Comparison · verified 2026-10-26
UncensoredAI vs DeepInfra
Low-cost inference on its own GPU fleet, with aggressive open-weight pricing.
- Our catalog
- 37
- models across 8 families
- DeepInfra
- 175
- model ids returned by their public /v1/openai/models endpoint
- Models both carry
- 9
- compared line by line below
- Cheapest here
- $0.04
- DeepSeek V4.1 Flash · per 1M input
Per-token rates on the same models
Both columns are the provider’s own published rate, normalised to USD per 1,000,000 tokens. Green means the lower price on that column; red means the higher one. Cached input is shown separately because a large repeated prefix can dominate a bill even when the headline input rate looks cheap.
| Model | Input · here | Input · DeepInfra | Cached · here | Cached · DeepInfra | Output · here | Output · DeepInfra |
|---|---|---|---|---|---|---|
| DeepSeek V4.1 Flash 1017deepseek-v4.1-flash-1017 | $0.04 | $0.20 | $0.01 | $0.006 | $0.04 | $0.60 |
| DeepSeek V4 Prodeepseek-v4-pro-0813 | $0.50 | $1.30 | $0.10 | $0.10 | $0.50 | $2.60 |
| DeepSeek V4 Flashdeepseek-v4-flash-0731 | $0.06 | $0.06 | $0.0053 | $0.015 | $0.06 | $0.18 |
| GLM-5.2glm-5.2 | $0.12 | $0.75 | $0.034 | $0.14 | $0.12 | $2.40 |
| GLM-5.3glm-5.3 | $0.12 | $0.90 | $0.034 | $0.20 | $0.12 | $4.00 |
| GLM-5.3 Flashglm-5.3-flash | $0.056 | $0.15 | $0.0056 | $0.03 | $0.056 | $0.50 |
| Kimi K3kimi-k3 | $0.30 | $2.85 | $0.06 | $0.285 | $0.30 | $14.25 |
| MiMo V2.6 Flashmimo-v2.6-flash | $0.06 | $0.14 | $0.004 | $0.0028 | $0.06 | $0.28 |
| MiMo V2.6 Promimo-v2.6-pro | $0.20 | $0.43 | $0.006 | $0.0036 | $0.20 | $0.87 |
Sources · read 2026-10-26
Competitor pricing changes without notice. These figures were read from the pages above on 2026-10-26; re-check before relying on them for a purchasing decision.
What that costs per month
Per-1M rates are hard to budget against, so here is the same DeepSeek V4.1 Flash 1017 priced across four modelled traffic shapes. Token counts are representative of common production patterns, not benchmarks; substitute your own numbers on the calculator.
| Workload | Here | DeepInfra | Difference |
|---|---|---|---|
| Support chatbot100k conversations/month · 500 prompt · 300 completion tokens | $3.40 | $28 | 8.3× cheaper here |
| Coding agent5k agent turns/month · 8k prompt · 1.5k completion tokens | $2.10 | $13 | 6.0× cheaper here |
| RAG answers20k answers/month · 20k prompt with a large cached prefix | $19 | $88 | 4.5× cheaper here |
| Batch classification1M calls/month · 200 prompt · 20 completion tokens | $8.80 | $52 | 5.9× cheaper here |
Commercial terms
The per-token rate is only part of the bill. Credit-purchase fees, minimums and funding methods change the effective cost, so they are listed here rather than buried in a pricing footnote.
- Pricing model
- Per-token list price
- No platform percentage on top. Some non-chat models are billed per second of GPU time instead.
- Infrastructure
- Own GPU fleet
- They run the hardware, so capacity is not dependent on another provider.
- Also offers
- GPU rental and fine-tuning
- Per-hour GPU instances sit alongside the token-priced inference API.
- Funding
- Card, invoicing for larger accounts
- Check their billing page for accepted methods in your region.
Where each one wins
A comparison with no downsides is an advertisement. These are the cases where DeepInfra is genuinely the better tool, next to the cases where this gateway is.
Where DeepInfra is the better choice
- Runs its own GPU fleet, so capacity and pricing are set directly.
- Very low rates on several open-weight models, with published cache-read pricing.
- GPU rental and fine-tuning on the same account, which matters if you also train.
- Transparent list pricing with no percentage fee layered on top.
Where UncensoredAI is the better choice
Closed frontier models on the same key
GPT, Gemini, Grok and the Claude family are not open-weight and cannot be served from a GPU fleet catalog. Here they sit beside the open models, so escalating from DeepSeek to GPT-6 Sol is a model-string change rather than a second integration.
Uncensored and abliterated builds
DeepInfra's catalog is standard open-weight releases; refusal-free weights are a separate distribution and are not part of it. If the workload needs an abliterated model, there is nothing on that platform to point at.
No email, account or KYC
A 16-digit key is the whole account. Nothing to verify, nothing to recover, and no personal data stored that could be breached.
One bill across eight families
Eight model families on a single prepaid balance, rather than one vendor's fleet plus a separate account for every closed model you also need.
The longer version
DeepInfra is the strongest structural competitor in this comparison, because it owns its hardware. That is not a marketing point: it means their economics come from utilisation, and it is why they can rent you the GPU by the hour as well as the tokens.
The trade-off is what a GPU fleet can serve. Open-weight models are distributable and can be hosted by anyone; closed frontier models cannot, so they are absent by definition. That single constraint explains most of the differences on this page — not pricing strategy, but what is physically hostable.
On the models both services carry, this platform is cheaper on output for all nine rows and on input for eight of nine, with DeepSeek V4 Flash equal at $0.06. Cached input is the exception: DeepInfra publishes lower cache-read rates on three rows, so a cache-heavy workload should read that column rather than the headline. The trade-off is capacity: a single model route here has no second fleet behind it, which DeepInfra's own hardware does.
For a team whose stack is entirely open-weight — a Qwen or Llama fine-tune, a DeepSeek coding model, an embedding model — DeepInfra is a natural fit and there is no reason to move. The case for this gateway starts when the workload needs something a GPU fleet cannot host: a frontier closed model for hard reasoning, or an abliterated build for red-teaming, or both behind one key so billing and SDK configuration do not multiply.
Latency is not comparable from published numbers. Both services publish rates, neither publishes a per-model latency SLA, and the honest answer is to measure with your own prompts. The status page here probes every model with a real one-token request and publishes rolling success rates and average latency, which is the closest either service comes to a public measurement.
Questions
- Is DeepInfra cheaper than UncensoredAI?
- It depends on the token mix. On input and output, this gateway is cheaper on almost every shared row; on cached input, DeepInfra is cheaper on three of the nine. The table above shows both rate cards side by side, and the workload table converts them into a monthly bill for four realistic traffic shapes, including a cache-heavy RAG pattern where the cached rate dominates.
- Does DeepInfra offer uncensored or abliterated models?
- Its catalog is standard open-weight releases. Abliterated weights are a separate distribution and are not listed there, so refusal-free builds are not available on that platform.
- Can DeepInfra serve GPT or Claude models?
- No. Those are closed weights and cannot be distributed to a third-party GPU fleet. If your application needs both open and closed models, that is two integrations unless you route them through one gateway.
- Do both services speak the same API?
- Yes, both expose a Chat Completions-compatible endpoint, so the request body is identical. Only the base URL and the model id change.
Verdict
Choose DeepInfra if your workload is entirely open-weight, you also want GPU rental or fine-tuning, and managing an account per vendor is acceptable. Choose this gateway if you need closed frontier models and refusal-free builds behind one key with no account to create.
Other comparisons
Try the cheaper rate card
Generate a key, point your existing SDK at the base URL, and send a request. There is no email, no card and no KYC step.