OpenAI Agents API: What Freelancers Should Actually Budget (Tokens, Tools, Sandbox)

A client watches a Codex demo and asks for “a custom agent.” ChatGPT agent mode or a thin Responses script might be enough. The new Agents API might be the right sell. Or it might be three meters on one bill — model tokens, tool calls, and a hosted sandbox ticking while you sleep — with US residency rules you cannot hand-wave. This is a buy/skip cost sheet for freelancers, not another agent-hype explainer.

If you only need the model $/MTok ladder for ordinary chat and API work, use our earlier sheet: API cost: Claude vs OpenAI for freelancers. This post is the harness + sandbox meter on top of that.

Affiliate disclosure: some tool or platform links may later be affiliate or referral. Prices below are from OpenAI’s live pricing and Agents API docs as of 24 Sep 2026 (SAST) — re-check before you quote a client. Not legal advice; not a guarantee of clients, savings, or income.

What shipped 10 Sep 2026 — managed Codex harness as an API

On 10 Sep 2026, OpenAI put the Agents API into public beta. The launch framing is blunt: the same harness that powers Codex — context management, tool use, subagents, durable sessions — is now available through an API, with OpenAI running the orchestration while you pick tools and an environment.

You create a session, name a model (samples use gpt-6-astra), attach tools (MCP, web search, programmatic tool calling), optionally enable multi-agent fan-out, and choose where code runs: OpenAI-hosted sandbox, your own compute, or a partner sandbox. Partners named on the launch post include Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel. Partner bills are separate — do not invent their prices into your quote.

This is not “a new chat personality.” It is infrastructure for long-running agent work: sessions that resume, context compaction across windows, and artifacts you can pull when a turn completes. For freelancers, the product question is whether the client owns an API product or whether you are still doing desk research inside ChatGPT / Codex in the browser.

IDE multi-agent shipping is a different surface — see Cursor Projects for freelancers. Voice meters are another sibling job: Gemini 3.8 Live.

“No additional Agents API fee” — what that sentence does and does not mean

OpenAI’s announcement says there are no additional fees for using the Agents API — you pay for the tokens and tools your agents use. The Agents API overview adds the third meter in plain language: model usage at the selected model’s API rates, OpenAI tools at standard tool rates, and OpenAI-hosted sandboxes at standard container rates.

So “no Agents API fee” does not mean “agents are free.” It means there is no separate “Agents SKU” on top of usage. Long research with web search, subagents, and a 16 GB container can still outrun a ChatGPT Plus month on a bad afternoon. Quote the three meters, not the marketing sentence alone.

The three meters on one bill

Meter 1 — model tokens. Quote only IDs you see on the live pricing table. As of this draft’s re-pull on developers.openai.com/api/docs/pricing (and the matching platform pricing page), standard short-context rows include gpt-6-astra at $10.00 / $50.00 per 1M input/output tokens, gpt-6-sol at $2.00 / $10.00, and gpt-6-luna at $0.10 / $0.50. Long-context columns and Batch / Flex / Fast tiers differ — re-check the table before you lock a fixed fee. Agents samples use gpt-6-astra; that is a quality choice, not a requirement to burn Astra on every overnight job.

Meter 2 — tools. Built-in web search on the same pricing page is listed at $10.00 per 1k calls, plus search content tokens billed at the model’s rates (preview and non-reasoning rows differ). File search and other tools have their own rows. Every “just let it research the web” brief is a tool meter, not only a token meter.

Meter 3 — containers (hosted sandbox). Hosted Shell and Code Interpreter: 1 GB $0.03, 4 GB $0.12, 16 GB $0.48, 64 GB $1.92 per 20-minute session per container. Eligible sessions are billed by the minute with a 5-minute minimum. Idle keep-alives matter — see the sandbox section below.

Self-hosted or partner sandboxes remove OpenAI’s container row for that compute, but do not remove model or tool meters, and they do not change residency rules.

Hosted sandbox lifetime gotchas

OpenAI-hosted sandboxes give the agent a Linux workspace under /workspace with Python, Node, and CLI tools. Network access can be enabled, disabled, or restricted to exact host names. Files under /workspace/outputs become downloadable artifacts when a turn completes — those copies can survive after the sandbox expires. Connected sandboxes get keep-alives; if activity and keep-alives stop for an hour, the sandbox can be deleted, and that timeout is not configurable. Delete the session when you are done. Closing an event stream does not cancel the task.

For freelancers: put secrets in vault credentials, not raw env strings the agent can casually print. Do not leave a “research overnight” session open on a fat container with web search enabled and no spend cap.

Chat / Codex product vs Agents API as a client deliverable

Stay on ChatGPT / Codex in the product when the agent is your desk: drafting, one-off repo checks, personal research packs. Subscription agent messages are a different commercial shape from API keys on a client invoice.

Use the Agents API when the agent is part of what you sell — a client-owned workflow, a webhook-driven research bot, an incident helper wired into their tools, or anything that must run without you babysitting a browser tab. Thin API without a full Agents session can still be enough for single-turn Responses work; escalate to Agents when you need the managed harness (sessions, sandbox, subagents, compaction).

If the SOW only says “use AI to speed research,” do not automatically upsell Agents API. If the SOW says “we need an agent that runs in our stack with audit trails and spend limits,” ChatGPT Plus is the wrong deliverable.

Buy/skip matrix (sheet)

Job	Stay on chat / Codex	Thin API / Responses	Agents API + hosted sandbox	Skip / wait
Proposal polish / rewrite	Yes	Rarely	No	—
Overnight public research pack	Maybe	Maybe	Yes — capped	If secrets in scope
Client-facing sandbox agent (their key)	No	Weak	Yes — with spend caps	If residency forbids US
“Custom agent” with no SOW	—	—	—	Yes — clarify first
Solo desk debugging	Yes	Sometimes	Overkill	—

Get the 1-page Agents API cost chooser

Optional printable of this free guide. Soft link until checkout. Not a “paid summary.” Not a guarantee of clients, savings, or income.

Hard gates before you quote

OpenAI’s Agents API docs state data residency is currently United States only, and the API is not Zero Data Retention (ZDR) eligible — choosing a self-hosted sandbox does not make it ZDR-eligible. If a client’s DPA or “local data” ask conflicts with that, skip or redesign (different product, different vendor, or human-only workflow). That is packaging and contract hygiene, not legal advice.

On your side: set API spend caps, use a client-scoped key or project, never paste production secrets into a personal key, and write the SOW so review time and overages are billable. Card/billing country quirks for platform signup are real for international freelancers — verify billing country and card at signup; do not invent decline rates. Long-running sessions need a stable uplink; for mobile-data packaging see fibre vs mobile data for AI work.

This week’s action — one non-secret workload, then read the bill

Pick one non-secret job (public docs, a sample CSV, a throwaway repo). Create a session with a spend limit. Prefer a small model ID if quality allows; add web search only if the brief needs it; start on 1 GB or 4 GB unless you know you need more. Save artifacts from /workspace/outputs, delete the session, then open usage and fill one sheet row: chat enough / thin API / Agents + hosted / wait for clearer residency or GA.

If the bill surprises you, the surprise is the product — not a reason to hide meters from the next client quote.

Sources: Introducing the Agents API (10 Sep 2026), Agents API overview, OpenAI-hosted sandboxes, API pricing — re-check model IDs and container rows at paste time.

Soft next step

If you want the job → meters → stay-on-chat / API / Agents / wait chooser on one page, that is the printable sheet. Everything above stays free.