MCP-Universe benchmark shows GPT-5 fails more than half of real-world orchestration tasks

The adoption of interoperability requirements, such because the Mannequin Context Protocol (MCP), can present enterprises with insights into how brokers and fashions perform outdoors their walled confines. Nevertheless, many benchmarks fail to seize real-life interactions with MCP. 

Salesforce AI Analysis developed a brand new open-source benchmark it calls MCP-Universe, which goals to trace LLMs as these work together with MCP servers in the true world, arguing that it’ll paint a greater image of real-life and real-time interactions of fashions with instruments enterprises really use. In its preliminary testing, it discovered that fashions like OpenAI’s just lately released GPT-5 are sturdy, however nonetheless don’t carry out as effectively in real-life situations. 

“Current benchmarks predominantly give attention to remoted facets of LLM efficiency, comparable to instruction following, math reasoning, or perform calling, with out offering a complete evaluation of how fashions work together with real-world MCP servers throughout various situations,” Salesforce stated in a paper. 

MCP-Universe captures mannequin efficiency via device utilization, multi-turn device calls, lengthy context home windows and huge device areas. It’s grounded on present MCP servers with entry to precise information sources and environments. 

Junnan Li, director of AI analysis at Salesforce, instructed VentureBeat that many fashions “nonetheless face limitations that maintain them again on enterprise-grade duties.”

“Two of the largest are: Lengthy context challenges, fashions can lose monitor of knowledge or wrestle to motive constantly when dealing with very lengthy or complicated inputs,” Li stated. “And, Unknown device challenges, fashions typically aren’t in a position to seamlessly use unfamiliar instruments or techniques in the way in which people can adapt on the fly. This is the reason it’s essential to not take a DIY strategy with a single mannequin to energy brokers alone, however as a substitute, to depend on a platform that mixes information context, enhanced reasoning, and belief guardrails to actually meet the wants of enterprise AI.”

MCP-Universe joins different MCP-based proposed benchmarks, comparable to MCP-Radar from the College of Massachusetts Amherst and Xi’an Jiaotong College, in addition to the Beijing College of Posts and Telecommunications’ MCPWorld. It additionally builds on MCPEvals, which Salesforce launched in July, which focuses primarily on brokers. Li stated the largest distinction between MCP-Universe and MCPEvals is that the latter is evaluated with artificial duties. 

The way it works

MCP-Universe evaluates how effectively every mannequin performs a sequence of duties that mimic these undertaken by enterprises. Salesforce stated it designed MCP-Universe to embody six core domains utilized by enterprises: location navigation, repository administration, monetary evaluation, 3D design, browser automation and internet search. It accessed 11 MCP servers for a complete of 231 duties. 

• Location navigation focuses on geographic reasoning and the execution of spatial duties. The researchers tapped the Google Maps MCP server for this course of. 
• The repository administration area appears to be like at codebase operations and connects to the GitHub MCP to show model management instruments like repo search, concern monitoring and code modifying. 
• Monetary evaluation connects to the Yahoo Finance MCP server to guage quantitative reasoning and monetary market decision-making.
• 3D design evaluates the usage of computer-aided design instruments via the Blender MCP.
• Browser automation, related to Playwright’s MCP, exams browser interplay.
• The online looking area employs the Google Search MCP server and the Fetch MCP  to verify “open-domain info looking for” and is structured as a extra open-ended process. 

Salesforce stated that it needed to design new MCP duties that replicate actual use circumstances. For every area, they created 4 to 5 sorts of duties that the researchers suppose LLMs can simply full. For instance, the researchers assigned the fashions a purpose that concerned route planning, figuring out the optimum stops after which finding the vacation spot. 

Every mannequin is evaluated on how they accomplished the duties. Li and his staff opted to comply with an execution-based analysis paradigm fairly than the extra widespread LLM-as-a-judge system. The researchers famous the LLM-as-a-judge paradigm “just isn’t well-suited for our MCP-Universe state of affairs, since some duties are designed to make use of real-time information, whereas the information of the LLM decide is static.”

Salesforce researchers used three forms of evaluators: format evaluators to see if the brokers and fashions comply with format necessities, static evaluators to evaluate correctness over time and dynamic evaluators for fluctuating solutions like flight costs or GitHub points.

“MCP-Universe focuses on creating difficult real-world duties with execution-based evaluators, which may stress-test the agent in complicated situations. Moreover, MCP-Universe affords an extendable framework/codebase for constructing and evaluating brokers,” Li stated. 

Even the massive fashions have bother

To check MCP-Universe, Salesforce evaluated a number of common proprietary and open-source fashions. These embody Grok-4 from xAI, Anthropic’s Claude-4 Sonnet and Claude 3.7 Sonnet, OpenAI’s GPT-5, o4-mini, o3, GPT-4.1, GPT-4o, GPT-oss, Google’s Gemini 2.5 Professional and Gemini 2.5 Fkash, GLM-4.5 from Zai, Moonshot’s Kimi-K2, Qwen’s Qwen3 Coder and Qwen3-235B-A22B-Instruct-2507 and DeepSeek-V3-0304 from DeepSeek. Every mannequin examined had at the very least 120B parameters.

In its testing, Salesforce discovered GPT-5 had the perfect success charge, particularly for monetary evaluation duties. Grok-4 adopted, beating all of the fashions for browser automation, and Claude-4.0 Sonnet rounds out the highest three, though it didn’t put up any efficiency numbers increased than both of the fashions it follows. Amongst open-source fashions, GLM-4.5 carried out the perfect. 

Nevertheless, MCP-Universe confirmed the fashions had problem dealing with lengthy contexts, particularly for location navigation, browser automation and monetary evaluation, with effectivity falling considerably. The second the LLMs encounter unknown instruments, their efficiency additionally drops. The LLMs demonstrated problem in finishing greater than half of the duties that enterprises usually carry out.

“These findings spotlight that present frontier LLMs nonetheless fall brief in reliably executing duties throughout various real-world MCP duties. Our MCP-Universe benchmark, due to this fact, supplies a difficult and vital testbed for evaluating LLM efficiency in areas underserved by present benchmarks,” the paper stated. 

Li instructed VentureBeat that he hopes enterprises will use MCP-Universe to realize a deeper understanding of the place brokers and fashions fail on duties in order that they will enhance both their frameworks or the implementation of their MCP instruments.