,000 out | $1.28 | $3.99 | Confirmed rates, calculated workload |
| 1,428 calls at 3,000 in + 500 out | $0.80 | $2.48 | Confirmed rates, calculated workload |
The cost formula separates input and output:
cost = (input_tokens / 1,000,000 x input_rate)
+ (output_tokens / 1,000,000 x output_rate)
These totals assume every input token is a cache miss. Repeated prompt prefixes can lower input cost sharply, but the exact hit rate must come from your own usage data. For broader budgeting, compare the same workload in our AI API pricing calculator guide.
How to Make a Granted Balance Last
The reliable levers are shorter repeated context, explicit output limits, cache-friendly prefixes, and daily usage measurement.
| Lever | Verified mechanism | Expected effect | Confidence |
|---|---|---|---|
| Set an output cap | Fewer generated tokens are billed | Lower output spend | Confirmed mechanism; saving varies |
| Keep repeated prefixes identical | Matching prefixes can hit disk cache | Lower input rate on cache hits | Confirmed mechanism; hit rate varies |
| Use V4 Flash by default | Flash has lower published rates | Lower cost per token | Confirmed |
| Disable thinking when unnecessary | Thinking produces additional reasoning content | Usually fewer generated tokens | Likely; workload-dependent |
Track returned usage |
API returns measured token consumption | Prevents estimate drift | Confirmed |
| Stop failed loops | Retries and agent loops can multiply calls | Avoids repeated spend | Confirmed arithmetic; incidence varies |
Three practical rules do most of the work:
- Put stable instructions and documents at the start of prompts so repeated prefixes have a chance to match the cache.
- Set the smallest output limit that still completes the task, then raise it only when truncation is observed.
- Log prompt tokens, cache-hit tokens, cache-miss tokens, completion tokens, model name, and HTTP status for every request.
We do not assign a universal savings percentage. Cache hit rate and output reduction depend on the application, and a percentage without first-party telemetry would be speculation.
What Happens When the Balance Runs Out
DeepSeek documents HTTP 402 Insufficient Balance when the account has no remaining balance.
| HTTP code | Official meaning | Recommended action |
|---|---|---|
| 400 | Invalid request format | Fix the request body |
| 401 | Authentication failure | Check or replace the API key |
| 402 | Insufficient balance | Check balance, then top up if required |
| 422 | Invalid parameters | Correct the parameters named in the response |
| 429 | Rate limit reached | Pace requests and add backoff |
| 500 | Server error | Retry after a brief wait |
| 503 | Server overloaded | Retry after a brief wait |
These definitions come from DeepSeek's official error-code table. A 402 is different from a 429: adding funds will not solve a concurrency problem, and retrying indefinitely will not solve a zero balance.
Use Case Matrix
Use V4 Flash for cost-sensitive general work, and test V4 Pro only where measured quality gains justify the higher rate.
| Use case | Starting model | Credit strategy | Why |
|---|---|---|---|
| API learning | V4 Flash | Small capped calls | Lowest published token rate |
| RAG prototype | V4 Flash | Stable document prefixes | Better chance of cache reuse |
| Agent experiment | V4 Flash, non-thinking first | Hard call and token budgets | Loops can consume balance quickly |
| Difficult reasoning | A/B test Flash vs Pro | Enable thinking only for hard cases | Quality benefit must be measured |
| Production migration | Explicit V4 model IDs | Do not rely on legacy aliases | Alias cutoff is documented |
| Multi-provider application | V4 Flash plus fallback | Record cost and error status | Avoid a single balance or endpoint dependency |
If you are new to the API, our DeepSeek API tutorial covers the current V4 model IDs. The Python setup guide shows the OpenAI-compatible client pattern.
Risks and Caveats
The largest risk is treating a promotional report as a contractual free tier.
| Risk | Status | Consequence | Mitigation |
|---|---|---|---|
| No grant appears on a new account | Confirmed possibility | Requests require paid balance | Check Billing before development |
| Grant amount differs by account or campaign | Likely | Token plan becomes inaccurate | Use the displayed amount only |
| Promotion expiry differs | Likely | Unused balance may disappear | Record dashboard terms and date |
| Prices change | Confirmed possibility | Cost projections become stale | Recheck official pricing before purchase |
| Legacy aliases stop working | Confirmed scheduled change | Requests can fail after cutoff | Move to explicit V4 model IDs |
| Cache savings are lower than expected | Confirmed possibility | Input spend exceeds estimate | Measure cache-hit fields |
| A permanent free API tier appears later | Speculation | This article would need another update | Require official documentation before claiming it |
Final Recommendation
Do not sign up expecting a guaranteed 5M-token DeepSeek API grant. Verify the granted balance and expiry in your own dashboard, then calculate calls from measured usage. If no grant appears, budget with current V4 prices and migrate off the retiring aliases before July 24.
FAQ
Does DeepSeek give every new API account 5 million free tokens?
No confirmed universal offer exists in the current official documentation. DeepSeek confirms that granted balances can exist, but your dashboard must confirm the amount for your account.
Does DeepSeek have a permanent free API tier?
No permanent public API quota is documented. The free DeepSeek chat product should not be treated as proof that API calls are free.
Where can I check DeepSeek free credits?
Check the Billing or Balance area of the DeepSeek API platform. Record the amount, its unit, and any expiry shown there before running a workload.
How many calls would 5M tokens support?
At 500 input and 300 output tokens per call, 5M tokens support 6,250 calls. This is arithmetic for a 5M-token allowance, not proof that your account has one.
What does 5M tokens cost at current DeepSeek rates?
It depends on the input/output mix and model. Across four example mixes in this article, the uncached cost is about $0.80-$1.28 on V4 Flash and $2.48-$3.99 on V4 Pro.
Do DeepSeek promotional credits expire after 30 days?
No universal 30-day rule is confirmed by the current pricing page. Use the expiry displayed in your own dashboard or account terms.
What happens when my DeepSeek balance reaches zero?
The official error table identifies HTTP 402 Insufficient Balance. Check the account balance and top up only if you choose to continue paid usage.
Can I keep using deepseek-chat and deepseek-reasoner?
Not as a long-term plan. DeepSeek says both compatibility aliases will be deprecated on July 24, 2026 at 15:59 UTC, so migrate to explicit V4 model IDs.
Sources
- DeepSeek Models & Pricing - Official prices, model limits, deduction rules, and alias cutoff
- DeepSeek Your First API Call - Official base URLs, model IDs, and SDK examples
- DeepSeek Token & Token Usage - Official token accounting guidance
- DeepSeek Error Codes - Official
402,429, and server error definitions - DeepSeek Rate Limit & Isolation - Official concurrency limits and
user_idbehavior - DeepSeek Context Caching - Official cache matching and usage fields
- DeepSeek Thinking Mode - Official thinking controls and response behavior
- DeepSeek Change Log - Official V4 launch and legacy alias migration notice
Related Articles
- DeepSeek API Pricing 2026
- DeepSeek V4 Review 2026
- DeepSeek API Tutorial 2026
- DeepSeek API Python Tutorial
- 15 Best Free LLM APIs 2026
About TokenMix
TokenMix.ai is an AI API relay for comparing and routing models through one OpenAI-compatible endpoint. Browse current model pricing, inspect the model catalog, or read the API documentation.