Comparison · verified 2026-10-26
UncensoredAI vs Together AI
Serverless and dedicated inference on open-weight models, plus fine-tuning.
- Our catalog
- 37
- models across 8 families
- Together AI
- 92
- token-priced rows in their published serverless pricing table
- Models both carry
- 6
- 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 · Together AI | Cached · here | Cached · Together AI | Output · here | Output · Together AI |
|---|---|---|---|---|---|---|
| DeepSeek V4.1 Flash 1017deepseek-v4.1-flash-1017 | $0.04 | $0.30 | $0.01 | $0.006 | $0.04 | $1.20 |
| DeepSeek V4 Prodeepseek-v4-pro-0813 | $0.50 | $1.32 | $0.10 | $0.13 | $0.50 | $3.96 |
| GLM-5.2glm-5.2 | $0.12 | $1.40 | $0.034 | $0.26 | $0.12 | $4.40 |
| GLM-5.3glm-5.3 | $0.12 | $1.40 | $0.034 | $0.26 | $0.12 | $4.40 |
| GLM-5.3 Flashglm-5.3-flash | $0.056 | $0.15 | $0.0056 | $0.03 | $0.056 | $0.50 |
| Kimi K3kimi-k3 | $0.30 | $3.00 | $0.06 | $0.30 | $0.30 | $15.00 |
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 | Together AI | Difference |
|---|---|---|---|
| Support chatbot100k conversations/month · 500 prompt · 300 completion tokens | $3.40 | $51 | 15.0× cheaper here |
| Coding agent5k agent turns/month · 8k prompt · 1.5k completion tokens | $2.10 | $21 | 10.1× cheaper here |
| RAG answers20k answers/month · 20k prompt with a large cached prefix | $19 | $134 | 6.9× cheaper here |
| Batch classification1M calls/month · 200 prompt · 20 completion tokens | $8.80 | $84 | 9.5× 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
- Serverless inference is token-priced; dedicated GPU clusters are billed per GPU-hour.
- Dedicated capacity
- Provisioned throughput
- Reserved capacity with SLAs, priced separately from serverless.
- Fine-tuning
- LoRA and full
- Billed per 1M training tokens with a per-model minimum charge.
- Batch inference
- Available
- Offline jobs are a first-class product rather than a client-side loop.
Where each one wins
A comparison with no downsides is an advertisement. These are the cases where Together AI is genuinely the better tool, next to the cases where this gateway is.
Where Together AI is the better choice
- Dedicated GPU clusters and provisioned throughput for load that cannot share capacity.
- LoRA and full fine-tuning on the same platform as inference, with published rates.
- Batch inference tooling for large offline jobs.
- A wide open-weight catalog across several vendors.
Where UncensoredAI is the better choice
Closed frontier models alongside open weights
GPT, Gemini, Grok and the Claude family are not open-weight and are not on a fine-tuning platform. Here they share an endpoint with the open models, so escalating from a cheap tier to a frontier model is one parameter.
Abliterated builds priced as model ids
Refusal-free weights are a separate distribution from the standard open releases. This catalog prices them on the same per-token basis, so reaching one needs no separate negotiation or deployment.
No email or identity to hand over
A random 16-digit key is the account. Nothing to verify and no profile to store, which matters when the workload itself is the reason you want an uncensored tier.
Lower per-token rates on the shared rows
On every model both services list, input and output rates here are materially lower, and there is no percentage fee on the balance you top up. Cached input is cheaper here on five of the six shared rows; the exception is DeepSeek V4.1 Flash 1017, where Together's cache rate is lower.
The longer version
Together AI is built for teams that want to own their inference stack: fine-tune a model, then serve it with reserved capacity. If that is the plan, the platform is a good fit and the per-token comparison is secondary to the training and provisioning tooling.
This gateway is built for the opposite shape of work: call a model you did not train, on hardware you do not manage, and change which model you call as the workload evolves. That is why the catalog spans eight families including closed frontier models, and why uncensored tiers are ordinary model ids rather than deployments you stand up yourself.
The price gap on the shared rows is wide, and the reason is structural rather than promotional. Together prices at a list rate with dedicated capacity behind it. This platform prices the same models lower — and correspondingly cannot give you a reserved GPU or a fine-tune. Read the cached column too: on five of the six shared rows the cache rate here is also lower, but not on all of them.
The practical question is whether you need training. If you do, Together is the answer and this page is not relevant to you. If you do not, and you want the same models at a lower per-token rate with uncensored builds available, the comparison favours this gateway on cost and simplicity.
Questions
- Is UncensoredAI cheaper than Together AI?
- On input and output, yes, on every shared model. The table above shows the same model ids at each provider's published rate, and Together does not add a percentage fee, so the difference is purely the per-token spread. Cached input is cheaper here on five of six rows; DeepSeek V4.1 Flash 1017 is the one where Together's cache rate wins.
- Does Together AI support fine-tuning?
- Yes, LoRA and full fine-tuning with published per-1M training rates and a per-model minimum charge. This gateway does not offer training at all, so if that is a requirement the decision is already made.
- Can I get dedicated throughput here?
- No. Requests share the model provider's capacity. If you need reserved GPU capacity or a throughput guarantee, a platform that sells dedicated instances is the right tool.
- Are the uncensored models the same as the standard ones?
- No. An abliterated model has the refusal direction removed from its weights, so the behaviour is baked in rather than applied by a system prompt. That is a different artifact from the standard release, not a configuration flag.
Verdict
Choose Together AI if you train or fine-tune models and want reserved capacity beside your inference. Choose this gateway if you consume models you did not train, want closed frontier models and uncensored builds behind one key, and would rather not manage per-vendor accounts.
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.