AI
OpenAI Puts GPT-6 Astra on Amazon Bedrock
2026-09-11 - ABikram Mondal
The Bedrock rollout
OpenAI made GPT-6 Astra generally available on Amazon Bedrock on September 8. The move gives enterprise users direct API access through AWS infrastructure without needing separate OpenAI accounts for every workload.
The model supports a 1 million token context window. AWS handles the inference layer and adds its own security controls and audit logging.
Developers can call the model through Bedrock APIs or route existing ChatGPT Work and Codex tools to it. OpenAI also added enterprise plugins that extend browser use across common business applications.
Access started with a limited group of organizations. It is expanding to ChatGPT Plus, Pro, Business and Enterprise subscribers over the coming days, plus the OpenAI API, Microsoft Azure and AWS Bedrock.
Indian companies already using AWS for compliance reasons can now test the model without new vendor approvals.
Benchmark numbers from the release
OpenAI reports 98 percent on FrontierMath Tier 4 v2. The same model hit 100 percent on ExploitBench, its internal cybersecurity test for developing working exploits.
On ARC-AGI-3 the score reached 99.9 percent under OpenAI's custom harness that keeps reasoning active across turns. The standard harness produced 62.7 percent.
Computer use benchmarks include 59.3 percent on Agents Last Exam and 72.6 percent on OSWorld 2.0 offline set. These figures sit above GPT-5.6 Sol numbers in OpenAI's own tables.
The company states Astra completes 41.4 percent of tasks on AutomationBench, up from 31.4 percent for Claude Fable 5.1.
These are the headline figures the company published. Independent runs on the same harnesses have not yet appeared in public leaderboards.
What the model actually does
Astra handles multistep computer and browser tasks with less step-by-step guidance than earlier versions. Examples include filling forms, navigating maps, and updating spreadsheets while following organizational templates.
Coding performance improved on agentic benchmarks. OpenAI lists gains in software engineering and troubleshooting workflows that run for hours without constant human input.
Science tasks include writing analysis code, executing it, reading errors and self-correcting. The model also works directly with specialist tools such as sequencing quality checks and cell-tracking pipelines.
Output quality targets professional documents, presentations and research summaries. The model aligns responses to company voice and standards when given the right prompts and plugins.
Context length of one million tokens lets teams load entire codebases or document collections in one pass.
Independent scores and price
Artificial Analysis placed GPT-6 Astra at 61 on its Intelligence Index. That matches the score for GPT-5.6 Sol and sits five points behind Claude Fable 5.1.
Pricing on the OpenAI API and Bedrock stands at 10 dollars per million input tokens and 50 dollars per million output tokens. Batch mode halves those rates.
The price is 2.5 times higher than GPT-5.6 Sol on both input and output. Cached input pricing is available on Bedrock but still leaves the model among the more expensive frontier options.
Enterprises that need the computer-use and cybersecurity edges may accept the cost. Teams running high-volume simple queries will likely stay with cheaper models.
Capacity pressure and limits
Demand for Astra has already forced OpenAI to pause new Pro tier signups. The Pro plan strains infrastructure more than other tiers, according to product lead comments reported in recent coverage.
Plus, Go and API plans remain open for now. Existing Pro users keep access.
OpenAI warned that usage limits may tighten further if capacity does not keep pace. This follows the pattern seen with earlier flagship releases that sold out quickly.
Bedrock availability gives AWS customers another route, but overall supply remains constrained for the highest-demand users.
Who should pay attention
Teams building autonomous agents or handling long-running coding and research workflows now have a new option with documented gains on computer-use benchmarks. Enterprises already on AWS gain a compliant path to test those capabilities.
Startups or small teams doing routine chat or summarization work will find the price and capacity limits unattractive compared with faster, cheaper alternatives still in wide use.
Indian developers working on automation for business processes can evaluate the model through Bedrock without new contracts. ABikram Mondal builds automation for exactly this kind of problem at https://abikrammondal.com/services/automation.
Anyone waiting for lower prices or broader availability should monitor usage limits and independent benchmark updates over the next few weeks.
Sources
- https://report-ai.org/indexes/technical-benchmarks/best-ai-models-2026/
- https://drawpie.com/blog/gemini-3-5-pro-vs-gpt-5-6-vs-claude-fable-5/
- https://9to5google.com/2026/09/03/openai-gpt-6-astra-launch/
- https://www.newsnow.com/ca/AI
- https://siliconangle.com/2026/06/26/openai-introduces-gpt-5-6-challenge-claude-mythos-5/
- https://agentlocker.ai/news/openai-releases-gpt-6-astra-ai-model
- https://www.nytimes.com/2026/07/09/technology/openai-sol-ai.html
- https://www.aichatdaily.com/ai-news
Reported from the sources above on 2026-09-11. Figures are as published at the time of writing. If something here has moved on, the linked source is the one to trust.
If you got here because you are actually thinking about putting models like this to work inside a real business, wired into the tools a team already uses, that is the work I do. I build for founders and small teams who want the thing to exist and work, not a deck about it.