OpenAI · Batch, Tools & Agents · 2026

OpenAI 2026 — Batch, computer_use, deep_research & MCP.

Fifteen deep-dive guides on OpenAI's async and agentic capabilities. Batch API for 50%-off nightly workloads, computer_use for browser and desktop automation, deep_research for multi-source cited reports, and MCP tools on Responses API to bring your own data and tools. Every page has tested Python code and the fixes that solve 90% of failures.

15Deep-dive error pages

4Advanced capabilities covered

50%Batch discount stacked with cache

100%2026-current code

Install & verify

All code samples target OpenAI Python SDK 1.66+ with Python 3.11+ for full support of Responses, MCP, computer_use, and deep_research.

setup.shbash

pip install --upgrade openai
# Version 1.66+ recommended for Responses, Conversations, computer_use, MCP support
python -c "import openai; print(openai.__version__)"

Batch API — 50% cheaper async workloads

The Batch API halves per-token cost for non-realtime workloads: classification, embeddings, offline analysis. Four pages cover the JSONL envelope, lifecycle polling, output reconciliation, and stacking the batch discount with Prompt Caching for 25% of sync cost.

Error #151

JSONL input format & per-request validation

Envelope shape, custom_id uniqueness, endpoint uniformity, per-line validation traps.

Error #152

Status polling, 24h expiration, cancellation

Six-status lifecycle, expired-batch recovery, cancellation semantics, orphan-batch reconciliation.

Error #153

Output parsing & partial failures (error_file_id)

output_file_id vs error_file_id, full custom_id reconciliation, classifying retryable errors.

Error #154

Enqueued token limits, rate limits & cost optimization

Queue-depth caps, 50% batch discount stacking with Prompt Cache for 25% of sync cost, reasoning-token budgeting.

computer_use — browser & desktop agents

computer_use turns Responses API into a browser or desktop agent. Four pages: correct tool configuration, the action loop with GA batched actions, the safety-check acknowledgment flow, and long-horizon context management for tasks with many iterations.

Error #155

Tool configuration (display, environment, model)

Preview vs GA shape, dimension alignment, environment selection, task-shaped input.

Error #156

Action loop (batched actions, screenshots)

GA batched actions[] vs legacy single action, computer_call_output shape, chaining via previous_response_id.

Error #157

pending_safety_checks acknowledgment flow

malicious_instructions / irrelevant_domain / sensitive_domain routing, human-in-loop patterns.

Error #158

Long-horizon context (screenshot inflation)

truncation="auto", session rotation, sub-task decomposition, screenshot compression.

deep_research — multi-source research reports

Deep research runs for 5-30 minutes, calls many sources, and produces structured cited reports. Four pages: model routing and the mandatory data source, MCP servers with the specific search/fetch interface, background job patterns, and citation extraction.

Error #159

Model routing & required data source

o3-deep-research vs o4-mini-deep-research, mandatory tools=[], timeout tuning, escalation pattern.

Error #160

MCP servers with search/fetch interface

The specific search/fetch tool contract, Azure allowed_domains, auth headers, background+MCP caveat.

Error #161

Background jobs, webhooks & long polling

background=true submission, poll intervals, webhook delivery, crash-safe resume, cost-gated cancel.

Error #162

Citation extraction & output structure

url_citation annotations, output item traversal, citation verification, preserving full response JSON.

MCP on Responses — bring your own tools

Model Context Protocol tools give Responses access to your APIs, databases, and third-party services. Three pages: declaration and caching mechanics, the approval flow for human-in-loop workflows, and failure/fallback patterns.

Error #163

MCP tool declaration (allowed_domains, headers, list_tools caching)

server_url vs connector_id, header stripping, mcp_list_tools caching via previous_response_id.

Error #164

Approval flow (mcp_approval_request human-in-loop)

require_approval per-tool policy, mcp_approval_request → mcp_approval_response chain, granular gates.

Error #165

MCP tool failures & fallback patterns

mcp_call.error monitoring, circuit breakers, degraded-mode operation, tool cache refresh on schema drift.

Quick reference — where to look first

If you see...	Start here
`files uploaded for use with batch must have purpose 'batch'`	#151 · Batch JSONL format
Batch stuck in `validating` for hours	#152 · Batch polling & expiration
Missing custom_ids after "successful" batch	#153 · Batch output parsing
Batch cost 10× estimate on reasoning models	#154 · Batch cost optimization
`Tool 'computer' is not supported for model 'gpt-4.1'`	#155 · computer_use setup
Model stuck asking for same action repeatedly	#156 · Action loop
`missing acknowledged_safety_checks`	#157 · Safety checks
Context length exceeded on long computer_use task	#158 · Context management
`deep_research models require at least one data source`	#159 · deep_research setup
Model can't find anything from your MCP	#160 · MCP search/fetch interface
Deep research timing out on sync call	#161 · Background jobs
Citations missing from output_text	#162 · Citation extraction
`MCP server url is not allowed` (Azure)	#163 · MCP declaration
Loop stalls with mcp_approval_request in output	#164 · Approval flow
Sustained MCP failures degrading output silently	#165 · MCP fallbacks