AWS Bedrock’s August 2026 Update: Million-Token GPT-5.6, 14-Day Agent Sessions, and OpenAI’s Offensive Cyber AI

AWS published a recap of Amazon Bedrock and AgentCore updates shipped in August 2026, covering expanded context windows, longer-running agents, new OpenAI cybersecurity models, and expanded government-cloud availability. According to AWS, Bedrock is now used by more than 225,000 active customers, including over 80% of Fortune 100 companies.

Expanded context and global routing

OpenAI's GPT-5.6 Sol, Terra, and Luna models now support million-token context windows on Bedrock, with prompt caching to reduce cost and latency on repeated context. A new Web Search capability lets these models pull in live information beyond their training cutoff and return cited responses without a separate search integration.

Cross-Region inference now spans more than 25 AWS Regions for the GPT-5.6 family. "Global profiles" prioritize lower per-token pricing and capacity; "Geo profiles" restrict processing to a defined geography for data-residency requirements. OpenAI also cut list prices across Luna, Terra, and Sol on Bedrock, though AWS did not disclose the new per-token rates in this post.

On the cost-governance side, IAM principal cost allocation now attributes inference spend to specific users, teams, or cost centers, and AWS Cost Anomaly Detection has been extended to monitor third-party foundation model spend on Bedrock with root-cause breakdowns.

New cybersecurity models from OpenAI

OpenAI's Daybreak Blue and Daybreak Red are now available to eligible customers on Bedrock. Daybreak Blue is positioned for defensive work — vulnerability discovery, detection engineering, incident response. Daybreak Red is described as supporting "advanced, authorized" offensive tasks including vulnerability research, exploit reproduction, and mitigation development. AWS says the models run with stronger identity verification, monitoring, access controls, and "zero-operator-access" infrastructure, though it did not publish independent benchmark data for either model.

Longer-running, more constrained agents

AgentCore runtime instances can now run on dedicated EC2 compute — including GPU-accelerated, memory-optimized, and compute-optimized configurations — with sessions lasting up to 14 days, up from Bedrock's prior limits on agent session duration. AgentCore also expanded to US West (N. California) and Asia Pacific (Hyderabad).

New governance controls include temporal policies (enforcing action sequences, approval gates, and data-freshness checks between agent calls), rate limiting by user or group across requests, tokens, and concurrent connections, and AgentCore payments, which lets agents transact against APIs, MCP resources, and paid content within infrastructure-enforced spending caps.

AgentCore memory can now extract long-term memories from structured JSON sources such as activity and system logs, not just conversation transcripts, with fine-grained access control to isolate memories by user or tenant. A new AWS Agent Registry provides a searchable catalog of agents, MCP servers, and skills, with organization-wide detection for existing AgentCore Runtime and Gateway resources across connected AWS accounts.

Regulated environments

Anthropic's Claude Opus 5 is now available in AWS GovCloud (US) Regions with zero data retention enabled by default. OpenAI's GPT-5.6 Terra and Luna are also available in GovCloud with million-token context and prompt caching. Amazon Nova Multimodal Embeddings, supporting retrieval across text, documents, images, video, and audio, is now available in GovCloud (US-West).

AWS also released the Strands Agent Harness SDK as open source, intended to let developers build agents independent of any single framework and deploy them outside Bedrock if desired.

What this means

None of these updates constitute a new model release — GPT-5.6, Claude Opus 5, and Daybreak Red/Blue all appear to be existing models gaining new access points, context limits, or regional availability rather than new trained checkpoints. The substantive news is infrastructural: AWS is betting that the bottleneck for enterprise AI adoption is no longer raw model capability but the surrounding system — session persistence, spend controls, data residency, and audit trails. The 14-day agent runtime and infrastructure-enforced payment limits suggest AWS expects agents to take on genuinely long-horizon, autonomous work, which raises the stakes on the governance tooling (temporal policies, rate limits) shipped alongside it. The arrival of an explicitly offensive-security model (Daybreak Red) inside a major cloud platform, gated by "authorized" use and enhanced access controls, is a notable first for mainstream cloud AI infrastructure and will likely draw scrutiny over how "authorization" is verified in practice.