Photo by Bridge for Billions on Unsplash
- As of July 8, 2026, Google has expanded Gemini API Managed Agents with four concrete additions: asynchronous background execution, remote MCP server integration, custom function calling, and credential refresh across long-running sessions.
- The Interactions API reached General Availability in 2026 and is now Google's recommended interface for all new agent projects — a signal that the platform is considered production-ready.
- Gemini 3.5 Flash processes 289 tokens per second — 4x faster than Claude Opus 4.7 and GPT-5.5 — priced at $1.50 per million input tokens, $9.00 per million output tokens, and $0.15 per million cached input tokens (a 90% discount).
- The Model Context Protocol has crossed 10,000 active public servers and 97 million or more monthly SDK downloads as of 2026, making it the de facto standard for connecting AI agents to external tools.
What Happened
10,000. That's the number of active public MCP servers running as of 2026 — a figure that would have read like speculation eighteen months ago. That ecosystem momentum forms the backdrop for Google's June 2026 expansion of Managed Agents in the Gemini API, covered by Google News and detailed on the official Google Developer Blog.
According to the Google Developer Blog, the update delivered four specific additions to the Managed Agents surface: background execution for async long-running tasks, remote MCP server integration for securely connecting agents to external tools, custom function calling at the managed agent layer, and credential refresh that persists authentication across multi-step interactions. On June 30, 2026, Google Cloud separately launched a fully managed remote MCP server for its Gemini Enterprise Agent Platform, giving enterprise customers a direct, secure path for external AI agents to reach Google Cloud resources without custom infrastructure.
The Interactions API — which underpins the new async execution capability — reached General Availability in 2026 and is now Google's recommended interface for all new Gemini agent projects. That shift from experimental to GA status is the quiet headline here: it signals that Google considers the managed agent runtime stable enough for production commitments, something enterprise procurement teams have been waiting to confirm before standardizing on the platform.
The Workflow Pain These Features Actually Solve
The problem is familiar to any team that has tried to run agents past a synchronous HTTP timeout. Enterprise use cases — autonomous market research for financial services, multi-document legal review, long-running quantitative analysis — routinely outlast the response windows that synchronous API calls can hold open. Teams building AI investing tools in fintech face this most acutely: an agent performing multi-step portfolio analysis or scraping live data feeds across dozens of sources can't complete inside thirty seconds. Background execution addresses this directly by letting agents run as asynchronous tasks that fire a callback when results are ready, rather than holding a live connection.
Remote MCP integration is the other headline. The Model Context Protocol — described in its own documentation as "USB-C for AI" — had already become a de facto standard before Google's June update. Anthropic originally developed MCP and donated it to the Linux Foundation's Agentic AI Foundation in December 2025, with OpenAI, Google, Microsoft, AWS, Cloudflare, GitHub, and Bloomberg joining as supporting members. OpenAI adopted MCP across its products including the ChatGPT desktop app in March 2025; Google DeepMind confirmed support in April 2025. As of May 24, 2026, GitHub reported 15,926 repositories tagged with the mcp-server topic, and the protocol had accumulated 97 million or more monthly SDK downloads across Python and TypeScript.
This ecosystem context explains why Google's managed remote MCP server isn't just a feature checkbox — it's plugging into an already-mature tooling layer. As AI Agents NewLens has documented, MCP is reshaping entire workflow categories, from content management to site automation. The enterprise adoption numbers back this up: as of Q1 2025, 28% of Fortune 500 companies had MCP servers in production, and according to Stacklok's 2026 report, 41% of surveyed software organizations had achieved either limited or broad MCP production deployment.
The Google Developer Blog frames the Managed Agents API as collapsing complex infrastructure into a single API call that spins up a secure, ephemeral Linux sandbox hosted by Google. That is accurate — and it competes most directly with OpenAI's Assistants API. CloudInsight's API comparison analysis describes OpenAI's offering as "more mature with Threads, tool use, code interpreter, and file search as a managed agent runtime," while noting that Gemini's equivalent features "exist but are newer and less battle-tested." That maturity gap is real, but it's narrowing fast.
Photo by Arnold Francisca on Unsplash
The Pricing Reality Nobody Markets
Gemini 3.5 Flash — the model powering the managed agent platform, announced at Google I/O 2026 on May 19–20, 2026 — benchmarks at 76.2% on Terminal-Bench 2.1, 1656 Elo on GDPval-AA, and 83.6% on MCP Atlas. It processes 289 tokens per second, 4x faster than Claude Opus 4.7 and GPT-5.5. The context window sits at 2 million tokens versus GPT-5's 400K, with a maximum output of 65,536 tokens per call. But the number that will drive financial planning conversations in procurement meetings is the cost structure.
Chart: Gemini 3.5 Flash API pricing as of July 8, 2026 — $1.50/M input tokens, $9.00/M output tokens, $0.15/M cached input (90% discount). Source: Google Developer Blog. Note: bar heights are proportional to $9.00 scale; cached input bar shown at minimum visibility.
The headline rates are $1.50 per million input tokens and $9.00 per million output tokens. Cached input drops to $0.15 per million — a 90% reduction. Research broadly indicates Gemini is 30–60% cheaper than OpenAI at equivalent model tiers. In my analysis, the cached-input discount is the cost lever that will actually move enterprise procurement decisions far more than any benchmark score. A knowledge retrieval workflow that repeatedly hits the same indexed documents runs nearly free on the input side at $0.15 per million tokens.
The real limit nobody markets: output tokens. Agentic loops generate substantial intermediate output — reasoning steps, tool call payloads, intermediate summaries — and those all accumulate at $9.00 per million. Teams should model their output-to-input ratios carefully before assuming the headline input price reflects their actual bill. The 2-million-token context window is genuinely differentiating for document-heavy workflows, but context window size and per-token cost are separate considerations that often get conflated in platform comparisons.
Three Steps for Teams Evaluating the Update
Background execution solves a specific problem: tasks that routinely exceed synchronous API timeout windows, or that require sequential tool calls with waits between them. Audit your current agent architecture for any workflow taking longer than 30 seconds end-to-end, or any multi-step pipeline where intermediate steps involve external API waits. These are the clearest candidates for migration to the Interactions API's async model. If all your agents complete in under ten seconds, background execution won't change your architecture today — and that's fine.
As of July 8, 2026, the MCP ecosystem hosts 10,000-plus active public servers. Before building a custom integration for an internal data source, verify whether a public or enterprise MCP server already covers your tool category. Google Cloud's fully managed remote MCP server, launched June 30, 2026, is the natural starting point for GCP-heavy teams — it handles authentication and secure connectivity to Google Cloud resources without custom infrastructure. For non-GCP internal systems, Stacklok's 2026 finding that 41% of software organizations already have some MCP production deployment suggests your infrastructure team may already have servers ready to connect.
Gemini 3.5 Flash's $1.50/M input rate and 90% cached-input discount look compelling in isolation. But agent workflows generate output tokens — reasoning steps, tool payloads, response synthesis — that accumulate at $9.00/M. Run a simulation: estimate average output tokens per agent session, multiply by expected session volume per month, and compare that number against OpenAI Assistants API pricing and Anthropic's tool-use tiers before finalizing vendor selection. The 2-million-token context advantage is real for document-heavy use cases, but only if your output generation costs don't offset the input savings.
Frequently Asked Questions
What is the difference between Gemini Managed Agents API and OpenAI Assistants API for enterprise use?
Both offer hosted agent runtimes handling tool use, persistent context, and file access. As of July 8, 2026, OpenAI's Assistants API is the more mature option — its Threads architecture, code interpreter, and file search capabilities have been in production longer and have more documented production deployments, according to CloudInsight's API comparison analysis. Google's Managed Agents surface introduces background execution and native remote MCP server support as genuine differentiators that OpenAI Assistants does not offer natively. Gemini's 2-million-token context window versus GPT-5's 400K is a meaningful architectural gap for document-heavy workflows, though teams should evaluate output token cost structures on both platforms before deciding.
Is the Gemini API free, and what does Gemini 3.5 Flash actually cost?
As of July 8, 2026, Google provides a free tier for Gemini API access with rate limits. Paid access to Gemini 3.5 Flash is priced at $1.50 per million input tokens and $9.00 per million output tokens, according to the Google Developer Blog. Cached input tokens receive a 90% discount at $0.15 per million tokens. Research broadly indicates Gemini is 30–60% cheaper than OpenAI at equivalent model tiers, though agentic workflows that generate heavy output token volume may see that discount erode in practice. Teams should model their specific output-to-input ratios before comparing sticker prices.
What is MCP and why does it matter for AI agent workflows in 2026?
The Model Context Protocol is an open standard for connecting AI agents to external data sources, databases, and tools. Originally developed by Anthropic and donated to the Linux Foundation's Agentic AI Foundation in December 2025 — with OpenAI, Google, Microsoft, AWS, Cloudflare, GitHub, and Bloomberg as supporting members — MCP has been adopted across the major AI platforms. As of May 24, 2026, GitHub reported 15,926 repositories tagged with the mcp-server topic, and the protocol had accumulated 97 million or more monthly SDK downloads. For AI agent workflows, MCP standardizes tool discovery and invocation, eliminating the need for custom connectors for each external integration. Gemini's June 2026 update adds native managed remote MCP server support, making it the first major managed agent platform to offer this natively.
Disclaimer: This article is editorial commentary for informational purposes only and does not constitute financial, investment, or technical advice. Research based on publicly available sources current as of July 8, 2026. The author may receive compensation through affiliate partnerships disclosed on this site.