# Amazon Bedrock vs Thinking Machines

> Bedrock wraps 100+ models in AWS security with managed fine-tuning. Thinking Machines gives researchers low-level control of post-training on open weights.

Canonical: https://www.subconscious.dev/compare/aws-bedrock-vs-thinking-machines · By The Subconscious Team · Updated September 30, 2026

## How they compare

Bedrock is a managed model service for enterprises already on AWS. One API reaches Claude, GPT-6 Astra, Nova, DeepSeek and 100+ other models, and every call inherits IAM, PrivateLink, KMS and CloudTrail. Customization comes through managed fine-tuning and Custom Model Import for open-weight checkpoints. Thinking Machines targets a different buyer. Tinker exposes forward_backward, optim_step, sample and save_state, so teams write their own SFT or RL loop instead of filling out a fine-tuning job, and the lab runs the distributed GPU work. Training is LoRA only, across bases like Kimi K2.6, GLM-5.3, Qwen3.5 and Inkling, billed per million prefill, sample and train tokens.

Bedrock wins on production serving and compliance. AgentCore adds a serverless agent runtime with memory, an MCP gateway and policy controls, and billing options include batch at 50% off and provisioned throughput. Third-party analyses put most Bedrock models 20 to 35% above direct prices, with Claude at parity. Thinking Machines cannot yet serve user traffic at scale: its serverless API is beta and Inkling-only, and checkpoint sampling is scoped to testing. Its edge is depth of control over training, including RL, on models that are hard to train in-house. One plausible path is training LoRA adapters on Tinker and bringing the result into a governed host.

## What each one does

### Amazon Bedrock

Amazon Bedrock is AWS's managed model service and has become the default AI control plane for many enterprises. One API reaches 100+ models from 18+ providers, including Anthropic's Claude family, Meta, Mistral, DeepSeek, Amazon's own Nova models, and, since an April 2026 partnership expansion, OpenAI models up to GPT-6 Astra. Switching models is usually just a new model ID. Every call inherits IAM, PrivateLink, KMS encryption and CloudTrail logging, and provider models never train on customer data.

### Thinking Machines

Thinking Machines Lab is the San Francisco lab Mira Murati, formerly CTO of OpenAI, founded in February 2025, with OpenAI co-founder John Schulman as chief scientist. It raised about $2 billion at a $12 billion valuation in July 2025 in a round led by Andreessen Horowitz, and in March 2026 signed a multi-year Nvidia deal for one gigawatt of Vera Rubin capacity. Its main developer product is Tinker, an API for post-training open-weight models that launched in October 2025 and is now generally available. Tinker exposes four low-level calls, forward_backward, optim_step, sample and save_state, so teams write their own supervised or reinforcement learning loops while Thinking Machines runs the distributed GPU work. Training uses LoRA adapters rather than full weight updates.

## Which is best, and when

### Choose Amazon Bedrock for

- Regulated teams that need IAM, KMS and CloudTrail on every call
- Multi-model agents on AgentCore
- Simple managed fine-tuning inside AWS

### Choose Thinking Machines for

- Custom RL loops that managed fine-tuning cannot express
- Post-training Kimi K2.6 or Inkling without a cluster
- Research iteration with mid-training checkpoint sampling

## At a glance

| Attribute | Amazon Bedrock | Thinking Machines |
|---|---|---|
| Model access | Closed and open, 100+ models | Open weights |
| Flagship models | Claude, GPT-6 Astra, Nova, DeepSeek | Inkling, Inkling-Small |
| Speed | Latency-optimized option on some models | - |
| Price | ~20–35% above direct; Claude at parity | Per 1M tokens by prefill, sample, train; Inkling $1.00 in, $4.05 out |
| Customization | Fine-tuning, Custom Model Import | LoRA SFT and RL via Tinker |
| Deployment | Managed on AWS, AgentCore | Training API, beta serverless (Inkling only) |
| Long context | Varies by model | Inkling up to 1M; Tinker 32K–256K |

## FAQ

### What is the difference between Amazon Bedrock and Thinking Machines?

Bedrock wraps 100+ models in AWS security with managed fine-tuning. Thinking Machines gives researchers low-level control of post-training on open weights.

### When should I choose Amazon Bedrock over Thinking Machines?

Regulated teams that need IAM, KMS and CloudTrail on every call; Multi-model agents on AgentCore; Simple managed fine-tuning inside AWS.

### When should I choose Thinking Machines over Amazon Bedrock?

Custom RL loops that managed fine-tuning cannot express; Post-training Kimi K2.6 or Inkling without a cluster; Research iteration with mid-training checkpoint sampling.

### Is Amazon Bedrock or Thinking Machines cheaper?

Amazon Bedrock: ~20–35% above direct; Claude at parity. Thinking Machines: Per 1M tokens by prefill, sample, train; Inkling $1.00 in, $4.05 out. The cheaper choice depends on the model and workload.

### Which has more context, Amazon Bedrock or Thinking Machines?

Amazon Bedrock: Varies by model. Thinking Machines: Inkling up to 1M; Tinker 32K–256K.

## Running long-horizon agents?

If your agents run past 200K tokens, compare both against Subconscious: [Subconscious vs Amazon Bedrock](https://www.subconscious.dev/compare/subconscious-vs-aws-bedrock.md), [Subconscious vs Thinking Machines](https://www.subconscious.dev/compare/subconscious-vs-thinking-machines.md).

Full profiles: [Amazon Bedrock](https://www.subconscious.dev/providers/aws-bedrock.md), [Thinking Machines](https://www.subconscious.dev/providers/thinking-machines.md).