AI Tools
ChatGPT Work vs Chat vs Codex — which OpenAI surface are you buying?
Plus/Pro Chat got an August Sol refresh. Work and Codex did not automatically inherit it. Treat them as separate buying decisions.
OpenAI’s August ChatGPT update improved Chat for consumers and Plus/Pro. OpenAI’s own note is explicit enough for buyers: Work and Codex keep previously released July Sol/Luna builds for now. That single sentence should kill the habit of treating “OpenAI seats” as one line item.
This guide is for UK teams deciding whether to buy ChatGPT Chat, ChatGPT Work, OpenAI Codex / coding agents, or raw API — and how those choices relate to Free vs Plus after the Luna/Think refresh. Related reading: ChatGPT Free vs Plus 2026, Sol / Luna / Terra tiers, and the OpenAI Codex harness.
Three products, three jobs
| Surface | Best for | Watch-out |
|---|---|---|
| ChatGPT Chat | Everyday Q&A, drafting, Think/slider UX for individuals | Not your API billing surface; tool limits still apply |
| ChatGPT Work | Team knowledge, workplace workflows, admin controls | Model build can lag consumer Chat after the Aug 6 refresh |
| Codex / coding agents | Delegated coding against repos | Separate harness decision — seat + usage, not “ChatGPT Pro covers it” |
Chat is the consumer and Plus/Pro conversation product. After 6 August 2026, Free/Go default to Luna with Think; Plus/Pro get refreshed Sol with an effort slider. That is a Chat story. It does not automatically upgrade Work tenants or Codex agents the same day.
Work is the workplace product: shared knowledge, admin, and team workflows. Buy Work when the job is organisational ChatGPT with controls — not when the job is “developers need an agent on the monorepo.” Expect feature and model lag versus consumer Chat after major Chat-only refreshes until OpenAI says otherwise.
Codex and related coding agents are for repository work. Compare them with Cursor, Claude Code, and GitHub Copilot on setup cost, review discipline, and metering — not on whether someone already has ChatGPT Plus. A Plus receipt is not a Codex licence.
Raw API (Sol, Luna, Terra IDs) is a fourth meter entirely. Product features that call the API should budget tokens separately from Chat seats. See the model profiles for Sol and Luna.
How teams usually get this wrong
The classic failure mode is a single spreadsheet cell labelled “OpenAI” that mixes:
- Founder Plus subscriptions used for email drafts.
- A Work rollout for customer support macros.
- A quiet Codex or Cursor BYOK spike on Sol.
- A backend feature calling the API on every page view.
When September pricing or weekly limits bite, nobody knows which surface to cut. Split the cell into Chat, Work, Codex/harness, and API before you renew anything.
Another failure mode: assuming the August Chat Sol refresh means Work answers got better overnight. Product owners promise “we’re on the new Sol” to stakeholders while Work still serves the prior build. Read the vendor note; schedule Work validation separately.
A practical buying sequence
- Write the job in one sentence: draft emails, answer employees with company knowledge, ship code with an agent, or power a product feature.
- Map the job to Chat, Work, Codex/harness, or API — only one primary surface.
- For Chat, decide Free/Go vs Plus/Pro using tool limits and Sol effort need, not vibes.
- For Work, run a pilot with real knowledge sources and admin policies before company-wide seats.
- For coding, run the AI coding assistant finder and compare setup guides — do not default to Codex because the brand is OpenAI.
- For API, pick model IDs and rate limits explicitly; never “whatever Chat uses.”
When to prefer something else
- Heavy IDE agent work with multi-file refactors: evaluate Cursor or Claude Code beside Codex.
- Strict data residency or BAA paths: check each vendor’s enterprise story; OpenAI Chat alone may not be the answer.
- Pure automation without chat UX: prefer API or an automation platform, not Chat seats.
Common mistakes
- Buying Plus for every employee when Work was the real requirement — or the reverse.
- Expecting Codex behaviour from ChatGPT Chat.
- Ignoring build lag between Chat and Work after Chat-only launches.
- Treating API spend as included in Chat subscriptions.
- Skipping harness comparison because “we’re already an OpenAI shop.”
Bottom line
Buy the surface that matches the job. Chat, Work, and Codex are different products with different model build timelines after August 2026. Keep API as its own meter. Use Free vs Plus guidance for Chat, Work pilots for company knowledge, and a real harness shortlist for coding — then your OpenAI bill finally matches how the work actually happens.
Editorial note
AI Choice Engine publishes editorial guides to help readers understand fit, trade-offs, and next steps before choosing a tool or provider.