# Together AI startup deal: $1,000 in credits

> Run and fine-tune open-source LLMs. Llama, Mistral, and custom models on scalable GPU infrastructure.

- **Offer:** $1,000 in credits
- **Save up to:** $1,000
- **Category:** AI & Data
- **How to get it:** Claim details are unlocked with SaaSOffers Premium. https://saasoffers.tech/offers/together-ai
- **Eligibility:** AI startup using open-source models

## What you get

- $1,000 in Together AI credits
- Inference for 50+ open-source models
- Fine-tuning with custom data
- Scalable GPU clusters
- OpenAI-compatible API

## Our Together AI listing: $1,000 in credits

**Offer:** $1,000 in credits. Our listing's requirement is "AI startup using open-source models". There is no promo code. Step 1 is to go to Together AI's startup accelerator page and apply there.

Together AI's own page shows credits that run well above that figure, so read the program section below before you decide.

We checked these plans and prices against Together AI's official pricing and help pages in November 2026.

![Together AI homepage, $1,000 in credits for startups](https://zlcbkfmbwwroglopnrhf.supabase.co/storage/v1/object/public/blog-images/deals/together-ai/tool1.png)

Together AI hosts open-weight models behind an API. You send a request, it runs on Together's GPUs, and you pay per token. The same account also rents dedicated GPUs, sells GPU clusters, runs fine-tuning jobs and serves image, video and speech models. Its docs call the API compatible with the OpenAI REST API and SDKs, with limits we list further down.

## The startup accelerator has three tiers, and the top one is $50K

Our listing says $1,000. Together AI's accelerator page describes a selection-based program with three tiers, set by how much funding you have raised:

| Tier | Funding raised | Platform credits | Forward-deployed engineering |
|---|---|---|---|
| Build | Up to $5M | Up to $15K | 3 hours |
| Scale | $5M to $10M | Up to $30K | 6 hours |
| Grow | Over $10M | Up to $50K | 10 hours |

The page says credit amounts are "up to" the listed values, so the real grant is decided by the Together AI team. It also says credits do not apply to Reserved GPU Clusters. They do cover serverless inference, dedicated endpoints, fine-tuning and instant clusters.

Beyond credits, the program lists forward-deployed engineering hours, content on AI engineering practices, joint go-to-market support aimed at 800K+ Together AI developers, community events and access to a VC network. The page gives no application deadline and no eligibility rules beyond funding raised.



## What a request actually costs

Serverless pricing is per million tokens, with separate input, cached input and output rates. These are the rows we checked on the pricing page:

| Model | Input | Cached input | Output |
|---|---|---|---|
| Ternary Bonsai 27B | $0.00 | none listed | $0.00 |
| Rnj-1 Instruct | $0.15 | none listed | $0.15 |
| gpt-oss-120B | $0.15 | none listed | $0.60 |
| GLM-5.3-Flash | $0.15 | $0.03 | $0.50 |
| Qwen3.5 9B | $0.17 | none listed | $0.25 |
| DeepSeek V4.1 Flash | $0.30 | $0.006 | $1.20 |
| MiniMax M3 | $0.30 | $0.06 | $1.20 |
| Qwen3.6-Plus | $0.50 | none listed | $3.00 |
| GLM-5.3 | $1.40 | $0.26 | $4.40 |
| Kimi K3.1 Pro | $2.00 | $0.20 | $10.00 |
| Qwen3.7-Max | $2.50 | $0.50 | $7.50 |
| Kimi K3 | $3.00 | $0.30 | $15.00 |

The spread from the cheapest paid row to the top row is large. Kimi K3 output costs 25 times what gpt-oss-120B output costs ($15.00 against $0.60).

Together's model pages list the same rates as standard serverless prices, for example gpt-oss-120B at $0.15 in and $0.60 out. The pricing page states no batch discount percentage.

### Worked example: gpt-oss-120B at 2,000 tokens in, 500 out

One request with 2,000 input tokens and 500 output tokens on gpt-oss-120B:

- Input: 2,000 x $0.15 / 1,000,000 = $0.0003
- Output: 500 x $0.60 / 1,000,000 = $0.0003
- Total: $0.0006 per request

If our $1,000 listing term applies, that is about 1.67 million requests (1,000 / 0.0006). The same request on Kimi K3 costs 2,000 x $3.00 / 1M = $0.006 plus 500 x $15.00 / 1M = $0.0075, so $0.0135. That is 22.5 times the gpt-oss-120B figure, and $1,000 covers about 74,000 requests.

Cached input matters for chat products that resend a long system prompt. On DeepSeek V4.1 Flash the cached rate is $0.006 per million tokens against $0.30 uncached, a 50x gap, so a 10,000-token prompt that hits the cache costs $0.00006 instead of $0.003.

![Together AI plans and features before the $1,000 in credits deal](https://zlcbkfmbwwroglopnrhf.supabase.co/storage/v1/object/public/blog-images/deals/together-ai/tool2.png)

## Image, video and speech are priced per unit, not per token

Together AI also serves non-text models, and the billing units change by modality.

**Images, per image:** FLUX.2 [dev] $0.0154, FLUX.2 [pro] $0.03, Qwen Image 2.0 $0.04, Gemini 3.1 Flash Image (Nano Banana 2) $0.05, GPT Image 2 $0.053 and Nano Banana Pro (Gemini 3 Pro Image) $0.134.

**Video, per video:** ByteDance Seedance 2.5 $0.115, HappyHorse 1.1 $0.14, ByteDance Seedance 2.0 $0.16, Vidu Q1 $0.22, PixVerse v5 $0.30 and MiniMax Hailuo 02 $0.49.

**Transcription, per audio minute:** Whisper Large v3 $0.0015, and the streaming version $0.0035.

**Text to speech, per million characters:** Kokoro-82M $4.00 and Cartesia Sonic-3 $65.00.

The pricing page showed "No matching models" for embeddings, rerank and moderation when we checked, so there are no prices to quote for those. Look at the models list in your account if you need them.

## Fine-tuning is priced by training tokens, with a minimum per job

Fine-tuning has its own rate card, per million tokens, for supervised fine-tuning (SFT) and direct preference optimization (DPO). Each model also carries a minimum charge per job.

| Model | SFT | DPO | Minimum |
|---|---|---|---|
| Llama 3.1 8B | $0.34 | $0.84 | $4.00 |
| Qwen3.5 9B | $0.34 | $0.84 | $4.00 |
| gpt-oss-20B | $0.40 | $1.00 | $4.00 |
| Qwen3.5 27B | $1.05 | $2.62 | $4.00 |
| Llama 3.3 70B | $2.03 | $5.08 | $4.00 |
| Llama 4 Scout | $2.10 | $5.25 | $6.00 |
| gpt-oss-120B | $2.50 | $6.25 | $6.00 |
| Llama 4 Maverick | $5.60 | $14.00 | $16.00 |
| DeepSeek-V3.1 | $7.00 | $17.50 | $20.00 |
| Kimi K2.6 | $15.00 | $37.50 | $60.00 |
| GLM-5.2 | $40.00 | $100.00 | $60.00 |

The page shows a second set of rates for a smaller list of models, for example Llama 3.3 70B at $2.24 SFT and Llama 3.1 8B at $0.38. The page has LoRA and full fine-tuning tabs but the table text does not say which rates belong to which tab, so check the job estimate in the console before you launch.

The bill is training tokens (dataset size times epochs) plus evaluation tokens (validation set size times number of evaluations).

### Worked example: a 5M-token dataset, 3 epochs

That is 15 million training tokens, before evaluation tokens.

- Llama 3.1 8B SFT at $0.34: 15 x $0.34 = $5.10
- Llama 3.3 70B SFT at $2.03: 15 x $2.03 = $30.45
- Llama 4 Maverick SFT at $5.60: 15 x $5.60 = $84.00
- GLM-5.2 SFT at $40.00: 15 x $40.00 = $600.00

Even the last line fits inside $1,000, though a few iterations on it would not. Serving the tuned model is billed separately, so budget for inference after training.

## Dedicated GPUs and clusters

When traffic is steady enough to keep a GPU busy, Together rents the hardware by the hour. The pricing page splits this into two products.

**Dedicated inference, per GPU per hour:** NVIDIA HGX H100 is $5.49 on demand and HGX B200 is $8.99. H200, B300, GB200 NVL72 and GB300 NVL72 say "contact us". Reserved pricing for all of them is "contact sales".

**GPU clusters, per GPU per hour:**

| Hardware | Preemptible | On-demand | Reserved 3 to 30 days | Reserved 31 to 90 days | Reserved 91 to 180 days |
|---|---|---|---|---|---|
| HGX H100 | $1.99 | $3.99 | $3.69 | $3.45 | $3.19 |
| HGX H200 | $2.99 | $5.99 | $4.99 | $4.15 | $3.99 |
| HGX B200 | $4.09 | $8.19 | $7.99 | $7.79 | $6.79 |
| HGX B300 | $4.99 | $9.99 | contact us | contact us | contact us |

For 181 days and longer the page says "Contact us" for every model.

The same H100 therefore costs $5.49 an hour as a dedicated endpoint and $3.99 as an on-demand cluster GPU. The cluster hour is the cheaper one, but you run the serving stack yourself.

### Worked example: how long $1,000 lasts on one H100

- Dedicated endpoint: $1,000 / $5.49 = about 182 hours, or 7.6 days of continuous running
- On-demand cluster: $1,000 / $3.99 = about 250 hours, or 10.4 days
- Preemptible cluster: $1,000 / $1.99 = about 502 hours, or 20.9 days

Remember the accelerator terms: credits do not apply to Reserved GPU Clusters. If our $1,000 listing term applies to your account, confirm which products it covers.

The page also lists post-training on an NVIDIA HGX B200 at $11.00 per GPU hour, reduced to $9.00 under a promotion valid until December 14, 2026.

## How OpenAI compatibility works, and where it stops

Together's docs say moving an OpenAI app takes two changes: your API key and the base URL, `https://api.together.ai/v1`. There is also a US-only base URL, `https://us-api.together.ai/v1`, which supports select models at 1.25 times standard serverless rates.

The compatibility covers chat, completions, vision, image generation, text to speech and embeddings. The docs list these gaps:

- Model IDs are namespaced, for example `openai/gpt-oss-120b`. OpenAI names like `gpt-4o` return a 404.
- Assistants, Threads and Runs are not supported. Neither are OpenAI-shaped fine-tuning and Batch APIs, and `moderations.create` returns a 404.
- `service_tier`, `store`, `metadata` and `prediction` are accepted but ignored. `logit_bias` is mostly unsupported and `seed` is best effort.
- Error objects follow OpenAI's shape but use Together's `type` and `code` values, so match on HTTP status instead.

If your code only calls chat completions, the swap is quick. If you use Assistants or the OpenAI fine-tuning endpoint, plan for a rewrite of that part.

## Rate limits

Together's rate limit page publishes no tier table, no requests-per-minute numbers and no free-credit amount. It says limits are not usually an issue for serverless inference. It warns that you may get `429 Too Many Requests` or `503 Service Unavailable` during high demand, and that Together may apply custom limits to models in high demand. Committed throughput goes through provisioned throughput and sales.

So you cannot size capacity from public numbers. Load-test the model you plan to use before launch day.

## Together AI next to Groq and OpenAI

The figures above use numbers we checked on each vendor's own pages in November 2026.

| | Together AI | [Groq](/offers/groq) | [OpenAI](/offers/openai) |
|---|---|---|---|
| gpt-oss-120B | $0.15 in, $0.60 out | $0.15 in, $0.60 out | not checked |
| Small proprietary model | none checked | none checked | gpt-6-luna $0.10 in, $0.50 out |
| Transcription | Whisper Large v3 $0.0015 per minute | Whisper Large v3 $0.111 per hour | gpt-4o-mini-transcribe $0.003 per minute |
| Fine-tuning | SFT and DPO on open models | No fine-tuning service, LoRA inference for enterprise | Closed to new users, being wound down |

Per request, Together and Groq charge the same for gpt-oss-120B on the pages we read, so the choice there is not price. It comes down to the catalog, the speed you measure and the extras. Together has image, video, fine-tuning and GPU rental in one account. Groq's gpt-oss-120B page lists 500 tokens per second.

Whisper Large v3 at Groq is $0.111 an hour, which is $0.00185 a minute. Together's $0.0015 a minute is lower by that arithmetic, and both are a small fraction of a cent.

## Who it fits

- Teams that want open models (Llama, Qwen, DeepSeek, gpt-oss, Kimi, GLM) through one API without running GPUs.
- Founders who may fine-tune a mid-size model on their own data and want training and serving in the same account.
- AI startups raising or about to raise, because the accelerator ties credits and engineering hours to funding stage.
- Products with mixed media, where text, image, video and transcription calls share one bill.

## Reasons to pass on Together AI
- Anyone whose code depends on OpenAI Assistants, Threads or the OpenAI fine-tuning API. Those are not supported here.
- Teams that need published rate limits to plan capacity. The docs give no numbers.
- Startups that need a fixed monthly price. Everything here is metered.

## Getting set up on Together AI
1. Start with one cheap model, such as gpt-oss-120B or GLM-5.3-Flash, and measure quality on your real prompts before you try Kimi K3 ($3.00 in, $15.00 out).
2. Put your long system prompt first so the cached input rate applies on the models that list one.
3. Log token counts per request from day one. Output tokens cost 4 times input on gpt-oss-120B, so a long answer moves the bill more than a long question.
4. Keep the model ID in a config value. Namespaced IDs change as models are replaced.
5. Estimate a fine-tuning job in the console before launching. Add evaluation tokens to the training tokens.
6. Do not buy a dedicated GPU until your steady load keeps it busy. At $5.49 an hour, an idle H100 costs $131.76 a day.

## How to get the Together AI offer
1. Go to Together AI's startup accelerator page at together.ai/startup-accelerator.
2. Fill in the application form with your company details and funding raised.
3. Wait for the Together AI team to review and respond. The page gives no timeline.
4. Once accepted, check your account for the credit balance and ask what it covers. The accelerator page says credits do not apply to Reserved GPU Clusters.
5. Create an API key, set the base URL to `https://api.together.ai/v1` and run your first benchmark.

Our listing's term is $1,000 in credits for an AI startup using open-source models. We have no way to confirm from Together AI's pages that this exact amount is what you will receive, so treat the figure as our listing's term and the tier table above as the vendor's own.

## FAQ

### How much credit does Together AI's startup accelerator give?

Together AI's accelerator page lists three tiers by funding raised: Build (up to $5M raised) gets up to $15K in credits and 3 forward-deployed engineering hours, Scale ($5M to $10M) up to $30K and 6 hours, Grow (over $10M) up to $50K and 10 hours. Amounts are 'up to'. Our listing's term is $1,000, so confirm the grant you receive.

### What do the startup credits not cover?

The accelerator page says credits cover serverless inference, dedicated endpoints, fine-tuning and instant clusters, and do not apply to Reserved GPU Clusters.

### What does a typical request cost on gpt-oss-120B?

Together AI lists gpt-oss-120B at $0.15 per million input tokens and $0.60 per million output tokens. A request with 2,000 input and 500 output tokens costs $0.0006, so $1,000 would cover about 1.67 million such requests if our listing's credit applies.

### Is the Together AI API compatible with OpenAI's?

Partly. Change the API key and the base URL to https://api.together.ai/v1 and chat, completions, vision, image generation, text to speech and embeddings work with the OpenAI SDK. Model IDs are namespaced (for example openai/gpt-oss-120b), and Assistants, Threads, Runs, OpenAI-shaped fine-tuning and Batch, and moderations are not supported.

### How is fine-tuning billed?

Per million training tokens, which is dataset size times epochs, plus evaluation tokens. Each model has a minimum charge per job. Examples from the pricing page: Llama 3.1 8B SFT $0.34 per million tokens, Llama 3.3 70B SFT $2.03, with a $4.00 minimum on both. A 15-million-token job on Llama 3.3 70B comes to $30.45 before evaluation tokens.

### How much does an H100 cost on Together AI?

A dedicated inference H100 is $5.49 per GPU hour on demand. In GPU clusters an HGX H100 is $1.99 preemptible, $3.99 on demand, and $3.19 to $3.69 reserved depending on length (3 to 180 days). Longer reservations say contact us.

---
Source: SaaSOffers, https://saasoffers.tech/offers/together-ai
