Short answer: AWS already offers OpenAI’s open-weight gpt-oss-120b and gpt-oss-20b. OpenAI released them under Apache 2.0 on August 5, 2025, and AWS made them available through Amazon Bedrock and Amazon SageMaker AI the same day. That did not amount to Amazon secretly bypassing Microsoft: the models were distributed with OpenAI’s knowledge and approval, and they are separate from OpenAI’s proprietary GPT products.
AWS later gained access to proprietary OpenAI models through a different development. OpenAI and Microsoft revised their partnership in April 2026, making Microsoft’s OpenAI intellectual-property license non-exclusive through 2032. AWS subsequently brought proprietary models to Bedrock, with GPT-5.5, GPT-5.4 and Codex generally available from June 1, 2026.
What Amazon actually offered
The original AWS announcement concerned two text-only, open-weight reasoning models:
- gpt-oss-120b
- gpt-oss-20b
They are available through Amazon Bedrock and Amazon SageMaker AI. OpenAI also made the weights downloadable through its open-models program and associated tooling. They are not ChatGPT, are not ordinary hosted GPT API endpoints, and are not available through ChatGPT or the standard OpenAI API, as OpenAI explains in its support documentation.
Free tools Windows power users keep installed
One-click scans. No signup required.
#1 Best Overall
- Use scikit-learn to track an example ML project end to end
- Explore several models, including support vector machines, decision trees, random forests, and ensemble methods
- Exploit unsupervised learning techniques such as dimensionality reduction, clustering, and anomaly detection
- Dive into neural net architectures, including convolutional nets, recurrent nets, generative adversarial networks, autoencoders, diffusion models, and transformers
- Use TensorFlow and Keras to build and train neural nets for computer vision, natural language processing, generative models, and deep reinforcement learning
Timeline: open weights first, proprietary access later
| Date | What happened | Why it matters |
|---|---|---|
| August 5, 2025 | OpenAI released gpt-oss-120b and gpt-oss-20b; AWS announced them on Bedrock and SageMaker AI. | This was AWS’s first offering of OpenAI models. |
| February 27, 2026 | OpenAI and Microsoft said Microsoft retained an exclusive license and access to OpenAI intellectual property across models and products, while collaborations such as Amazon’s were contemplated under their agreements. | The statement addressed how the Amazon collaboration fit within the existing relationship. |
| April 27–28, 2026 | OpenAI said Microsoft’s license would continue through 2032 but become non-exclusive; OpenAI announced models, Codex and Managed Agents for AWS. | This changed the contractual basis for AWS access to proprietary OpenAI products. |
| June 1, 2026 | AWS announced general availability of GPT-5.5, GPT-5.4 and Codex on Bedrock. | These are managed proprietary models, not downloadable gpt-oss weights. |
Sources: OpenAI’s gpt-oss announcement, AWS’s launch announcement, OpenAI’s February partnership statement, the April partnership update, OpenAI on AWS and AWS’s June availability notice.
What “open-weight” means
OpenAI published the trained numerical weights, allowing developers to download, run, fine-tune and deploy the models on infrastructure they control or through a hosting provider. That is materially different from a closed API: the deployer can choose the runtime, hardware, network boundary and tuning process.
“Open-weight” does not mean every part of the system is open source. Training data, training infrastructure, some tooling and hosted product layers can remain proprietary. OpenAI uses “open models” and “open-weight models” rather than claiming that the entire surrounding stack is unrestricted.
Rank #2
What Apache 2.0 permits—and what it does not
Apache 2.0 is a permissive license. In general, it allows commercial use, modification and redistribution without the copyleft obligations associated with licenses such as the GPL. A business can therefore incorporate the weights into a product or run them for internal workloads, subject to the license’s notice and attribution requirements.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Apache 2.0 is not a waiver of every obligation. OpenAI distributes gpt-oss under the license together with a separate gpt-oss usage policy. Organizations must also consider privacy, copyright, export controls, sector rules, security requirements, contracts and laws applicable to their deployment. The license does not guarantee safe outputs, remove liability or authorize harmful use.
Model specifications and capability claims
The following are specifications and benchmark claims reported by OpenAI, not independent production testing:
| Specification | gpt-oss-120b | gpt-oss-20b |
|---|---|---|
| Approximate total parameters | 117 billion | 21 billion |
| Approximate active parameters per token | 5.1 billion | 3.6 billion |
| Maximum context | 128,000 tokens | |
| OpenAI comparison | Approaches o4-mini on core reasoning benchmarks | Similar to o3-mini on selected evaluations |
| Memory target stated by OpenAI | Approximately 80 GB | Approximately 16 GB |
Both support adjustable reasoning effort, tool use and structured outputs. The memory figures reflect OpenAI’s stated deployment assumptions, including MXFP4 quantization; they are not universal minimum hardware requirements. Runtime overhead, context length, batch size, KV-cache use and CPU/GPU choice affect actual capacity, latency and throughput. Benchmark parity also does not establish equal reliability, hallucination rates, tool behavior or total cost in a production application.
Bedrock, SageMaker or self-hosting?
| Option | What you get | Control and work | Best fit |
|---|---|---|---|
| Amazon Bedrock | Managed model access, AWS billing and application integrations. | Least infrastructure work; subject to regions, quotas, service behavior and AWS dependence. | AWS-centric organizations needing centralized identity, governance, networking and procurement. |
| Amazon SageMaker AI | Model deployment, endpoint management, customization and MLOps workflows. | More operational control, but the team manages more of the serving lifecycle. | ML engineering teams already using SageMaker for development and deployment. |
| Self-hosting | Downloaded weights on private or rented infrastructure. | Maximum control and customization; the customer owns GPUs, monitoring, upgrades, security and reliability. | Controlled environments, fine-tuning, predictable high-volume workloads or local inference. |
| Other hosted providers | Alternative runtimes and infrastructure economics. | Portability and regional choices vary; inspect data handling, uptime, quotas and contracts. | Teams not tied to AWS or seeking specialized serving stacks such as vLLM or Ollama. |
Bedrock’s managed layer can reduce GPU and endpoint administration and can consolidate AWS identity, billing and audit controls. SageMaker is better suited to teams that need deployment and MLOps control. A direct download is not “free inference”: storage, GPU time, electricity or rental, engineering, monitoring and support remain costs.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →Did Amazon sidestep Microsoft?
Only in a limited, and potentially misleading, sense. Apache 2.0 enabled broad distribution of the gpt-oss weights; it did not independently cancel Microsoft’s contractual rights or compel AWS to receive proprietary GPT models. Amazon said the 2025 offering was made with OpenAI’s knowledge and approval.
Rank #4
The later AWS access to proprietary models followed the April 2026 contractual change. OpenAI said Microsoft retained a license to OpenAI intellectual property through 2032, but that license was no longer exclusive. Microsoft therefore did not lose all access, and the change should not be conflated with the 2025 open-weight release.
gpt-oss versus proprietary OpenAI models on Bedrock
| Issue | gpt-oss | Proprietary OpenAI models on Bedrock |
|---|---|---|
| Examples | gpt-oss-120b and gpt-oss-20b | GPT-5.5, GPT-5.4 and Codex |
| Weights | Downloadable | Not released as open weights |
| Access model | Self-hosting or provider-managed inference | Managed service access |
| License or terms | Apache 2.0 plus the gpt-oss usage policy | Service and model-provider terms |
| Self-hosting | Yes, subject to applicable terms | No ordinary self-hosting of the model weights |
| Pricing signal | Infrastructure or provider-dependent | AWS says listed Bedrock pricing matches OpenAI first-party rates; verify model and region details |
What AWS customers actually pay for
The downloadable weights do not eliminate the rest of a production budget. Depending on the design, costs can include:
- Bedrock inference or SageMaker endpoint capacity.
- GPU instances, storage and networking for self-hosted deployments.
- Data transfer, logging, monitoring and security services.
- Fine-tuning, evaluation and model-serving engineering.
- High availability, incident response, upgrades and support.
AWS says eligible usage for the generally available proprietary Bedrock offerings can count toward existing AWS commitments. Check the Bedrock pricing page, SageMaker AI pricing, region availability and quotas before budgeting. No deployment choice is automatically cheapest: predictable volume may justify owned or reserved capacity, while variable demand often favors managed inference.
Best Value
Operational and security limits to check
Availability and access
If a model is missing in the Bedrock console, check the selected AWS Region, account permissions, model-access or marketplace settings, and whether the listing is preview or generally available. Availability can differ by model and region.
API compatibility
OpenAI-related formats and tools do not make gpt-oss identical to the OpenAI API. Runtime adapters, system prompts, tool wiring, tokenization and safety behavior can differ. Test your application against the exact Bedrock or local runtime.
Safety and governance
OpenAI’s model card notes that open weights have a different risk profile: determined users can fine-tune them to bypass refusals, and access cannot be revoked after release. Validate structured outputs, authorize every tool call, restrict secrets, monitor prompts and outputs, and apply your organization’s privacy and retention controls.
Data handling
Managed inference still requires review of retention, logging, regional processing, encryption, private networking and contractual data-processing terms. Do not assume controls advertised for a later proprietary Bedrock integration apply identically to every gpt-oss deployment mode.
Recommended Free Tools
Quick Recap
Who should choose which route?
Choose Bedrock when
- Your organization already operates mainly on AWS.
- IAM, auditability, networking and consolidated billing are priorities.
- You want managed inference and the ability to evaluate multiple providers through one AWS service.
Choose SageMaker AI when
- You need endpoint, customization and MLOps control.
- Your ML team already trains, evaluates and deploys models in SageMaker.
Self-host gpt-oss when
- Data must remain in a tightly controlled environment.
- You need custom fine-tuning or inference behavior.
- Workload volume is high and predictable enough to support GPU operations.
Use another provider when
- You need a different region, latency profile, runtime or hardware price.
- Your application is not otherwise dependent on AWS.
- You have assessed the provider’s data handling, uptime, quotas and contract terms.
What this does not mean
- gpt-oss is not GPT-5, ChatGPT or a drop-in replacement for every OpenAI hosted product.
- Apache 2.0 does not remove the gpt-oss usage policy or other legal duties.
- Open weights do not mean the training data, infrastructure and entire product stack are open.
- The stated 16 GB and 80 GB figures do not guarantee production throughput or latency.
- Self-hosting is not automatically cheaper than Bedrock after hardware, people and operations are included.
- Microsoft did not lose all OpenAI access; its license continues through 2032.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
