GPT-6.1 Sol lands in Microsoft Foundry, as Routines and Tool Search go GA

Microsoft’s September 2026 Foundry roundup makes Routines and Tool Search generally available, and adds GPT-6.1 Sol, a mid-tier OpenAI model priced at $2 per million input tokens, to the platform’s model lineup.

What is new in Foundry

The GPT-6 family now has three members in Foundry: GPT-6 Astra (released September 3), GPT-6 Luna (September 22) and GPT-6.1 Sol (September 29). All three are generally available. Microsoft describes Astra as the most capable model, Sol as near-Astra performance at lower cost, and Luna as the most efficient option for focused, high-volume tasks.

Prices listed for Global Standard short-context requests, per million tokens, are:

• 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.
• Luna: $0.10 input, $0.50 output, $0.01 cached input.

Long-context rates are higher. Each model has a 1.05M-token context window and up to 128K output tokens, shared with reasoning. Knowledge cutoffs are April 2026 for Astra and Sol, and May 2026 for Luna. Sol tool calls require the Responses API. According to the post, GPT-6.1 Sol improves on its predecessor GPT-6 Sol for coding, computer use and professional tasks, with unchanged standard input and output rates and a 50% cut for cached input.

Beyond OpenAI, the roundup notes Claude Opus 5.5 availability in Foundry (announced September 24), public previews of MAI-Image-2.6 and MAI-Image-2.6-Flash, and MAI-Transcribe-2 demo availability. Microsoft cautions that OpenAI client parameters should not be assumed to carry over to Claude.

Agent tooling moves toward production

Routines, which start one agent from a timer, recurring schedule, GitHub issue event or supported Teams channel event, are now generally available. Their configuration, identity and run history stay with the Foundry project. Microsoft stresses that Routines do not replace multi-agent orchestration or approval workflows, and that the agent-scheduled reminder tool is still in preview.

Tool Search and Agent-to-Agent (A2A) are also generally available. Instead of loading every tool definition into each request, a Toolbox exposes discovery and invocation through tool_search and call_tool. Microsoft says this leaves more context for the task and could reduce input-token costs, but it also states that its own demonstration proves discovery and does not measure token savings. Toolboxes for prompt agents enter public preview; hosted-agent Toolboxes were already generally available.

Insights enters public preview. It groups recurring behavior in agent traces into findings with supporting evidence, likely causes and proposed fixes. In Microsoft’s example, it flagged sample-heavy prompts, large prompts resubmitted after rate-limit errors, and an earlier unconfirmed delivery receipt. Voice agents and long-running hosted-agent resilience are also in public preview.

Model Router gains preview session affinity for related Chat Completions requests, plus routing metadata that shows selection and fallback. Microsoft says affinity can help prompt-cache reuse but guarantees neither a cache hit nor the same model. The router also expands to 32 regions, adding Canada Central, North Europe, Norway East and UAE North.

Local, speech and developer updates

Foundry Local 2.1, published September 30, lets developers register compatible ONNX Runtime GenAI models without copying files, with support across C++, C#, Python, JavaScript/TypeScript and Rust. Microsoft adds that compatibility is not guaranteed for every model format.

On speech, an update called Speech LLM 2607 improves multilingual and mixed-language recognition and adds a phrase-list field for domain terminology. Content Understanding gets a preview toolkit and command-line tool. Foundry Toolkit shipped four releases in September, and the Dev Pack now has an installer. SDK releases span Python, JavaScript/TypeScript, .NET and Java, with several migration notes. According to the post, Azure Developer CLI extension azure.ai.agents beta.18 rejects legacy manifest discovery and requires azd 1.34.2 or later.

On pricing, a deployment policy effective September 1 applies Standard premiums only to models launched on or after that date. Microsoft lists two September retirement dates: MAI-Transcribe-1 version 2026-01-23 on September 15, and the gpt-chat-latest snapshot 2026-06-24 on September 24.

Why it matters

The price list is the clearest signal, and it looks sharper against the market than against Microsoft’s own lineup. Sol’s $2 input and $10 output rates are one-fifth of Astra’s. According to eesel’s pricing breakdown, they are also half of the $4 and $20 charged for the previous mid-tier model, GPT-5.6 Sol. The same analysis puts the earlier GPT-6 Sol cached-input rate at $0.20, which fits Microsoft’s statement that the new version halves it to $0.10. Cached input matters most for agent loops, which resend long tool and instruction prefixes on every turn.

The competing option in Foundry is priced higher. Eesel’s Opus 5.5 guide, which cites Anthropic’s pricing documentation, lists the model at $4 per million input tokens and $20 per million output, twice Sol on both. The comparison is rough, since the two models are not equivalent in capability, and these are secondhand figures that were not checked against Anthropic’s own page.

One caveat applies to the headline rate. Eesel reports that GPT-6 Sol prompts above 272K input tokens are billed at a higher long-context rate of $4 input and $15 output. Microsoft says its long-context rates are higher too, and each model has a 1.05M-token window. Agents that load large documents or long histories may therefore pay more than the $2 figure suggests. The same source cites Artificial Analysis as finding the earlier Sol roughly level with GPT-5.6 Sol on its Intelligence Index, so cost, not capability, is the main change there. Whether 6.1 shifts that depends on benchmarks that Microsoft’s post does not supply.

What’s next

The September platform announcement also targeted rubric evaluation, synthetic and trace-derived dataset generation, and Agent Optimizer general availability for later in the month. The roundup says these remain announced targets rather than confirmed releases.