Tokens per second
Whole beta, all responses ≥ 40 tokens.
The slowest tenth still ran at 118 tok/s.
Reserved inference, booked by the hour. $0.20 per lane-hour.
Dedicated capacity for coding agents and long-context work: a 1M context window, images and tool calling on an OpenAI-compatible endpoint, available every hour of every day. Your requests never queue behind public API traffic, and your context stays cached between turns.
Your first 5 hours are on usApproved accounts start with free credits to test the lane. Limited capacity.
Beta report
Reserved lane · DeepSeek V4.1 Flash-1017 · free beta report
Numbers from the free beta that preceded this launch, measured at the API with real coding-agent traffic. A reserved lane is dedicated capacity: your requests never queue behind public API traffic, your context stays cached between turns, and the speed you measure on the first day is the speed you get on the thirtieth.
64,381
over 30 continuous hours
6.06 B
prompt tokens
49.9 M
completion tokens
97.2%
of prompt tokens, billed at the cache-hit rate
97.2% of prompt tokens were served from cache; only new tokens were processed, which keeps warm time to first token at 1.7 s median.
99.97%
22 failed requests, none from load
Traffic was real agentic coding work: long repository contexts, tool calls, compaction turns, and long code generations. A typical turn carried a 100K to 260K-token context (the window holds 1M) of which a few thousand tokens were new, and produced a few hundred tokens of code or a tool call.
Speed
Measured at the API
Every completed response of at least 40 tokens, timed from first token to last — everything between your client and the model included.
Whole beta, all responses ≥ 40 tokens.
The slowest tenth still ran at 118 tok/s.
Warm turns resume a cached context. Cold turns load one fresh.
Part of the warm figure is upload: a long conversation is sent as JSON on every turn, and a 1M-token one runs to several MB.
Per-request speed as more requests run on the lane at once. Controlled runs, 512-token answers, temperature 0.7.
tok/s per request
Code outruns prose because the speculative decoder predicts it more accurately. Real agent traffic sat between the two, at a 196 tok/s median.
Specification
Concurrency
Concurrency rowBooking
Booking row| Capability | Detail |
|---|---|
| Model | DeepSeek V4.1 Flash-1017, official weights. OpenAI-compatible chat completions with streaming; any OpenAI client or coding harness works by changing the base URL, key and model id. |
| Context window | 1,000,000 tokens. Set your harness to 1000000. If a turn would overflow, max_tokens is trimmed to fit; a prompt already at the limit returns a clear error so your agent can compact and continue. |
| Tool calling | Native function calling with parallel calls; tool calls arrive as standard tool_calls objects, streamed. |
| Images | Up to 8 images per request as data URLs or public URLs (PNG, JPEG). Screenshots, diagrams, UI captures. |
| Reasoning | Off by default for fast tool loops. Turn it on per request with reasoning_effort (low, high, xhigh, max, or a number 1–100); reasoning text arrives in the reasoning field. |
| Prompt caching | Automatic and free of charge. Repeated context is served from cache and reported in usage.prompt_tokens_details.cached_tokens. |
| Concurrency | Two simultaneous requests per lane, with a short queue behind them. Book more lanes in the same hour for more parallelism. Beyond the queue you get a 429 with Retry-After. |
| Booking | One-hour slots, any hour of the day, up to seven days ahead. Cancel free until an hour before the slot; join a live hour for the rest of it. Your API key is permanent and is admitted only during hours you hold. |
| Billing | Prepaid credits at $0.20 per lane-hour, bought in packs from $10 with bonus credits on larger packs. No subscription, no minimum. Disrupted hours are credited back automatically. |
| Data retention | Zero. Prompts, images and completions are processed in memory and never stored, logged or used for training. Only billing metadata (token counts, timestamps) is kept. Read the policy. |
Credits
Reserved-lane credit and Inference API credit are separate balances. The two run on different platforms, each with its own credit system, so topping up one does not fund the other. Check which account you are adding credit to before you pay.
Field notes
Testimonials
What people are saying who have already run the reserved lane.
The experience with Singularity has been really good! From API Inference (which I use daily now) to the Lane testing and now the paid lanes, everything top-notch. Speed is blazing fast, price is the minimum you can get, quality is just like the original one (GLM, DeepSeek, GPT and more) and the customer service.. 5 star!
The per hour lane system will soon explode in popularity and I'm glad to be part of it from the start!
ClassifiedI have good experience on Singularity API and doing an A/B test on my production application at losan.ai. I like the speed and the performance of model without sacrifice quality. I believe this is the good path because I don't have to choose between speed/quality/pricing. This is a good choice, hope you can maintain it ❤️
hiepxanhThe concepts in the beta were genuinely pretty good, and the results were impressive as well. The problems of token cost and time gates are very real, and the concept of a paid lane addresses those quite well. I had a few hiccups while testing, but it was consistent most of the time. Being able to test a model's capabilities at low or no cost before committing is another huge advantage. Definitely worth trying. 👍
XenoS1996Singularity API is the best provider I found till date for models like DeepSeek and GLM. There is no per-M-token pricing now — you use tokens on an hourly basis without worrying about API costs, usage limits, or the 5-hour waiting period every other provider has. It serves the direct model instead of quants, which gives the best performance in intelligence, and it's blazing fast. No worrying about your data being taken for training. I'd recommend anyone to use it.
yashSingularity API manages to provide a stable and reliable service for DeepSeek. During the beta tests, everything was very stable and completed my tasks successfully. The connection is fast and direct, and the pricing is also very affordable. Overall, I've had a great experience with it so far.
mintPricing
Beta · pricing
A lane costs $0.20 for an hour and carries two concurrent requests; book more lanes for more parallelism. Credits come in packs from $10, with bonus credits on the larger ones. Booked hours that are disrupted are credited back automatically.
Your first 5 hours are on us.
Every approved account starts with enough free credit for five lane-hours, so you can test the lane with your own harness before buying anything. Capacity is limited while we onboard in batches.
Running ten or more hours a day? Talk to us on Discord (opens in a new tab) about a committed plan.
This is a beta. Purchased credits never expire during it, and when the beta ends any unused credits are merged automatically into your account on the Singularity inference API.
Inference API
Singularity inference API
Sister product / singularityapi.dev / billed per request
A reserved lane is committed capacity for teams running constant traffic. If your usage is smaller or spiky, the inference API is the cheaper way in — the same platform, billed per request, with nothing to reserve and no commitment.
Separate credit balanceInference API
20+
One endpoint reaches all of them.
One
A single integration, not one per provider.
Semantic
Repeat work skips the round trip.
Edge
Steered to the fastest healthy region.