Codex CLI
Codex CLI
One-line summary: OpenAI's terminal-native coding agent. Strong on benchmark throughput (leads Terminal-Bench), but bundled into ChatGPT subscriptions rather than sold as a standalone coding product — leading to low visible adoption.
What it is
OpenAI's terminal/CLI/desktop coding tool. Per Uvik, a "Desktop app launched February 2026" covering terminal, browser, and CLI surfaces. Bundled into ChatGPT Plus/Pro/Business at no incremental cost, not a separately-purchased coding product (2026-04-21-autoresearch-best-ai-coding-tools).
Why it matters to this thread
Codex CLI sits in the terminal/agentic tier alongside claude-code and has the strongest throughput claims of any tool reviewed. Its bundling model makes apples-to-apples pricing comparisons against $20/mo standalone tools impossible, and its low surveyed adoption (3% JetBrains) likely undercounts the ChatGPT-subscriber population using it.
Key facts (from 2026-04-21-autoresearch-best-ai-coding-tools)
- Pricing: Bundled into ChatGPT Plus/Pro/Business — no incremental cost.
- Desktop app: Launched February 2026.
- Throughput: MorphLLM reports "240+ tokens per second" — the fastest of the 15 agents tested.
- Benchmark: Terminal-Bench 2.0 (April 2026 MorphLLM): ~77.3% (tied with Gemini 3.1 Pro) — led the terminal-agent field in that snapshot per ai-coding-benchmarks.
- Benchmark: Terminal-Bench 2.0 (official table, 2026-08-27 pass): Codex CLI + GPT-5.5 is rank 4 at 82.2%±2.2 (23 Apr 2026) — highest official vendor CLI on the fetched page, not the TB winner (NexAU-AHE + GPT-5.5 84.7%). Different vintage/pairing from the MorphLLM row. See terminal-bench-harness-swing.
- Benchmark: SWE-bench Verified: 75.2%.
- Adoption (JetBrains April 2026): 3% adoption / 27% awareness — very low, but likely undercounted given the ChatGPT bundling.
Harness layer (official docs, 2026-08-27 pass)
From 2026-08-27-best-agent-harnesses-for-programming — sandbox and instruction chain, not another product recap.
- Permissions as policy language. Official permissions treat sandbox as first-class: built-ins
:read-only,:workspace,:danger-full-access; filesystem read/write/deny; optional network proxy + domain allowlists; macOS Seatbelt, Linux/WSL bubblewrap+seccomp, Windows elevated sandbox users. Profiles do not compose with oldersandbox_mode. Default is least-privilege local execution that refuses a command if the policy cannot be applied. - AGENTS.md. Official instruction chain: global
~/.codex, then repo-root-to-cwd walk, override files, 32 KiB default cap, fallback filenames. See agents-md-vs-claude-md. - Tools. Schema-first exec plus a separate permission engine — the Codex encoding of Anthropic’s “bash + editor is load-bearing; typed tools for the rest” claim.
- OpenAI Codex product overview (
developers.openai.com/codex) timed out in this pass — do not invent a product-overview claim.
Strengths
- Top Terminal-Bench score; best proxy for autonomous terminal-agent behavior.
- Highest reported token throughput.
- Zero marginal cost for existing ChatGPT subscribers.
Weaknesses / concerns
- Bundling obscures true adoption; no separate "Codex user" metric.
- Low awareness as a coding-specific product.
- Open question: is this the same "Codex" product across the API, CLI, and desktop app? The autoresearch source doesn't disambiguate cleanly.
GPT-5.6 Sol credit scope (official X, August 2026)
From 2026-08-25-x-ai-news-25-aug-2026-openai-sol-price-cut-claude-security (official @OpenAI reply, 2026-08-21): the GPT-5.6 Sol API/credit price drop of over 20% for 3 months is "Now available on the API and rolling out across eligible plans for ChatGPT Work and Codex credits." "Pro, Plus, and Business subscription usage remains unchanged." Linked URL not fetched. The post names Codex credits, not this CLI page's product surface — do not treat as a Codex CLI price change. Full treatment: gpt-5-6-sol-pricing. No dollar rate in the official posts.
Astra named in Codex (2026-09-08/09) — not a CLI SKU rewrite
From 2026-09-09-overnight-x-openai-astra-full-work-codex-rollout-aa-model (@OpenAI, 2026-09-08 21:06:10Z): gpt-6-astra “fully rolled out” to Plus, Pro, Business, and Enterprise users in Codex and ChatGPT Work. Access claim on the model page, not a Codex CLI permissions / AGENTS.md / price change. Chat-picker vs Work/Codex left open on gpt-6-astra.
Defense Factory “100% Codex-based” remediation (2026-09-09) — not a CLI SKU rewrite
From 2026-09-10-overnight-x-openai-defense-factory-uk-aisi-mythos-51-access (issuer Defense Factory): remediation described as 100% Codex-based; Codex Desktop / CLI / Security CLI named on the page. Issuer workflow self-report — not a permissions / AGENTS.md / price change and not “fleet patched” (merged ≠ deployed). Same clip notes separate Unity plugin for Codex docs — do not conflate with the Unity Claude Code video announce. Full treatment: openai-defense-factory.
Sources
- 2026-04-21-autoresearch-best-ai-coding-tools
- 2026-08-25-x-ai-news-25-aug-2026-openai-sol-price-cut-claude-security — official @OpenAI: Sol price drop rolls out to eligible Codex credits (not a CLI SKU change)
- 2026-08-27-best-agent-harnesses-for-programming — permissions / AGENTS.md / TB 2.0 82.2%
- 2026-09-09-overnight-x-openai-astra-full-work-codex-rollout-aa-model — Astra “fully rolled out” in Codex; not a CLI SKU rewrite
- 2026-09-10-overnight-x-openai-defense-factory-uk-aisi-mythos-51-access — issuer “100% Codex-based” remediation; not a CLI SKU rewrite