Microsoft Foundry adds GPT-6.1 Sol, makes Routines and Tool Search generally available

GPT-6 lineup grows to three generally available models

According to Microsoft, GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna are all generally available in Foundry. GPT-6 Astra was released on September 3, GPT-6 Luna on September 22 and GPT-6.1 Sol on September 29.

Microsoft describes Astra as the most capable model for the most demanding work, Sol as offering near-Astra performance for complex work at a lower cost, and Luna as the most efficient model for focused, high-volume tasks. The September-announced prices for Global Standard short-context requests, per 1M tokens, are:

• GPT-6 Astra: $10.00 input, $50.00 output, $1.00 cached input
• GPT-6.1 Sol: $2.00 input, $10.00 output, $0.10 cached input
• GPT-6 Luna: $0.10 input, $0.50 output, $0.01 cached input

Long-context rates are higher, the post says. Microsoft states that the GPT-6.1 Sol launch reports better coding, computer use and professional work than GPT-6 Sol at the same uncached-input and output prices, with 50% lower cached-input pricing.

All three models have a 1.05M-token context window and up to 128K output tokens, including reasoning, within one shared budget. They support Responses, Chat Completions, streaming, function calling, structured outputs, and image input with text output. Knowledge cutoffs are April 2026 for Astra and GPT-6.1 Sol, and May 2026 for Luna. GPT-6.1 Sol tool calls require Responses.

Other models and Model Router controls

September also brought announcements for Claude and Microsoft's MAI models. Foundry availability for Claude Opus 5.5 was announced on September 24, and Microsoft advises checking the hosting option and API before deployment. Users can choose between Azure-hosted and Anthropic-hosted options, and should not assume OpenAI client parameters transfer unchanged. MAI-Image-2.6 and MAI-Image-2.6-Flash are in public preview as of September 4, and MAI-Transcribe-2 has Foundry demo availability confirmed from September 3.

Model Router gets two preview controls: session affinity for related Chat Completions requests, and request-level metadata for routing attempts and fallback. Session affinity asks the router to prefer the same eligible model across related requests, which can help with prompt-cache reuse. Microsoft says neither a cache hit nor the same model is guaranteed, because safety, capability, quota, availability and fallback requirements still apply. Model Router also expands to 32 regions, adding Canada Central, North Europe, Norway East and UAE North.

Agents: Routines, Tool Search and Insights

Routines are now generally available. They let teams run agents on a timer, a schedule, or a supported GitHub or Teams event. The agent-scheduled reminder tool remains in preview.

Tool Search and Agent-to-Agent (A2A) are also GA. Tool Search loads tool definitions on demand rather than all upfront, which leaves more context for other material and can potentially reduce input-token costs. Hosted-agent Toolboxes remain GA, while prompt-agent Toolboxes enter public preview. Voice agents and long-running hosted-agent resilience are in public preview as well.

Insights enters public preview. It surfaces recurring issues in an agent's traces, such as repeated retrieval or lost context, and shows supporting evidence and recommended next steps so developers can investigate before changing an agent.

The post illustrates these features with a developer changelog workflow built from two hosted agents, one using the GitHub Copilot SDK and one using Microsoft Agent Framework. A Monday routine starts the research, and a second agent delivers the checked report. Microsoft recommends choosing the external protocol (Responses, Invocations or Activity) based on the caller's contract rather than the framework.

Local and developer tooling

Foundry Local 2.1 adds bring-your-own-model support, allowing compatible ONNX Runtime GenAI models to be registered without copying files. Content Understanding adds preview tooling. On the developer side, Dev Pack launches an installer, and Foundry Toolkit adds templates, skill-only Toolboxes and controls for unfinished Responses work.

For teams building on Foundry, the update moves scheduling and on-demand tool loading into production-ready territory, while trace-based analysis and routing visibility remain in preview. Microsoft's guidance throughout is to keep test inputs fixed when comparing models and to check each model's API and availability limits before deploying.