OpenAI Launches GPT-6 Astra With 1.9x Faster Computer Use and Multi-Cloud Support

·1 views

OpenAI has begun a phased rollout of GPT-6 Astra, its new flagship AI model, succeeding GPT-5.6 Sol with claimed improvements in reasoning, coding, cybersecurity, and computer-use tasks. The model completed tasks 1.9 times faster than its predecessor on the Mind2Web benchmark and also showed gains on OSWorld and FrontierMath evaluations. Astra is being made available to ChatGPT Plus, Pro, Business, and Enterprise subscribers, as well as developers via the OpenAI API and enterprise customers on Microsoft Azure and AWS Bedrock. OpenAI has priced API access at $10 per million input tokens and $50 per million output tokens, with usage included in existing subscription tiers. The announcement also highlights dedicated production misalignment monitoring and a Codex Auto-review safeguard as part of the model's alignment controls.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Agentic RAG Lets AI Systems Reason, Verify and Retry Before Answering

Traditional RAG pipelines follow a fixed retrieve-then-generate approach, which can fail when retrieved documents are irrelevant or insufficient to answer a query. Agentic RAG addresses this by enabling AI systems to reason about retrieval, verify answers, refine searches, and escalate to humans when confidence is low. A developer has published a set of open-source notebooks demonstrating various Agentic RAG patterns built using LangGraph. The notebooks are designed as self-contained, practical examples that developers can study and adapt for their own AI projects. A free version is available on GitHub, with an advanced paid version offered separately.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Guide: How Agent-to-Agent AI Marketplaces Work in 2026

By 2026, autonomous AI agents are expected to transact with each other programmatically through structured marketplaces, eliminating the need for human mediation. These agent-to-agent (A2A) marketplaces are built on composable layers covering agent discovery, service definitions, payments, and reputation management. Payments are handled via the x402 micropayment protocol using USDC on Base, an EVM-compatible Layer 2 blockchain, with transactions authenticated through ECDSA cryptographic signatures. Agents are identified through decentralized identifiers (DIDs) linked to on-chain registries and IPFS-based documents, enabling verifiable discovery and capability negotiation. The guide, targeting Node.js and Python developers, provides runnable code examples illustrating minimum viable integrations for both exposing and consuming agent services.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

kagent 0.10 Adds Audience-Scoped Agent Tokens to Curb Confused Deputy Risk

Kagent v0.10.0, released on September 4, 2025, introduces two environment variables—KAGENT_STS_RESOURCE and KAGENT_STS_AUDIENCE—that allow agent authentication tokens to be scoped to specific backends via RFC 8707 and RFC 8693 OAuth 2.0 Token Exchange. Both variables default to empty, meaning tokens remain unscoped unless explicitly configured, which silently leaves agents with credentials valid across all trusted backends. This creates a confused deputy risk: a single compromised token obtained for one tool, such as a GitHub issue reader, could also grant access to a sensitive backend like a payments server. The fix requires operators to set the variables, after which the backend can reject tokens whose audience claim does not match the intended resource. A public demo repository, themsquared/kagent-sts-audience, illustrates the difference in token behavior across all three configuration states without requiring a Kubernetes cluster or cloud account.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

MCP Agent Identity Roadmap: Only One of Four Specs Ready, and It Targets Humans

The Model Context Protocol roadmap, published on August 22, lists agent identity and enterprise security as priorities, outlining four components: DPoP, Workload Identity Federation, the ID-JAG grant, and standard token exchange. Of these, only one specification is currently stable — the enterprise-managed-authorization extension — but it authenticates human employees via browser-based IdP login, not autonomous agents. A draft client-credentials extension exists for machine-to-machine use, but it relies on pre-registered credentials stored where the agent can access them, a known security risk. The two proposals that would grant agents a native identity independent of human delegation remain open pull requests with no stable spec yet. Developers deploying agents in production this quarter must bridge the gap themselves at a different architectural layer, as the protocol's own identity infrastructure is still incomplete.

0 comments Read more at DEV Community