AWS Bedrock
AWS service exposing an integrated training + hosting chain to fine-tune and deploy foundation models (Claude Haiku, Llama, Titan, etc.). API via boto3 SDK: `create_model_customization_job` for training, `create_provisioned_model_throughput` for deploy. As of 2026, only Claude Haiku is fine-tunable on Anthropic side (Sonnet/Opus not exposed to public FT).
Strengths
- Only official path to fine-tune Claude Haiku as of 2026
- Training job + CloudWatch monitoring + deploy integrated in a single AWS stack
- Documented hyperparams (learningRateMultiplier, epochCount, batchSize) with conservative defaults
- Standard boto3 SDK, familiar vocabulary for any team already on AWS
Limitations
- Mandatory provisioned throughput for deploy = ~$1440/month minimum per PTU, kills ROI on small volumes
- Only Claude Haiku fine-tunable (not Sonnet or Opus), base model quality ceiling
- 1-month (On-demand) or 6-month (discounted) commitment, significant cash lock-in
- Cold-start latency on first call after inactivity, not suited for rare spikes
Best for
- Teams shipping a custom Claude agent with monthly volume ≥ 5M tokens
- Niche cases where advanced prompting + RAG plateau (ultra-strict format, business terminology)
- Teams already on AWS infra who want to avoid a new AI provider