Written from 63 named sources · Sep 14 · first result Cross-Vendor Terminology Audit: "Deep Research," "Agent," "Reasoning," and "Grounding" Author: AI Documentation Researcher & Terminology Analyst Date: September 14, 2026 Executive Summary & Stack Architecture The primary point of friction in contemporary AI systems engineering is not that vendors disagree on colloquial definitions, but that major AI providers (OpenAI, Google, Anthropic, Microsoft, Meta) operationalize the terms "deep research," "agent," "reasoning," and "grounding" across fundamentally different layers of the software and hardware stack [15]. A term that represents an infrastructural model primitive in one ecosystem is frequently branded as a managed catalog application or an evaluative metric in another [12, 14, 15]. Analytical Comparative Synthesis 2.1 Deep Research Explicit Definitions vs. Descriptive Usage OpenAI: Operates "deep research" simultaneously across three adjacent registers: Specialized Model Family: Standalone checkpoints such as o3-deep-research and o4-mini-deep-research designed to plan, browse, execute code, and synthesize hundreds of online sources into structured, citation-backed reports [16, 20]. Product Workflow: An interactive ChatGPT capability combining clarification, prompt rewriting, plan review, progress tracking, interruption, and autonomous investigation into cited reports [16, 18, 19]. Deeper Web-Search Mode: A distinct path within the standard Responses API web_search tool for GPT-5.5-class models, configured with high reasoning budgets (reasoning.effort: "high" or "xhigh") and, for long research runs, return_token_budget: "unlimited" [46]. Note that these parameters belong to the GPT-5.5 web-search path, not to the dedicated o3-deep-research model family, which is configured through model selection and its own tool set [16, 46]. Google: Formally defines "Gemini Deep Research" as a discrete, managed catalog Agent (e.g., deep-research-preview-04-2026) accessed through the Interactions API [42, 45]. It is explicitly executed as an asynchronous, multi-step background job (background=True) with native tool support for Google Search, URL Context, Code Execution, File Search, and Remote Model Context Protocol (MCP) servers [42]. Google's documentation is emphatic on the ontological point: "Deep Research is an agent, not just a model" [42, 45]. Microsoft: Inherits OpenAI's framing within Azure OpenAI Foundry, treating Deep Research as an "agent-driven mode" for extended investigations powered by o3-deep-research or gpt-5.5 at high/xhigh reasoning effort [48]. Anthropic & Meta: Maintain a documentation gap for "deep research" as a formal branded product or API primitive. Anthropic delivers comparable multi-step investigative capabilities through subagent fan-out and dynamic workflow scripts via the Claude Agent SDK [27], and its research-agent eval guidance treats "research agents" as an evaluative category rather than a named API product [28]. Meta shows no branded "deep research" API in its supplied documentation; the closest capabilities are Muse Spark's long-context agentic coding loops, tool calling, and search grounding [58, 59], though Meta does not market a research-analyst report product comparable to OpenAI's or Google's. Irreconcilable Conflicts The fundamental conflict is ontological: Model vs. Agent vs. Tool Mode. For Google, Deep Research is an instantiated, hosted agent entity with built-in execution state, invoked by agent ID and polled asynchronously [42, 45]. For OpenAI and Microsoft, it is simultaneously an inference model identifier (o3-deep-research) and an execution tier of the general web_search tool within the Responses API [16, 46, 48]. Anthropic and Meta do not brand it at all. This is a genuine packaging and ontology conflict, not merely a branding difference. Documented & Quiet Shifts OpenAI (Agent to Mode/Model Re-parenting): In its initial February 2025 release, OpenAI positioned Deep Research as "OpenAI's next agent" [17]. By July 2025, OpenAI stated that deeper browsing was available through ChatGPT "agent mode," while "the original deep research functionality remains available via the 'deep research' option in the tools menu" [17]. By February 2026, deep research could connect to any MCP or app, restrict web searches to trusted sites, and support real-time progress tracking and interruption [17]. This supports a quiet repositioning from standalone "next agent" toward the broader ChatGPT agent umbrella, without a formal deprecation of the original "agent" nomenclature in the supplied material. 2.2 Agent Explicit Definitions vs. Descriptive Usage Anthropic (The Strict Architectural Boundary): Anthropic provides the most formal and restrictive definition in the industry [26]. It establishes "agentic systems" as the umbrella term and draws an explicit architectural boundary: Workflows: Systems where LLMs and tools are orchestrated through predefined, deterministic code paths (e.g., prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer) [26]. Agents: Systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks [26]. OpenAI: Adopts an SDK-centric operational definition: an agent is "the core unit of an SDK-based workflow" packaging a base model, instructions, tools, guardrails, MCP servers, handoffs, and structured outputs into a managed execution runner [21]. Its Agents SDK guide frames agents as "applications that plan, call tools, collaborate across specialists, and keep enough state to complete multi-step work" [25]. Google: Defines agents as "systems that leverage Gemini models, a set of tools, and reasoning capabilities to perform complex, multi-step tasks and achieve specific goals" [44], and its managed-agent docs put that harness in an isolated Linux sandbox with code execution, file management, and web access [41, 43]. Meta: Defines agents operationally through the coding-agent loop: agents act as orchestrators that decompose instructions into tool calls—read/edit file, run command—and loop until completion [58]; its cookbook names the "basic agent loop" as perceive-decide-act [60]. Microsoft: Presents Foundry Agent Service agents as cloud-hosted entities connected to tools such as web grounding [53, 55]. The supplied pages strongly support managed web-grounded agents and their governance (RBAC role renames, admin controls) [53, 55]; the fuller packaged formulation including memory and compliance filters is only partially evidenced in the supplied extracts and should be treated as a reasonable but not fully documented characterization. Irreconcilable Conflicts The Autonomy Threshold Conflict (with qualification): A deterministic code pipeline chaining three LLM function calls is classified as a Workflow (and explicitly not an Agent) under Anthropic's taxonomy [26]. OpenAI and Google do not clearly state that such a fixed pipeline is an "agent"; their definitions consistently emphasize planning, tool use, reasoning, and goal-directed multi-step behavior rather than mere chaining [21, 44]. The conflict is therefore real in emphasis and normativity—Anthropic's boundary is explicit and prescriptive where the others' are operational and inclusive—but it is not as cleanly binary as a strict reading suggests. In practice, the same system built with OpenAI's Agents SDK or Google's managed-agent harness is packaged and billed as an "agent" regardless of whether its control flow is predetermined, which is precisely the outcome Anthropic's taxonomy forbids. Documented & Quiet Shifts Anthropic SDK Refactoring: Anthropic's agent tooling was rebranded in September 2025, with the SDK standardized around the Claude Agent SDK name and the Model Context Protocol (MCP); a third-party glossary corroborates the September 29, 2025 rename from Claude Code SDK [6], and current Anthropic documentation uses the Claude Agent SDK name throughout [26, 27]. OpenAI (Assistants API to Agents SDK / Responses API): OpenAI transitioned architectural guidance away from the stateful, server-managed Assistants API (now listed under "Legacy APIs" with a migration guide) toward the Agents SDK and the Responses API [25, 49, 61]. Note that the Responses API should not be described as "stateless": OpenAI documents stateful persistence via previous_response_id and Conversations, including response forking [47, 49, 61]. The distinction is better stated as: the Agents SDK runs the agent loop for you, while the Responses API is for when you want to own the loop yourself [25]. 2.3 Reasoning Explicit Definitions vs. Descriptive Usage OpenAI: Formally defines reasoning models as architectures that emit invisible internal reasoning tokens prior to generating user-visible output tokens; those tokens occupy context and are billed as output tokens [1, 23]. Reasoning compute is exposed as an adjustable parameter (reasoning.effort: none, minimal, low, medium, high, xhigh, max — seven values, with model-dependent subsets) and distinct operational execution modes (reasoning.mode: standard vs. pro on GPT-5.6) [23]. OpenAI documents reasoning.effort as guidance for how much the model thinks, with models reasoning adaptively and actual token spend remaining task-dependent and bounded by max_output_tokens [23]. Anthropic: Explicitly uses the term "Thinking" (extended thinking or adaptive thinking) rather than "reasoning models" [31, 32]. Thinking tokens are billed as output tokens, but the model output separates internal reasoning summaries (thinking content blocks) from the final response (text blocks) [31, 32]. Each thinking block carries an opaque encrypted signature — a continuity/integrity artifact passed back to the API in multi-turn and tool-use conversations, not a user-auditable proof of the raw chain of thought; no display setting returns the raw chain of thought, and the API explicitly rejects reasoning-extraction attempts on some models [32]. Anthropic strictly separates whether Claude thinks (thinking.type) from how much work Claude puts into the whole response (output_config.effort) [30, 32, 34]. Google: In the supplied Google pages, reasoning appears less as a standalone parameter page and more through "thought" steps with signatures and configurable thinking_summaries: "auto" for streaming intermediate reasoning in Deep Research and Interactions [36, 42]. Third-party comparison material describes Gemini "thinking models" using chain-of-thought [11], but the supplied Google pages support only the narrower claim: thought summaries can be streamed where enabled. Microsoft: Directly mirrors OpenAI's API schema, exposing reasoning.effort inside Azure OpenAI model deployments and using the same high/xhigh guidance for deeper search and research [48]. Meta: Defines reasoning as a parameterised inference state within Muse Spark and the Model API: reasoning: true, reasoningEffort including high/xhigh/max, reasoningSummary, and encrypted reasoning replay via include: ["reasoning.encrypted_content"] to preserve reasoning across turns [58]. Llama 4 is marketed with advanced reasoning and agentic capabilities [56]. Irreconcilable Conflicts Token Budget vs. Behavioral Steering: In Anthropic's system, effort is a behavioral steering knob governing overall response thoroughness across both thinking and output tokens, while token budgets (budget_tokens, max_tokens) are distinct controls [30, 31, 32]. In OpenAI's system, reasoning.effort guides how much the model thinks, with actual token volume remaining task-dependent and capped by max_output_tokens [23]. The conflict is parametric rather than ontological — both vendors expose graded effort controls — but the semantics of "effort" differ: response-level steering (Anthropic) versus thinking-depth guidance within a mode (OpenAI). Meta adds a third variant: encrypted reasoning continuity as a first-class round-trip payload [58]. Documented & Quiet Shifts Anthropic (Extended to Adaptive Thinking Transition): Anthropic deprecated manual extended thinking with fixed token allocations (thinking.type: "enabled" with budget_tokens) on Claude 4.6 models (requests still succeed) and hard-rejected it on Claude 4.7+ (returning an HTTP 400 error). It was replaced by Adaptive Thinking (thinking.type: "adaptive"), where the model dynamically decides whether and how much to reason per turn, with depth controlled via output_config.effort [31, 32]. OpenAI (Parameter Granularity Expansion): OpenAI's reasoning surface now includes a seven-value reasoning.effort hierarchy, independent reasoning.mode ("standard" vs. "pro"), mid-conversation configuration_update items that change effort while preserving the prompt-cache prefix, and reasoning.context controls (current_turn vs. all_turns) for cross-turn reasoning rendering [23]. (The historical "from a boolean flag" framing is not evidenced in the supplied page and is omitted here.) 2.4 Grounding Explicit Definitions vs. Descriptive Usage Microsoft: Maintains the most rigorous multidimensional taxonomy of grounding: Grounding Data (Structural): Contextual non-training data injected into the prompt at inference time via databases, search indexes, or vector stores [51]. Groundedness (Evaluative Quality): A measurable Azure AI Content Safety detection metric defining the degree to which an LLM's generated output faithfully adheres to the provided source material without fabrication, with optional automated correction [52, 54]. Managed Service: Commercial endpoints such as Grounding with Bing Search and Grounding with Bing Custom Search, plus the native Web Search tool in Foundry Agent Service [53, 55]. Google: Treats grounding primarily as a provider-managed platform capability ("Grounding with Google Search") [36, 37, 38]. The model decides when to execute search queries, returning an integrated response decorated with structured groundingMetadata (groundingChunks, groundingSupports, webSearchQueries, and inline citation offsets) in the generateContent API, or inline url_citation annotations in the Interactions API [36, 37, 40]. OpenAI: Avoids "grounding" as a formal API namespace. It operationalizes the concept through discrete tool primitives: web_search (retrieval over the live web) [46] and file_search (hosted vector stores and chunking retrieval) [50], returning citation annotations directly on the message payload [16, 46]. Anthropic: Omits "grounding" as a branded product term entirely. It discusses grounding functionally as obtaining "ground truth from the environment" via tool execution at each step of an agent's work [26], and separately documents Citations as a feature that grounds responses in source documents [30]. Meta: Employs "search grounding" as a listed model capability within the Meta Model API for Muse Spark [58, 59]. The supplied Meta pages support the capability flag and MCP/tool connectivity; a formal architectural pairing with RAG is not stated in the supplied extracts and should not be asserted. Irreconcilable Conflicts Product Pipeline vs. Data Layer vs. Verification Metric: Grounding is simultaneously used to describe: A proprietary, provider-managed search-and-generation pipeline whose retrieval, cost, latency, and citation behavior are bundled and largely opaque (Google; cf. the "native search grounding" critique [14]) [36, 37]. Non-training data injected at inference time (Microsoft) [51]. A safety and quality evaluation metric measuring output faithfulness against source tokens, with optional correction (Microsoft Content Safety) [52, 54]. Environmental feedback and document citations (Anthropic) [26, 30]. Discrete tool primitives with per-call audit records (OpenAI) [46, 50]. A model capability flag (Meta) [58, 59]. These are different layers of the stack, not interchangeable definitions of one thing. Documented & Quiet Shifts Google (Tool Evolution & Billing Metric Shift): Google's current pages use the google_search tool identifier, while the legacy google_search_retrieval identifier appears in migration notes on the same pages; the supplied extracts list both identifiers but do not explicitly narrate a formal deprecation, so the transition should be characterized as an identifier migration evidenced by the docs rather than a documented deprecation statement [36, 37]. Far more explicitly documented is the billing change: for Gemini 3 models, projects are billed for each search query the model generates, whereas Gemini 2.5 and older remain billed per prompt — a quiet but consequential shift that ties cost to iterative model reasoning [36, 37, 38]. Microsoft (Tool Consolidation): Microsoft promoted native Web Search to general availability (GA) in Foundry Agent Service, superseding the requirement to manually provision separate Bing Search resources for basic web grounding; Grounding with Bing Search remains a GA option with more parameters and broader model support [53]. Separately, Foundry RBAC roles (Foundry User, Foundry Owner, etc.) were recently renamed from their Azure AI equivalents — a governance-level terminology shift with unchanged permissions [53]. Cross-Vendor Comparison Matrix Term Vendor Current Formal / Operational Definition Primary Divergence Points Revision / Change Status Deep Research OpenAI Specialized model family (o3-deep-research, o4-mini-deep-research) plus a ChatGPT workflow; separately, a deeper GPT-5.5 web-search path at high/xhigh effort with optional return_token_budget: "unlimited" [16, 18, 46]. Positioned as a model family and product workflow plus a parameterised search tier, rather than an isolated agent catalog entity [16, 46]. Quiet Shift: Re-positioned from "OpenAI's next agent" (Feb 2025) toward the ChatGPT agent umbrella, with expanded MCP/site controls and interruption by Feb 2026 [17]. Google Named managed agent (deep-research-preview-04-2026) executing asynchronous planning, browsing, coding, and report generation; "an agent, not just a model" [42, 45]. Built as an asynchronous, hosted catalog agent requiring background execution and polling [42, 45]. Active Preview: Maintained as a first-class agent catalog resource [42]. Microsoft Agent-driven mode within Azure OpenAI Foundry inheriting o3-deep-research and high/xhigh-reasoning gpt-5.5 [48]. Treated as a search mode configuration mirroring OpenAI's API parameters [48]. Stable: Aligned with Azure OpenAI Responses API updates [48]. Anthropic No formal branded primitive. Implemented via subagent orchestration and dynamic workflow scripts [27]. Rejects single-term branding; treats deep investigation as a structural workflow pattern and an eval category [26, 27, 28]. No Primitive: Capability handled via Claude Agent SDK [27]. Meta No formal branded primitive. Closest capabilities are Muse Spark long-context agentic loops and search grounding [58, 59]. Addressed through general model long-context reasoning rather than branded search modes; no research-analyst product documented [58, 59]. No Primitive: Addressed through the Muse model line [58, 59]. Agent Anthropic Systems where LLMs dynamically direct their own processes and tool usage, explicitly distinguished from deterministic workflows [26]. Strictest taxonomy; explicitly excludes predefined code paths/chains from the definition of an agent [26]. Formalised: "Workflows vs. Agents" framework remains the architectural benchmark [26]. OpenAI The core unit of an SDK-based workflow packaging a model, instructions, tools, guardrails, MCP servers, handoffs, and structured outputs [21]. SDK configuration wrapper emphasizing planning, tools, and multi-step goal pursuit [21, 25]. Shift: Migration from Assistants API (legacy) to Agents SDK and the stateful Responses API with previous_response_id/Conversations [25, 47, 49, 61]. Google Systems combining Gemini models, tools, and reasoning to execute multi-step tasks within managed Linux-sandbox harnesses (Antigravity runtime) [41, 43, 44]. Centred on isolated environment runtimes and managed sandbox infrastructure [41, 43]. Evolution: Consolidated around the Gemini Agents API and the Antigravity harness [41, 44]. Microsoft Hosted execution entities within Foundry Agent Service connected to tools such as web grounding, with renamed RBAC governance [53, 55]. Focuses on enterprise infrastructure, data boundary management, and RBAC governance; full packaged definition (memory, compliance filters) only partially evidenced in supplied extracts [53]. Refactored: Roles renamed within Foundry Agent Service; Web Search promoted to GA [53]. Meta Multi-step perceive-decide-act loops driving workspace file edits, shell commands, and MCP tools via Muse Spark [58, 60]. Focuses on coding-agent orchestration and open model harnesses [58, 60]. Expanded: Standardised around Muse Spark agentic patterns [58, 60]. Reasoning OpenAI Latent inference-time compute using invisible reasoning tokens, guided by a seven-value reasoning.effort ladder and standard/pro modes; effort is guidance, with spend bounded by max_output_tokens [23]. Governed by graded effort levels, modes, mid-conversation configuration_update, and reasoning.context controls [23]. Expanded: Broadened to granular effort levels, mode controls, and cross-turn reasoning rendering [23]. Anthropic Step-by-step internal reasoning termed "Thinking" (adaptive by default on latest models), returning summaries and opaque encrypted signature fields for continuity — not user-auditable proofs of raw chain of thought [31, 32]. Distinguishes thinking modes from response-level effort; rejects raw chain-of-thought extraction [30, 31, 32]. Documented Shift: Deprecated manual budget_tokens on 4.6 (still succeeds); hard-rejected on 4.7+ in favour of Adaptive Thinking [31]. Google "Thought" steps with signatures and streamable thinking_summaries in agent interactions, where enabled [36, 42]. Exposed as streamable thought summaries and thought/signature steps rather than a standalone effort parameter page in the supplied docs [36, 42]. Shift: Integrated thinking-summary streaming across Deep Research and Interactions [42]. Microsoft Latent reasoning tokens configured via reasoning.effort inside Azure OpenAI deployments [48]. Direct mirror of OpenAI reasoning parameters in Azure infrastructure [48]. Aligned: Synchronised with OpenAI model releases [48]. Meta Multi-step reasoning with reasoning: true, effort levels, and encrypted cross-turn reasoning replay via include: ["reasoning.encrypted_content"] in Muse Spark [56, 58]. Focuses on cross-turn reasoning preservation via encrypted payload round-tripping [58]. Enhanced: Added encrypted reasoning continuation across multi-turn sessions [58]. Grounding Microsoft 1. Grounding data (non-training inference data) [51]. 2. Groundedness (evaluation metric with optional correction) [52, 54]. 3. Managed Bing tools plus GA native Web Search [53, 55]. Clearest separation between structural data injection, output evaluation, and managed search tools [51, 52, 53]. Shift: Native Web Search promoted to GA, superseding manual Bing resource setups; RBAC roles renamed [53]. Google Managed provider tool ("Grounding with Google Search") returning structured citations and metadata [36, 37, 38]. Integrates retrieval, generation, and citation into a provider-managed pipeline [36, 37]. Quiet Shift: Tool identifier migrated from google_search_retrieval to google_search (deprecation not explicitly narrated in supplied extracts); billing shifted from per-prompt to per-generated-search-query on Gemini 3 [36, 37, 38]. OpenAI No formal namespace. Delivered via discrete web_search and file_search tool citations [46, 50]. Avoids "grounding" branding; decomposes functionality into independent search and retrieval tools [46, 50]. Evolution: Replaced preview search models with Responses API web_search and hosted vector stores [46, 50]. Anthropic No formal branded primitive. Addressed as obtaining "ground truth from the environment" via tools, plus Citatio