Is ChatGPT Codex Actually Free? What Your Engineering Budget Needs to Know

I read three different write-ups this week about OpenAI’s Codex merger, and every single one led with the same word: free. “Available on every plan, including Free,” the Coursiv team wrote on July 10, describing the new unified ChatGPT desktop app that folded the standalone Codex product into Chat and a new agentic mode called Work. Tech Times ran a similar headline the same day. That is a real product change, announced July 9, and it matters for anyone evaluating coding-agent budgets right now. It is also not quite the whole story, and the gap between the headline and the rate-limit table is where I want to spend this piece.

OpenAI just told every ChatGPT user they have a coding agent

Here is what actually shipped. On July 9, OpenAI merged its standalone Codex app into the main ChatGPT desktop app for macOS and Windows, the same day GPT-5.6 went generally available. One app now holds three modes: Chat, Work, and Codex. Existing Codex users keep their projects and settings. The framing from OpenAI and from the write-ups covering it was consistent: every plan, including Free, gets access.

That timing is not an accident. Engineering budgets have had a rough few months. Uber capped employee spending on AI coding tools at $1,500 per employee per month per tool back in early June, after burning through its entire annual AI budget in about four months, with some individual engineers running $500 to $2,000 a month on token-billed tools. Microsoft reportedly wound down direct Claude Code licenses around the same window. When a major lab announces “free, every plan” into that environment, it lands differently than it would have a year ago. It reads like relief.

What free actually includes once the rate limit table gets checked

I went looking for the fine print, because “free” from a frontier lab usually means “free within a shape we control.” The clearest breakdown I found came from DigitalApplied on July 10, published the same week as the merger, and it is worth quoting directly because it contains the exact statistic that matters here.

That single sentence does most of the work. Free and Go users get GPT-5.6’s smallest model, Terra, and nothing else. Plus, Pro, Business, and Enterprise users can pick between Sol, Terra, and Luna and set an effort level. Every message draws from a shared agentic credit pool, estimated at 5 to 40 credits depending on the task, with Terra running at half the credit cost of Sol and Luna at a fifth. Plus-tier users get roughly 15 to 90 Sol messages every five hours before the pool resets. And the detail that undercuts the headline the most: ChatGPT Work itself is not available on Free or Go tiers at all. Only the Codex mode is, and only with Terra.

None of that makes the announcement false. Codex genuinely is on every plan, including Free. It just means “free” describes distribution, not capacity. The door is open to everyone. The room behind it still has a meter running.

Why free headlines do not survive a real sprint

This is the part most teams get wrong, and I do not think it is because anyone is being careless. It is because “free” is a word with wildly different meanings depending on which coding tool is using it, and nobody pauses long enough to ask which meaning applies.

Free is not a capacity number. It is a marketing word attached to a capacity number, and the two rarely match.

Take the budget-tier tools most engineering orgs are already piloting alongside a primary harness. A tool like Codeium (now folded into Windsurf) built its free tier around unlimited tab-complete, a genuinely open-ended autocomplete allowance, with agentic tasks metered separately. Tabnine’s positioning leans the other direction entirely, built for enterprise governance and context control rather than a generous free layer at all. Codex’s version of free is a third shape again: full app access, but real coding work capped by a shared credit pool that resets every few hours. Three tools, three definitions of the same word. An evaluation spreadsheet with a column labeled “Free tier” is comparing apples to a fruit basket unless someone also writes down what “free” is actually metering in each row.

I have watched this play out the same way more than once. Someone on the team tries the free tier over a weekend, ships a few small fixes, and reports back that it is “basically as good as the paid one.” Two sprints later, the same person hits the pool limit mid-refactor and either pays quietly out of a personal card or stalls the task. Neither outcome shows up in the pilot report, and both of them are the actual cost of the free tier nobody measured.

The budget math a Q3 memo should actually show

Put a number next to the claim before it goes in a memo. Three numbers, specifically.

First, the free-tier ceiling in real terms: for Codex, that is Terra-only access through a shared credit pool that most teams will exhaust inside a single focused coding session, not a full sprint. Second, the paid-tier real cost once token billing kicks in, which is the number Uber’s finance team learned the hard way when engineers were quietly running $500 to $2,000 a month per person before the company capped spend at $1,500 per employee per tool. Third, the switching cost if “free” turns out to be a trial funnel rather than a durable tier. All three numbers belong on the same slide, because a free tier that quietly becomes a paid one six weeks into a rollout is not a budget line, it is a budget surprise with a delay timer on it.

None of this means Codex’s free tier is a trap. It genuinely lowers the floor for trying a coding agent, and that is worth something, especially for smaller teams or early pilots. It just means the floor and the ceiling are two different things, and only one of them made the headline.

Three questions before free goes in the budget line

1. Ask what free is actually metering.

   Before anyone quotes a free tier in a planning doc, write down whether it caps messages, credits, tokens, or nothing at all. Codex free tier means Terra-only, shared-pool access. That is not the same free as unlimited autocomplete.

2. Time-box the pilot past the honeymoon week.

   A weekend trial will not surface a credit-pool ceiling. Run any free-tier pilot across at least two full sprints, including a heavy refactor week, before anyone calls it representative.

3. Name the owner of the re-check.

   Free tiers move. Put one person's name on checking whether Codex, or whatever budget-tier tool your team picks, is still free in the same shape three months from now, before it quietly becomes a line item nobody approved.

That is the whole exercise. Not skepticism for its own sake, just the habit of reading the rate-limit table before the headline goes in the deck. The free tier is real. So is the meter underneath it.