IBM watsonx.ai
Listed at https://www.ibm.com/products/watsonx-ai
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | Per-1M-token pricing ranges from $0.06 (Granite 4 H Small input) to $3.00 (Mistral Large / Mistral Medium 2505) | per million tokens |
| Output | Per-1M-token output pricing ranges from $0.10 (Granite Guardian 2B / Llama 3.2 1B) to $10.00 (Mistral Large / Mistral Medium 2505) | per million tokens |
🤖 Supported Models (28)
✨ Pros
- ✓Granite 4.2 family (released Aug 25, 2026): 3B / 8B / 30B sizes with native chain-of-thought reasoning, tool calling, 128K native context (extendable to 512K), Apache 2.0 license
- ✓28 Foundation Models: IBM Granite (proprietary) + Meta Llama (1B-90B) + Mistral (Small / Medium / Large / Pixtral) + OpenAI gpt-oss-120B + Google Flan-T5 + SDAIA Allam
- ✓Granite Speech 5.0 Turbo CTC: 470M-parameter speech recognition model with CTC architecture, suitable for on-device / edge deployment
- ✓Enterprise compliance: HIPAA / FedRAMP / Financial Services Validated / EU Supported multi-cert, supports on-prem deployment on IBM Z / LinuxONE
- ✓128K-512K long context: Mistral Large 131K, Llama 3.2 11B/90B Vision 128K, Granite 4.2 128K native + 512K extended
- ✓Watsonx.ai Runtime integrates watsonx Orchestrate / watsonx Governance / watsonx Data: enterprise-grade Agentic workflow orchestration + data governance + audit trail
⚠️ Cons
- ×IBM Cloud IAM auth is more complex than OpenAI API Key: requires API Key → Bearer Token exchange before calling ml/v1 endpoints (~50ms auth overhead)
- ×28 models is far fewer than Eden AI (1,009) / OpenRouter (400+) / Cloudflare AI Gateway (aggregating all Workers AI models)
- ×Mainland China access requires overseas egress (200-400ms); no ICP filing or domestic mirror
- ×No permanent free tier; new users must register IBM Cloud first to get trial credits
- ×Mistral Large pricing ($3/$10) is notably more expensive than Mistral Large routed via OpenRouter; not ideal for cost-sensitive pure-LLM workloads
- ×Granite Speech 5.0 supports limited languages; SLA on watsonx.ai platform is less explicit than OpenAI Whisper's
🎯 Best For
Enterprise agentic workflows with strict compliance (HIPAA / FedRAMP / FS Validated); on-prem deployment on IBM Z / LinuxONE; hybrid Granite + Llama + Mistral multi-model strategy; long-context (128K-512K) RAG pipelines; speech-to-text on edge devices
💰 Pricing & Plans
| Model | Input ($/M) | Output ($/M) | Context |
|---|---|---|---|
| IBM Granite 4 H Small | $0.06 | $0.25 | 20K |
| Granite Guardian 3.2 2B / Granite Vision 3.2 2B / Llama 3.2 1B | $0.10 | $0.10 | 8K-128K |
| Llama 3.2 3B Instruct | $0.15 | $0.15 | 128K |
| Granite 3 8B / Granite 3.3 8B / Granite Guardian 3.3 8B | $0.20 | $0.20 | 8K |
| OpenAI gpt-oss-120B | $0.15 | $0.60 | 8K |
| Llama 3.3 70B Instruct | $0.71 | $0.71 | 128K |
| Llama 4 Maverick 17B | $0.35 | $1.40 | 128K |
| Granite 13B Chat / 13B Instruct | $0.60 | $0.60 | 8K |
| Mistral Small 2503 / Small 3.1 24B | $0.10 | $0.30 | 32K |
| Mistral Large / Mistral Medium 2505 | $3.00 | $10.00 | 128K-131K |
🔧 API & Developer Experience
- •Base URL: https://{region}.ml.cloud.ibm.com (region: us-south / us-east / eu-de / eu-gb)
- •Chat Completions: POST /ml/v1/text/chat (OpenAI-compatible payload + IBM-specific parameters)
- •Text Generation: POST /ml/v1/text/generation (legacy endpoint)
- •List Foundation Models: GET /ml/v1/foundation_model_specs (returns all 28 models + capabilities)
- •Embeddings: POST /ml/v1/embeddings (Granite + Slate embedding models)
- •Authentication: IBM Cloud IAM API Key → Bearer Token via POST /identity/token (~50ms auth overhead, token TTL 1h)
- •Python SDK: ibm-watsonx-ai on PyPI (pip install ibm-watsonx-ai); Node.js SDK available
- •Streaming: Yes (Server-Sent Events on /ml/v1/text/chat_stream)
🏢 Enterprise & Compliance
IBM watsonx.ai ships five enterprise compliance stamps out of the box: HIPAA Enabled (US healthcare), FedRAMP Moderate (US federal agencies), Financial Services Validated (banks/insurance), EU Supported (GDPR + EU AI Act data residency in eu-de/eu-gb), and IAM-enabled role-based access. The Enterprise tier unlocks private deployment on IBM Z (z16 / z17) and LinuxONE (III / IV) mainframes. The Granite 4.2 family (released Aug 25, 2026) targets agentic workflows -- the 8B and 30B models add reasoning-augmented tool calling trained on agentic traces via multi-stage RL (foundational RL + agentic RL + RLHF) on an NVIDIA GB200 NVL72 cluster hosted in CoreWeave. Combined with watsonx Orchestrate for orchestration, watsonx Governance for risk management, and watsonx.data for the lakehouse, watsonx.ai is the compliance-first foundation-model platform rather than a general-purpose LLM gateway.
🌐 Regional Availability & Latency
IBM watsonx.ai exposes four production regions: us-south (Dallas, TX), us-east (Washington DC), eu-de (Frankfurt, Germany), and eu-gb (London, UK). Each region runs an independent ml.cloud.ibm.com endpoint with full data residency -- traffic originating in eu-de never leaves the EU, satisfying GDPR Art. 44 transfer-impact assessments. Cross-region routing is not supported; select region at IBM Cloud project creation. Mainland China access requires overseas egress (200-400ms typical). IBM does not operate a China-domestic ICP-registered mirror, so Chinese-facing workloads should pair watsonx.ai with a domestic provider (Volcano Ark, SiliconFlow, or FreeModel). Enterprise tier unlocks on-prem deployment on IBM Z16/Z17 and LinuxONE III/IV, letting regulated banks and healthcare networks run Granite 4.2 entirely inside their own data center with no internet egress.