Alpha
Contact
Assess · Together AI

Can you own it?

Ownership levelPartialnone·limited·partial·substantial·fullAnalytical input C ยท 69.6/100

This page is a projection of the one entry record, the Use & modify and Transparency factors that Assess covers. The full verdict is set by all four factors together, floor-weighted so the weakest caps the whole.

Which domain expands which factor
  • AssessUse & modify + Transparency
  • ImplementData control + Reliability
  • UseReliability
  • SupportTransparency

Data governance - retention, ZDR & training

Together stores your prompts by default and may use them for product improvements; zero data retention is an opt-in. The product docs say so directly: "By default, Together stores the prompts you send and the responses models return, and may use them for product improvements." An organisation admin turns off Prompt storage in Privacy settings to enable ZDR, and ZDR applies only to requests after it is enabled. Under ZDR request content is not persisted. Request metadata, account data and files you uploaded for fine-tuning or batch jobs are still retained. Training on your data is a separate opt-in that is off by default (Privacy Policy Sec 1 and 2.2). A second default matters: passthrough models are ON. Some models forward your prompts and responses to an upstream provider that handles them under its own data policy, and turning ZDR on switches passthrough off. All of this is Together's own statement and has not been audited. Terms Sec 9 adds that "The parties will have no confidentiality obligations to each other unless otherwise agreed in writing", which caps this dimension at 3.

Data & IP ownership

The Terms of Service (Sec 7, updated 2026-05-19) settle ownership: as between you and Company, you exclusively own all right, title, and interest in Your Content and Output, with a limited licence to operate them "solely as necessary to provide the Services". Together owns Usage Data (aggregated or de-identified technical data it may use to improve and develop services). Two limits apply. By default stored prompts may be used for product improvements, which sits awkwardly with the narrow licence, and passthrough models place your data under a third party's policy. The DPA's Annex III names nine sub-processors, including OpenAI OpCo, LLC for "Generative AI". Annex 1 scopes the DPA around account and employee data, so confirm how personal data inside prompts is covered.

Compliance & attestations

Together announced SOC 2 Type II (blog, 2025-07-08, auditor not named) and ISO 27001:2022 from A-LIGN, an ANAB-accredited body (blog, 2026-06-10). Both are vendor announcements. The SOC 2 report and the ISO certificate are available only on request, and the trust centre is a JavaScript-only page with no readable content. A GDPR DPA with EU SCCs and the UK Addendum is published as a PDF. HIPAA and BAAs are stated in the SOC 2 blog, yet ToS Sec 4 forbids sending "financial or medical information" or sensitive personal data under the standard terms, so obtain a signed BAA and enterprise terms before sending PHI. ISO 27701 is not claimed.

Data residency & jurisdiction

Serverless has a US-only guarantee and no EU option. The docs offer one region control: send requests to us-api.together.ai (1.25x the standard rate, 8 models) to keep inference processing in the US. That guarantee "doesn't cover where data is stored, where logs are written, or how network traffic is routed", and requests to api.together.ai "do not guarantee US processing". EU residency means a dedicated endpoint or VPC deployment arranged with sales. The support article that says so is a year old and gates it to Scale and Enterprise tiers, which the rate-limit docs call legacy contracts. As a US-headquartered company Together also carries standing US CLOUD Act exposure that no region setting removes. The DPA PDF has no statement of default North America routing. The nearest statement is the docs line that DeepSeek models are hosted in North America data centres.

Security controls

The DPA's Annex 2 commits Together to a written security programme, MFA for remote access, encryption of controller data in transit and storage, daily vulnerability scans, annual external penetration tests, 30-day remediation of critical findings and USD 2 million of cyber insurance. Vulnerability reports go through a private HackerOne programme. The SOC 2 and ISO announcements describe independent audits. No penetration-test report, SOC 2 report or ISO certificate is public, so the posture is solid but partial. The docs line "models are hosted at full precision" is contradicted by the catalogue's FP4, FP8, NVFP4 and MXFP4 column, so treat marketing claims on that page with care.

Pricing & cost model

Serverless per-token rates with cached-input prices are published for 24 chat models. Examples per 1M tokens: gpt-oss-120B $0.15 in and $0.60 out, DeepSeek V4.1 Flash $0.30 and $1.20, GLM-5.3 $1.40 and $4.40, Kimi K3.1 Pro $2.00 and $10.00, MiniMax M3.1 Flash $0.15 and $0.60. US-only processing costs 25% more. Dedicated single-tenant GPUs are billed on demand per GPU hour (H100 $5.49, H200 $5.99, B200 $9.29), GPU Clusters start at $3.99 an hour for H100, and provisioned throughput is sold through sales. Rates and promotions (Qwen4 Max to 2026-11-23) move often, so confirm before committing.

Reliability posture

A public status page exists and shows per-model uptime, but there is no self-serve SLA. status.together.ai reports 90-day availability from 99.28% (Kimi K3) to 100%, for example gpt-oss-120B 99.66% and GLM-5.3 99.91%. Its feed has two 2026 maintenance notices and no incident posts since January 2025. The Terms (Sec 1) give "no guarantees ... uptime, or reliability" unless an order form says otherwise, the docs call serverless "best-effort", and a defined SLA comes only with provisioned throughput through sales.

How this scores

The ownership factors this domain covers, drawn from the one entry record.

1

Use and modify freelyCan you use it freely and leave without lock-in?

Strong

OpenAI-compatible /v1 API serving portable open-weight models - the same checkpoints run on other providers or self-hosted, so exit cost is low.

How this scores (AOI sub-dimensions)
Transparency & lock-in4/5how portable it is and how easily you can leaveOpenAI-compatible /v1 API (ev6; embeddings and moderation are not served on serverless) over portable open-weight models (DeepSeek, GLM, Kimi, Qwen, gpt-oss, MiniMax).
Cost4/5how the pricing model compares and how predictable it isPricing is published directly (ev5): per-token serverless rates with cached-input prices for 24 chat models, a published 25% premium for US-only processing, on-demand per-GPU-hour dedicated and cluster rates, fine-tuning and PTU pricing, with dated promotions.
2

TransparencyAre the binding terms published, legible and independently checkable?

Moderate

The Terms, Privacy Policy and DPA are published, but the posture is Together's own self-attestation. Sub-processors appear only in the DPA Annex III, and the trust centre has no readable content. Some docs claims are contradicted by Together's own catalogue (full precision versus FP4 and FP8 quantisation).

How this scores
Not a scored AOI dimension. For a hosted provider, transparency is whether the binding terms are published, legible and were actually read - the read/unverified evidence below, not a certification. A strong rating here must trace to a retrieved binding document.
What this means for adoptionYou own your inputs and outputs under the Terms, training is opt-in, and the open-weight catalogue on an OpenAI-compatible API gives a clean portable exit. Ownership is partial. By default Together stores your prompts and may use them for product improvements, and zero data retention is an admin opt-in. Passthrough models send prompts to third parties until you switch them off, and the Terms disclaim any duty of confidence. Residency is US-only on serverless, with EU limited to sales-led dedicated deployments. Turn prompt storage off for ZDR, negotiate confidentiality and EU terms in an enterprise contract, and do not rely on the default terms for trade secrets or regulated data.

Sources

The same evidence records as the entry sheet. Read means the text was verified; unverified means it is known to exist but not yet read.

Privacy Policyread2026-11-19
Privacy Policy (updated 2025-12-17).
Documentationread2026-11-19
Privacy and security docs: see ev-docs for the default-storage, passthrough and ZDR clauses.
Vendor announcementread2026-11-19
SOC 2 Type II announcement (published 2025-07-08): completed via an independent multi-month audit, auditor not named; HIPAA adherence and business associate agreements stated.
Documentationread2026-11-19
Support knowledge base article, "Last updated 1 year ago": EU data centres exist, and dedicated endpoints on targeted data centres including EU locations are available only on Scale and Enterprise tier plans.
Vendor announcementread2026-11-19
Pricing page: serverless per-token rates with cached-input prices for 24 chat models plus vision, image, video and audio models; on-demand dedicated GPU endpoints (H100 $5.49 per GPU hour); GPU clusters; fine-tuning; provisioned throughput via sales; promotions with end dates.
Documentationread2026-11-19
OpenAI-compatibility docs: an OpenAI-compatible REST API and SDKs across chat, completions, vision, image generation, text-to-speech and speech-to-text, used by pointing an OpenAI client at Together's base URL.
Terms of serviceread2026-11-19
Terms of Service (updated 2026-05-19).
Third-party analysisread2026-11-19
Artificial Analysis provider page: independent third-party analysis of 18 Together AI models by intelligence, speed and price (Kimi K3.1 Pro ranked first), showing the same open-weight models compared across providers.
Data Processing Addendumread2026-11-19
Data Processing Addendum PDF (no effective date on the PDF): Together (US) Inc.; EU SCCs and UK Addendum; the conflict clause prefers the DPA for personal data it covers; 24-hour breach notice (Annex says 48).
Documentationread2026-11-19
Privacy and security and zero-data-retention docs.
Documentationread2026-11-19
Geo-specific inference docs: only a United States endpoint (us-api.together.ai, 1.25x standard rates, 8 models) keeps inference processing in the US.
Vendor announcementread2026-11-19
ISO 27001:2022 blog (published 2026-06-10): Together AI "has received an ISO 27001:2022 certification from A-LIGN Compliance and Security, Inc., an ANAB-accredited certification body", scoped to the ISMS for the global platform including third-party data centres.
Slaread2026-11-19
Public status page (rendered page, index.json and feed): "All services are online".
Securityunverified2026-11-19
Trust centre (trust.together.ai): a Vanta single-page app shell of about 7.9 KB with no readable content.