What are OpenAI Files, Codex, and Claude Files in June 2026?

Three different AI file-and-code surfaces, all maturing fast in mid-2026. (1) OpenAI File Search is the built-in document RAG tool inside OpenAI Agents SDK and the Assistants API — you upload files, OpenAI handles chunking, embedding, and retrieval. (2) Codex is OpenAI's GitHub-integrated coding agent, available via Codex CLI and the ChatGPT Codex web surface, which reads and writes against your repositories directly. (3) Claude Files is Anthropic's document-handling capability inside Claude.ai and the Claude API — file upload, citation, and retrieval across uploaded knowledge bases. The three are not direct replacements for each other; they target different surfaces (general agent file search, GitHub-resident code, conversational document QA) and overlap only at the edges.

Which is best for general document RAG: OpenAI File Search or Claude Files?

OpenAI File Search is the more production-tested option for general document RAG inside agentic workloads. It is built into the OpenAI Agents SDK and the Assistants API, handles chunking and embeddings automatically, and integrates with the rest of the OpenAI tool ecosystem (Operator, Code Interpreter, function calling). Claude Files is excellent for conversational document QA in Claude.ai and is integrated with Claude's strong long-context handling and citations, but it is less production-feature-rich than OpenAI File Search for SDK-driven agents. Pick OpenAI File Search if you're building an agent product. Pick Claude Files if conversational document QA at the chat UI level is the primary surface, or if you specifically need Claude's citation quality.

When should I use Codex instead of OpenAI File Search for code?

Use Codex when your codebase lives on GitHub and you want an agent that operates directly against the repository — clone, read, edit, commit, PR. OpenAI File Search is for uploading documents (including source files) into a separate index. Codex is for operating on the live repo. Practically: Codex via the ChatGPT Codex surface and Codex CLI is the right pick for AI-assisted development workflows where the unit of work is a PR. OpenAI File Search is the right pick when you want to ask questions across a snapshot of code plus documentation in one agent context. For the long tail of build-time and CI-time code-search workloads, Codex is materially more capable in mid-2026 than File Search.

How does Anthropic's Claude Code fit alongside Claude Files?

Different surfaces. Claude Code is Anthropic's terminal-and-IDE-resident coding agent, similar to Codex CLI — it reads and writes against your local repo and has rich sub-agent and skill support via the Claude Agent SDK. Claude Files is the document-handling capability inside Claude.ai (the chat product) and the Claude API for general file QA and citation. The two complement each other: Claude Code for engineering workflows, Claude Files for everything else. Note that on June 15, 2026 Anthropic moved Claude Agent SDK and `claude -p` headless usage off the Pro/Max subscription billing pools onto separate API-credit billing — Claude Code in interactive use is still covered, but headless and SDK usage costs you API credits.

What about Kimi K2.7 Code's file and repo handling?

Kimi K2.7 Code (Moonshot, June 12, 2026) and the Kimi Code workspace at kimi.com/code both handle repo connection and codebase indexing, similar in spirit to Codex and Claude Code. The advantages: 256K context lets you load substantially more code into a single agent turn without summarization, the MCP tool-use scores are best-in-class for open-weight models, and the cost is roughly 5x lower than Claude or GPT-equivalent surfaces. The trade-offs: forced thinking mode, locked sampling, and a brand-new product that is operationally unproven. For cost-sensitive bulk codebase work, K2.7 Code via its file-and-repo surfaces is a credible alternative to Codex. For highest-quality production engineering, Codex with GPT-5.5 or Claude Code with Opus 4.7 / 4.8 remains stronger.

What's the right multi-tool stack for AI-assisted document and code work in mid-2026?

For most teams: use Codex (GPT-5.5 backend) or Claude Code (Opus 4.7 backend) for engineering workflows on live repos, use OpenAI File Search inside Agents SDK for product-embedded document RAG, use Claude Files for high-citation-quality document QA inside Claude.ai or Claude-backed customer-facing chat, and use Kimi K2.7 Code (via Cursor or Kimi Code) as a cost-optimization route for the long tail of bulk agentic coding tasks. Hold at least two model-line backends in your routing layer to absorb policy or supply shocks like the June 12 Fable 5 / Mythos 5 suspension. Avoid lock-in to a single file-search API by abstracting your retrieval layer behind a stable interface.

Quick Answer

OpenAI Files vs Codex vs Claude Files: Which AI File-Search Stack in June 2026?

Published: June 18, 2026

OpenAI Files vs Codex vs Claude Files: Which AI File-Search Stack in June 2026?

Three different AI file-and-code surfaces are competing for the same workloads in mid-2026 — but they target different deployment surfaces and overlap only at the edges. OpenAI File Search is for product-embedded RAG. Codex is for GitHub-resident code work. Claude Files is for citation-quality document QA. Here’s how to pick for your stack.

Last verified: June 18, 2026.

TL;DR

OpenAI File Search: Inside Agents SDK / Assistants API. Best for production RAG inside agentic products.
Codex: GPT-5.5-backed coding agent on live GitHub repos. Best for AI-assisted dev workflows.
Claude Files: Anthropic’s document-handling in Claude.ai and API. Best for citation-quality QA.
Claude Code: Terminal/IDE-resident coding agent with Agent SDK; complements Claude Files.
Kimi K2.7 Code: Cost-optimization route for bulk agentic coding. Newest, ~5x cheaper.

Quick comparison

Surface	Owner	Primary model	Best surface	Typical use
OpenAI File Search	OpenAI	GPT-5.5	Agents SDK, Assistants API	Product-embedded document RAG
Codex CLI / ChatGPT Codex	OpenAI	GPT-5.5 (Codex-tuned)	Terminal + ChatGPT web	GitHub repo work, PRs
Claude Files	Anthropic	Claude Opus 4.7 / Sonnet 4.5	Claude.ai chat + API	Citation-quality document QA
Claude Code	Anthropic	Claude Opus 4.7 / Sonnet 4.5	Terminal + IDE	Engineering workflows with sub-agents
Kimi K2.7 Code	Moonshot AI	Kimi K2.7 Code	Kimi Code web + API	Cost-optimized bulk coding

Decision matrix by use case

”I’m building a customer-facing AI product with document upload”

Use OpenAI File Search inside Agents SDK as your default. It is the most production-tested of the options, handles chunking and embeddings automatically, integrates with Operator and Code Interpreter for cross-tool agent loops, and has the broadest deployment ecosystem.

Use Claude Files if citation quality is the primary success criterion (regulated industries, legal-tech, healthcare summarization) and you can accept the slightly heavier integration cost vs the OpenAI Assistants surface.

”I’m building an AI-assisted development workflow”

Use Codex when your codebase is on GitHub and the agent should operate directly against the live repo — clone, read, edit, branch, commit, PR. Codex CLI is the production tool today; the ChatGPT Codex web surface is the user-friendly entry point.

Use Claude Code when you want sub-agents, skills, and the Agent SDK’s deeper composition primitives — particularly if your team is building agent-of-agents patterns. Note the June 15, 2026 billing change moved Agent SDK and headless claude -p usage to API-credit billing.

Use Cursor + Kimi K2.7 Code routing for cost-optimized bulk codegen on lower-stakes tasks (see our Cursor vs Kimi Code comparison).

”I need conversational document QA with citations for a regulated industry”

Use Claude Files as the primary surface. Anthropic’s citation quality and constitutional-AI alignment posture are differentiators for legal, healthcare, and government deployments. The ENISA-Anthropic meeting on June 18 reinforces the EU regulatory acceptance story.

”I need product-embedded RAG that calls other tools mid-conversation”

Use OpenAI File Search inside Agents SDK. The integration with Operator, Code Interpreter, and the broader OpenAI tool ecosystem is unmatched for tool-using agents that also need document retrieval. Claude Files is improving fast but is still the second-best fit for this specific workload pattern in June 2026.

”I have a massive codebase and cost is the dominant constraint”

Use Kimi K2.7 Code via Cursor routing or the Kimi Code web workspace. 256K context lets you load substantially more code per turn. $0.95 / $4.00 per million tokens is roughly 5x cheaper than Claude Opus 4.8 routes. Trade-offs: forced thinking mode, brand-new product, Chinese-origin vendor (US procurement question for some buyers).

When you should hold multiple backends

The June 12, 2026 US export-control order on Claude Fable 5 and Mythos 5 — and the broader pattern of policy moves shaping access (G7 Evian “trusted partners,” ENISA-Anthropic on June 18, Meta-Manus unwind) — argue strongly for holding at least two model-line backends in your routing layer:

OpenAI primary + Anthropic fallback for the most common production-grade stack.
Anthropic primary + OpenAI fallback for the safety-and-alignment-led stack.
Closed-frontier primary + Kimi or DeepSeek fallback for cost-and-supply resilience.

Architect your retrieval layer behind a stable interface (LangChain, LlamaIndex, or a thin in-house adapter) so swapping File Search providers does not require rewriting your agent logic.

Cost comparison (approximate, per million tokens)

Provider	Input	Output	Notes
OpenAI File Search (GPT-5.5)	~$10	~$30	Plus storage and search costs
Codex (GPT-5.5)	~$10	~$30	Per-token; bulk discounts on Codex CLI
Claude Files (Opus 4.7)	~$15	~$75	Highest quality, highest price
Claude Files (Sonnet 4.5)	~$3	~$15	Strong default for production
Claude Code (via Agent SDK from June 15)	API credits	API credits	No longer covered by Pro/Max subscription
Kimi K2.7 Code	$0.95	$4.00	Cheapest credible option

Pricing is approximate; check provider pages for current rates.

Honest caveats

Codex and Claude Code overlap with Cursor. If you’re already in Cursor, you may not need either — Cursor routes to both backends already. Pick the IDE-level surface or the standalone CLI surface, not both.
OpenAI File Search’s chunk and retrieval quality is opaque. You don’t control the chunking strategy. For some workloads, a custom retrieval layer over your own vector store beats it.
Claude Files citations are good but not infallible. Treat them as evidence to verify, not as guaranteed accuracy.
Kimi K2.7 Code’s benchmarks are vendor-reported. Independent SWE-bench Verified scores are not yet available as of June 18.
Agent SDK billing has shifted. Anthropic’s June 15, 2026 move of Agent SDK and headless claude -p to API credits affects total cost math for Claude-based production agents.

Sources

OpenAI documentation on File Search inside Agents SDK and Assistants API.
Anthropic Claude Files announcements and Claude Code documentation.
Reuters and CNBC coverage of OpenAI June 8 IPO filing, Anthropic June 1 IPO filing.
Codersera, Lushbinary, Kingy AI, CometAPI — Kimi K2.7 Code reporting June 12-15, 2026.
Cursor SDK June 2026 release notes.

This page summarises the file-and-code-search landscape as of June 18, 2026. Expect significant change in Q3 as IPO timelines tighten and providers reposition.

OpenAI Files vs Codex vs Claude Files: Which AI File-Search Stack in June 2026?

TL;DR

Quick comparison

Decision matrix by use case

”I’m building a customer-facing AI product with document upload”

”I’m building an AI-assisted development workflow”

”I need conversational document QA with citations for a regulated industry”

”I need product-embedded RAG that calls other tools mid-conversation”

”I have a massive codebase and cost is the dominant constraint”

When you should hold multiple backends

Cost comparison (approximate, per million tokens)

Honest caveats

Sources

Related pages