GILD verdict
Pick Together AI when you want one host that runs the model and sells dedicated GPUs, and an admin will turn prompt storage off.
Pick OpenRouter when you want to route an open model across hosts and pay the lowest listed rate, and test what each host serves. At the 1,000,000-request gpt-oss-120b workload above, Together costs US$600, OpenRouter pinned to Together's endpoint about US$633, and OpenRouter's lowest listed host about US$61.72 on an fp4 endpoint.
Together AI hosts open models itself. OpenRouter routes each request to one of many hosts, Together among them, and bills the host's list price from prepaid credits with a fee on each credit purchase. For gpt-oss-120b, Together lists US$0.15 per 1M input tokens and US$0.60 per 1M output tokens, OpenRouter lists the same rates for Together's endpoint, and OpenRouter's model page starts at US$0.015 and US$0.057 on an fp4 endpoint. Together keeps prompts by default until an admin turns storage off, while OpenRouter keeps none itself by default but may route to hosts that store data unless you deny them.
Choosing between Together AI and OpenRouter
Choose Together AI if
- You want one host that runs the model itself, with dedicated endpoints billed by the minute when traffic is steady.
- Your contract needs fixed rate limits, which Together keeps in place for the term of an Enterprise or Scale contract.
- An admin can turn prompt storage off before production traffic.
Choose OpenRouter if
- You want one API key for every model and provider you approve, with backup providers allowed by default.
- You want Zero Data Retention routing, which is available on every plan.
- You will set a provider order or a list of providers to allow where host quality matters.
Side by side
| Dimension | Together AI | OpenRouter |
|---|---|---|
| gpt-oss-120b list price and billing unitList prices in US dollars on Nov 1, 2026. OpenRouter's endpoints API prints prices per token: 0.00000015 and 0.0000006 are US$0.15 and US$0.60 per 1M. | US$0.15 per 1M input tokens and US$0.60 per 1M output tokens on serverless, billed per input and output token12 | Billed per token at the serving host's list price, with no markup. The model page lists gpt-oss-120b from US$0.015 per 1M input and US$0.057 per 1M output tokens, on an fp4 endpoint. Together's endpoint on OpenRouter lists US$0.15 and US$0.60. A fee of 5.5% on Standard or 8% on Business is charged when you buy credits789 |
| Monthly cost for one gpt-oss-120b workload (GILD's calculation from published rates)GILD's calculation in US dollars from list prices stored on Nov 1, 2026. Inputs: 1,000,000 requests a month of 2,000 input and 500 output tokens, Together's US$0.15 and US$0.60 per 1M, OpenRouter's rates for Together's endpoint and its lowest listed rates, and the 5.5% Standard fee on credits bought. It leaves out cached-input discounts, taxes and BYOK. | At 1,000,000 requests a month of 2,000 input and 500 output tokens: 2,000M input tokens at US$0.15 per 1M is US$300 and 500M output tokens at US$0.60 per 1M is US$300, so US$600 on Together serverless1 | The same workload served by Together's endpoint on OpenRouter Standard: US$600 of tokens plus the 5.5% credit fee of US$33, so US$633. At the model page's lowest listed rate of US$0.015 and US$0.057 per 1M: US$30 plus US$28.50 is US$58.50, plus a US$3.22 fee, so about US$61.72789 |
| Prompt and response storage by default | Stores prompts and responses by default and may use them for product improvements. Organization admins can turn storage off, which enables zero data retention. Passthrough models, allowed by default, forward prompts and responses to the upstream provider under its own data policy3 | Does not store prompts or responses unless you opt in to logging or content is flagged as potential CSAM. Requests may go to providers that store data unless you deny it, and provider-side retention can be turned off for the account or per request, with Zero Data Retention routing on every plan71011 |
| Dedicated capacity | Dedicated endpoints bill by the minute for each running replica, with one H100 80GB at US$5.49 an hour, and have no hard rate limits beyond the selected hardware and autoscaling bounds45 | Private Endpoints, an Enterprise add-on, route a fine-tuned model, a dedicated deployment of a public model or capacity reserved with a provider through OpenRouter. The deployment must sit with a provider OpenRouter supports for BYOK, reached with your own key712 |
| Rate limits | Serverless performance is best effort. When demand is high Together may limit requests, and it may apply custom rate limits or access restrictions to a model in especially high demand. Purchased rate limits on Enterprise and Scale contracts stay in place until the contract expires6 | The Free plan allows 50 requests a day. Standard and Business pass through the rate limits of the providers, and Enterprise can get higher allowances. Extra accounts or API keys do not change rate limits, which OpenRouter governs globally713 |
| Routing and fallback | Third-party models hosted on Together run on Together's own infrastructure at full precision, apart from passthrough models. When demand is high Together may return 429 or 503, and its docs tell you to retry with exponential backoff36 | By default, routes each request based on each provider's measured price, throughput, latency and reliability. Backup providers are allowed by default when the primary is unavailable, a list of models falls back to the next model on an error, and an attempt that errors is not charged for its model tokens71114 |
| Reported by practitionersFirst-hand accounts, each attributed and dated. They state no product facts. | The OpenHands evals team got a tools-unsupported error for Qwen3-Coder-Next on Together through LiteLLM and moved that model to OpenRouter (Feb 4, 2026). Mo Moustafa, who runs Olly, saw Together's endpoint on OpenRouter return empty 200 responses on the DeepSeek V4 Flash 0731 checkpoint (Sep 7, 2026)1617 | Mo Moustafa, who runs Olly, reports OpenRouter's per-host board scoring the same DeepSeek V4 Flash 0731 weights 90% on GPQA at DeepSeek and 75% at DigitalOcean, and pinning three hosts with fallbacks off took Olly down (Sep 7, 2026)17 |
| Choose it if | Choose Together AI if you want one host that runs open models on its own infrastructure at full precision, apart from passthrough models, with dedicated hardware from the same vendor, and can have an admin switch prompt storage off35 | Choose OpenRouter if you want one API key for every model and provider you approve, with backup providers allowed by default and Zero Data Retention routing on every plan, and will set a provider order or a list of providers to allow where host quality matters71115 |
What practitioners report, side by side
Each product's figures come from its own practitioners, for the group you pick.
Overall signal
All practitioners rate OpenRouter +10, 59 points above Together AI at −49. Most of the gap is in complaints, with 26 of Together AI's 40 sources negative against 30 of OpenRouter's 101.
The score counts the authors from all practitioners that take a side, one vote each with their latest view, and is how far positive votes lead negative ones, per 100 votes, so Together AI's −49 rests on votes split 5 positive, 8 mixed and 22 negative and OpenRouter's +10 on 40, 30 and 30.
A lead of a fifth of the votes reads positive and a lead of half strong positive, with the same lines below zero. The dots show each source, positive, mixed, negative. Run a pilot on your own prompts at expected concurrency to measure time to first token, output speed, error rate and cost per million tokens.
Footprint on the AI Application Platform stack
They differ on four of 9 layers. Together AI owns one that OpenRouter does not, and OpenRouter owns one that Together AI does not. Coverage is each product's placement on this stack, as its product page shows it. Core is a job the product owns; Supports and Touches are jobs it helps with. Coverage says nothing about how well a product does them.
| Layer | Together AI | OpenRouter |
|---|---|---|
| Model | ||
| Model gateway | Touches | Core |
| Operate | ||
| Observability | Touches | Supports |
| Cloud deployment | Core | Touches |
| Guardrails | Touches | Supports |
The 5 layers where they match
| Layer | Together AI | OpenRouter |
|---|---|---|
| Model | ||
| Model APIs | Supports | Supports |
| Build | ||
| Orchestration | Touches | Touches |
| Tool use | Touches | Touches |
| Data | ||
| Retrieval | Touches | Touches |
| Evaluate | ||
| Evaluation | Touches | Touches |
What practitioners praise
Among all practitioners, OpenRouter alone has three praise themes that reach 5 sources from 3 authors. One API and One Balance and Fast Access to New and Open Models lead, at 25 of 25 each.
| Praised for | Together AI | OpenRouter |
|---|---|---|
| One API and One Balance | Not a theme in its sources | 25 of 25 |
| Fast Access to New and Open Models | Not a theme in its sources | 25 of 25 |
| Spend Limits & Cost Data | Not a theme in its sources | 13 of 16 |
Each count is the sources that praise a theme, or raise it as a concern, out of those that discuss it. A row pairs the two products when both carry a theme under the same or an equivalent name.
A count ending in "so far" has under 5 sources from 3 authors among all practitioners. "Not a theme in its sources" means GILD found no theme of that name in that product's sources.
What practitioners raise as concerns
Among all practitioners, OpenRouter alone has four concern themes that reach 5 sources from 3 authors. Fees & Billing Surprises and Provider Quality Variance lead, at 20 of 25 and 16 of 18. Together AI alone has three concern themes that reach 5 sources from 3 authors. Tool Calling & API Compatibility and Python SDK & Integration Packages lead, at 9 of 9 and 5 of 5.
| Raised as a concern | Together AI | OpenRouter |
|---|---|---|
| Tool Calling & API Compatibility | 9 of 9 | Not a theme in its sources |
| Python SDK & Integration Packages | 5 of 5 | Not a theme in its sources |
| Errors, Rate Limits & Outages | 5 of 6 | Not a theme in its sources |
| Fees & Billing Surprises | Not a theme in its sources | 20 of 25 |
| Provider Quality Variance | Not a theme in its sources | 16 of 18 |
| Latency, Errors & Rate Limits | Not a theme in its sources | 14 of 16 |
| Free-Model Limits | Not a theme in its sources | 3 of 5 |
17 sources
- Together AI: gpt-oss-120b model page
- Together AI docs: Serverless overview
- Together AI docs: Privacy and security
- Together AI docs: Dedicated endpoints pricing
- Together AI docs: Dedicated endpoints overview
- Together AI docs: Serverless rate limits
- OpenRouter pricing and plan comparison
- OpenRouter: gpt-oss-120b model page
- OpenRouter API: gpt-oss-120b endpoints
- OpenRouter docs: Data collection
- OpenRouter docs: Provider routing
- OpenRouter blog: Private Endpoints (Oct 14, 2026)
- OpenRouter docs: Limits
- OpenRouter docs: Model fallbacks
- OpenRouter Enterprise
- OpenHands/software-agent-sdk issue #1899: move qwen3-coder-next from together.ai to OpenRouter (Feb 4, 2026)
- Mo Moustafa: So you want to use OpenRouter? (Sep 7, 2026)