Grok Build vs OpenAI Codex: Prices and When I Use Each

Operators keep asking me which model should run the business. The searches I actually see are narrower: Grok Build versus OpenAI Codex.

Short answer: Grok Build and OpenAI Codex both write code. They do not talk to your customers. Codex is included in ChatGPT Plus at $20 a month, and the API model gpt-5.3-codex lists $1.75 in and $14 out per million tokens. Grok Build 0.1 lists $1 in and $2 out per million tokens under 200,000 prompt tokens. I build with one of those two. Claude Sonnet 5, at $2 in and $10 out per million tokens, runs the customer-facing agent.

I still deploy phone receptionists, intake, and ops agents. The coding tool and the runtime model are two invoices. Here is how I split them.

What does Grok Build cost versus Codex?#

Grok Build 0.1 is the cheaper API when the agent writes a lot of code. xAI lists $1 in and $2 out per million tokens under 200,000 prompt tokens. OpenAI lists gpt-5.3-codex at $1.75 in and $14 out. ChatGPT Plus includes Codex at $20 a month.

Tool	What I use it for	Price I quote from	Where I leave it
OpenAI Codex	Webhooks, CRM plumbing, tests, deployment scripts	ChatGPT Plus $20/month, or API gpt-5.3-codex at $1.75 in / $14 out per 1M tokens	Live customer conversations
Grok Build 0.1	Long coding loops where output tokens dominate the bill	$1 in / $0.20 cached / $2 out per 1M under 200k; double those rates at 200k+	Anything a caller hears
Claude Sonnet 5	Receptionist flows, lead intake, layered instructions	$2 in / $10 out per 1M tokens	High-volume jobs where Haiku 4.5 at $1 / $5 is enough
Grok 4.7	Runtime when Web Search or X Search has to be on	$2 in / $0.50 cached / $6 out per 1M under 200k	Strict medical or legal intake

Sources, checked 25 September 2026: OpenAI Codex plan pricing, OpenAI API pricing for gpt-5.3-codex, Grok Build 0.1, xAI models, and Claude API pricing. Prompts that cross 200,000 tokens on Grok Build are billed at the higher rate for every token in that request. Context on Grok Build 0.1 is 256,000 tokens.

When do I open OpenAI Codex?#

I open Codex when I am already in ChatGPT and the job is scaffolding: Twilio hooks, calendar writes, CRM fields, and tests. Plus is $20 a month. Pro starts at $100 a month for a larger allowance. Heavy automation goes on an API key.

Codex is the coding agent inside ChatGPT, not a second product you buy beside it. OpenAI’s Codex pricing includes it on Free ($0), Go ($8), Plus ($20), and Pro (from $100). An API key runs Codex in the CLI or IDE and bills standard API rates, with no cloud features such as GitHub review. On the API price card, gpt-5.3-codex is $1.75 per million input tokens, $0.175 cached input, and $14 per million output tokens.

GPT-5.5 leaves ChatGPT, ChatGPT Work, and Codex plans on 14 October 2026. The API is not part of that retirement. I do not plan a client deployment around a plan model that is dated.

I still read the diff. Codex invents an API signature often enough that I will not paste a webhook into production unreviewed. If you are a solo operator building your own lane, it is a real accelerator. It is the wrong thing to put on the phone.

When do I open Grok Build?#

I open Grok Build 0.1 when the coding loop is long and output tokens dominate the bill. At $2 per million output tokens it sits well under Codex’s $14 API output rate. I still review every diff before it touches a client system.

xAI’s own picker says to use Grok 4.7 for code because it is the most capable model they list. Grok Build 0.1 is the coding-priced model on the same page, with aliases grok-code-fast-1 and grok-code-fast. I treat that as a cost choice, not a quality trophy. The CLI install for SuperGrok subscribers is a different surface. I wrote that up in what Grok Build is. This page is the API comparison people are actually searching.

Grok Build does not get a customer persona. Same rule as Codex. The small-business AI map is still trigger, action, system of record, human escalation. A coding model belongs in the action box only when the trigger is “change this integration.”

Which model talks to the customer?#

Claude Sonnet 5 runs nine out of ten of my production agents. It holds a long instruction list and declines when a caller goes out of scope. Anthropic lists it at $2 in and $10 out per million tokens. Grok 4.7 is the runtime when the answer depends on search.

Sonnet 5’s $2 / $10 rate is the standard price. Anthropic’s note says the planned 1 September 2026 increase to $3 / $15 did not happen. Opus 4.8 is $5 in and $25 out when the thread is dense enough to need it. Haiku 4.5 at $1 / $5 is what I look at when call volume would make Sonnet hurt. I put that range in the quote before anyone signs. The packaged product, separate from this model pick, is in Claude for small business.

Grok 4.7, under 200,000 prompt tokens, is $2 in, $0.50 cached, and $6 out per million tokens, with a May 2026 knowledge cutoff. xAI says it has no live-event access unless Web Search or X Search is enabled. I use it for jobs where freshness is the point: a rate that changed this week, an X-native workflow, a listing question. I do not use it as the primary model on medical or legal intake, where a clean refusal matters more than a fresh paragraph.

If you are still deciding whether any of this belongs in the business, start with evaluating AI tools without getting burned. The model is the cheap part of a bad deployment. The prompt, the logs, and the human escalation path are the expensive part.

What do I charge after the model is picked?#

The model bill stays on your API account. My bill is the workflow. One Lane is $3,500 once. Business MCP is Look Up $3,500, Desk $7,500, or Floor $14,000. If the leak is the phone, the AI Receptionist is $8,000 once.

A named job you already do by hand every week is One Lane. Customers who should book, pay, or check a schedule from ChatGPT, Claude, or Grok belong on Business MCP. Missed calls belong on the AI Receptionist.

If you are staring at Grok Build versus Codex and you still do not know which sentence is yours, send the workflow. The audit form is short. I reply with the replacement map within 24 hours: which coding tool I would use, which runtime model I would run, and which of those three offers matches the leak.