DEV Community

dubleCC
dubleCC

Posted on • Originally published at heycc.cn

Claude Code vs OpenAI Codex CLI vs Gemini CLI in 2026: Pricing, Models, Sandboxing, and the Antigravity Plot Twist

#claudecode #codexcli #gemini #antigravity

Originally published at heycc.cn. This is a mirrored copy — the canonical version is kept up to date at the source.

Claude Code vs OpenAI Codex CLI vs Gemini CLI in 2026: Pricing, Models, Sandboxing, and the Antigravity Plot Twist

Twelve months ago this would have been a straightforward three-way comparison: three vendors, three terminal agents, three npm packages. In July 2026 it is anything but. One of the three contenders — Gemini CLI — effectively stopped existing for individual developers on June 18, 2026, when Google cut off access for everyone without an enterprise license and pointed the community at a closed-source successor called Antigravity CLI. Meanwhile Claude Code and OpenAI's Codex CLI have diverged into two very different philosophies of what a coding agent should be: one a closed-source product with an unusually candid security model and a premium subscription ladder, the other a fully open-source Rust binary with an $8 entry tier and default-deny networking.

This deep dive compares all three across the axes that actually determine adoption decisions: pricing and subscription mechanics (including the fine print about shared usage pools and per-5-hour message windows), the underlying model lineups, open-source status (which is more nuanced than star counts suggest), MCP and extensibility, sandbox architecture down to the syscall-filter level, and enterprise readiness. Every fast-moving number below was checked against the vendors' primary documentation on the run date — see the verification note at the end. This comparison stays within the terminal-agent category; if you are weighing a terminal agent against an IDE-native tool, see Claude Code vs Cursor in 2026.

The headline comparison

Claude Code OpenAI Codex CLI Gemini CLI / Antigravity CLI
Vendor Anthropic OpenAI Google
Open source? No — repo is distribution + issue tracker + plugins Yes — Apache-2.0, 96.4% Rust Gemini CLI: yes (Apache-2.0); Antigravity CLI: no (closed source, Go)
GitHub stars (2026-07-02) 135,357 94,947 105,699
Latest release (run date) v2.1.198 (npm) 0.142.5 (July 1, 2026) 0.49.0 (npm; individuals cut off June 18, 2026)
Default frontier models Opus 4.8, Sonnet 5 (1M context), Fable 5 GPT-5.6 Sol, GPT-5.6 Terra, GPT-5.6 Luna (GPT-5.5 now previous-gen) Enterprise-licensed Gemini models only
Cheapest paid tier with the agent Pro, $20/mo ($17/mo annual) Go, $8/mo Gemini Code Assist Standard (org license)
Sandbox on macOS / Linux Seatbelt / bubblewrap + socat Seatbelt (sandbox-exec) / bwrap + seccomp n/a for individuals post-transition
Native Windows sandbox No (WSL2 only) Yes (native sandbox or WSL2) n/a
Default network posture in sandbox Zero pre-allowed domains; per-domain approval prompts via proxy Off entirely unless enabled in config.toml n/a
MCP support claude mcp add — http/stdio (SSE deprecated, ws via add-json) config.toml; plus subagents, web search, cloud tasks ~/.gemini/settings.json; carried into Antigravity as plugins

Three-column architecture and timeline comparison of Claude Code, Codex CLI, and the Gemini CLI to Antigravity transition

First, the earthquake: what happened to Gemini CLI

Any 2026 comparison that treats Gemini CLI as a going concern for individual developers is out of date. We covered the shutdown mechanics in detail in our best-AI-coding-assistants roundup, so the short version here: per Google's official deprecation notice, on June 18, 2026 Gemini CLI stopped serving requests for Google AI Pro subscribers, Google AI Ultra subscribers, and users of the free "Gemini Code Assist for individuals" tier. Only organizations with Gemini Code Assist Standard or Enterprise licenses — plus anyone paying for a Gemini API key or the Gemini Enterprise Agent Platform — retain access.

What that roundup did not cover, and what matters most for this head-to-head, is what came next. The designated successor, Antigravity CLI, became "available to everyone" on May 19, 2026. It is written in Go rather than TypeScript, and it carries over the extensibility surface Gemini CLI users had built against: Agent Skills, Hooks, Subagents, and Extensions (rebranded as Antigravity plugins). But there is one enormous difference: Antigravity CLI is closed source.

Google's handling of the open-source question rewards a careful read, because every individual statement is accurate while the overall picture changed completely. In the official transition discussion (gemini-cli discussion #27274), Google confirmed the gemini-cli repository "remains available to the community as an Apache 2.0 licensed repository with no changes" and will continue receiving model releases plus bug and security fixes — for enterprise customers. Each clause is true: the code is still open, the license did not change, maintenance continues. What the framing omits is that the hosted brain the code talks to no longer answers for individuals, and the actively developed successor is proprietary. The community's verdict on the announcement post was unambiguous: 291 downvotes against 6 upvotes, with recurring complaints about the closed-sourcing and about Antigravity shipping without features Gemini CLI users relied on, such as chat history compression and per-model quota controls.

The strategic lesson for tool selection is bigger than Google: a coding agent is not just a binary, it is a binary plus a subsidized model endpoint. The Apache-2.0 license on gemini-cli protected the code; it protected nothing else. Keep that in mind when weighing "open source" as a column in any comparison table, including the one above.

Open-source status: one command tells the story

The star counts in the headline table obscure a more interesting reality, and you can surface it straight from npm registry metadata. Querying version and license for all three packages (run 2026-07-02):

$ npm view @anthropic-ai/claude-code version license
version = '2.1.198'
license = 'SEE LICENSE IN README.md'

$ npm view @openai/codex version license
version = '0.142.5'
license = 'Apache-2.0'

$ npm view @google/gemini-cli version license
version = '0.49.0'
license = 'Apache-2.0'
Enter fullscreen mode Exit fullscreen mode

That SEE LICENSE IN README.md is the whole Claude Code open-source story in one field. The github.com/anthropics/claude-code repository — the most-starred of the three — contains no product source code. It is a distribution and issue-tracker repo: installation scripts, example settings, and a plugins/ directory. The agent itself ships as a compiled npm artifact under a proprietary license. Anthropic gets community issue triage and a plugin ecosystem without opening the core.

OpenAI took the opposite bet. The openai/codex repository is the actual product, Apache-2.0 licensed, with its latest release 0.142.5 shipped July 1, 2026 — the day before this article's verification run, which says something about release cadence. It is also overwhelmingly a Rust codebase: GitHub's languages API reports 38,620,595 bytes of Rust out of 40,053,610 total — 96.4%. The Rust rewrite matters practically, not just aesthetically: a single static binary with no Node runtime dependency is easier to audit, easier to vendor into locked-down environments, and faster to start.

Gemini CLI sits in the strangest position: genuinely open source (npm install -g @google/gemini-cli, three auth paths — Google OAuth, Gemini API key, Vertex AI), still receiving enterprise-targeted fixes, but functionally orphaned for the individual developers who starred it.

So the honest 2026 taxonomy is: one closed-source product with an open plugin surface (Claude Code), one fully open-source product (Codex CLI), and one open-source husk with a closed-source successor (Gemini CLI/Antigravity). If forkability or self-auditing is a hard requirement, Codex CLI is the only live option among the three.

The models underneath

The two surviving vendors run their model lineups on opposite lifecycle philosophies, and the difference has operational consequences beyond the spec sheets.

Claude Code uses alias indirection. Per the official model-config docs (July 2026), opus resolves to Opus 4.8, sonnet to Sonnet 5 — which ships a native 1M-token context window — and a third alias most comparison posts miss: fable selects Claude Fable 5, which the docs describe as "the most capable model in Claude Code, suited to tasks larger than a single sitting." Fable 5 is not the default — you opt in with /model fable — and it is positioned explicitly for multi-session, long-horizon work rather than interactive quick edits. Each model also gates on a minimum CLI version: Sonnet 5 requires v2.1.197+, Opus 4.8 requires v2.1.154+, Fable 5 requires v2.1.170+. The npm registry's current 2.1.198 clears all three bars.

Codex CLI uses explicit model ids with scoped lifecycles — and that lifecycle keeps moving. OpenAI shipped a new flagship generation, GPT-5.6, on July 9, 2026 (one week before this correction), rolling out across ChatGPT, Codex, and the API simultaneously. The current recommended lineup is a three-tier family: gpt-5.6-sol (the flagship, for the hardest and most open-ended coding, computer-use, and research work), gpt-5.6-terra (a balanced everyday tier OpenAI positions as GPT-5.5-class quality at roughly half the cost), and gpt-5.6-luna (the fastest, cheapest tier for high-volume, repeatable tasks). gpt-5.5 remains selectable as the previous-generation frontier model; gpt-5.4, gpt-5.4-mini, and the restricted-preview gpt-5.3-codex-spark — current at our original July 2 check — have dropped off the officially recommended list. Separately, gpt-5.2 and gpt-5.3-codex are deprecated when signing in with ChatGPT — the models page scopes the deprecation to ChatGPT sign-in, and API-key usage is not covered by that statement. That scoping is the tell: in the Codex world, model lifecycle is coupled to how you authenticate, which is an under-appreciated operational detail for teams mixing subscription seats and API keys.

The comparative takeaway: aliases and explicit ids fail differently. Anthropic's aliases mean your config never goes stale — but the model underneath can change without you touching anything, and (as the enterprise section shows) the same alias resolves to different models on different platforms. OpenAI's explicit ids mean deprecations force config churn — but nothing swaps silently, and the docs state Codex can be pointed at other providers exposing Chat Completions or Responses APIs, a real hedge against the exact platform risk Gemini CLI users just ate. Pick which failure mode your team tolerates better: silent drift or scheduled breakage.

Gemini/Antigravity: post-transition, model access simply follows the license — organizations on Code Assist Standard/Enterprise keep receiving Gemini model releases through the CLI; everyone else routes through Antigravity CLI or paid API keys.

Pricing: three very different ladders

Claude (claude.com/pricing, verified 2026-07-02)

Tier Price Notes
Pro $20/mo ($17/mo billed annually, $200 upfront) Includes Claude Code
Max From $100/mo Choice of 5x or 20x Pro usage
Team standard $20/seat/mo annual ($25 monthly)
Team premium $100/seat/mo annual ($125 monthly) 5x more usage than standard seats
Enterprise Seat + usage at API rates SSO, SCIM, audit logs, compliance API, HIPAA-ready option

Two pieces of fine print matter more than the sticker prices. First, the shared pool: per Anthropic's support docs, "Both Pro and Max plans offer usage limits that are shared across Claude and Claude Code, meaning all activity in both tools counts against the same usage limits." A heavy chat-research morning literally consumes your afternoon coding budget. Second, Anthropic publishes actual cost telemetry for API-billed usage: across enterprise deployments Claude Code averages about $13 per developer per active day, roughly $150–250 per developer per month, with 90% of users staying under $30 per active day. Those numbers make the $100+ Max tier legible: if your API-metered equivalent is $200/month, a $100–200 flat subscription with 5x–20x headroom prices out sensibly.

Codex (developers.openai.com/codex/pricing, verified 2026-07-02)

Tier Price GPT-5.6 Sol messages / 5h window
Free $0 —
Go $8/mo —
Plus $20/mo 15–90 (Terra: 20–110; Luna: 50–280)
Pro 5x $100/mo 75–450
Pro 20x $200/mo 300–1,800
Business $20/user/mo annual ($25 monthly) Plus-level limits
Enterprise/Edu Credits-based "Scales with credits," no fixed rate limits

Codex CLI ships in all ChatGPT paid plans ("ChatGPT Plus, Pro, Business, Edu, and Enterprise plans include Codex"), plus straight API-key auth. The credits math for API-style billing on the now-previous-generation flagship: GPT-5.5 costs 125 credits per 1M input tokens and 750 per 1M output; GPT-5.4 is 62.50/375; GPT-5.4-mini is 18.75/113. For the current GPT-5.6 family, OpenAI's July 2026 launch post lists per-1M-token pricing of $5 input / $30 output (Sol), $2.50/$15 (Terra), and $1/$6 (Luna) — we could not independently confirm the exact Codex-credit conversion for GPT-5.6 as part of this correction, so treat those as directional and verify current credit rates on the live pricing page before budgeting. OpenAI's older per-message estimate for GPT-5.5 — "usage averages 5–45 credits per message" — no longer maps directly to the Plus tier's current GPT-5.6 Sol limits shown above.

Structurally, the ladders differ in one key way: OpenAI has a Free tier and an $8 Go tier below the $20 line; Anthropic has nothing below $20. For hobbyists and students, Codex is the only one of the three with an on-ramp under a twenty. At the top end both ladders converge on "from $100/month" with 5x/20x multipliers — near-identical shapes, which at minimum tells you both vendors are pricing against the same heavy-usage cohort with the same headroom tiers.

Gemini

There is no individual price anymore. Access requires an organizational Gemini Code Assist Standard/Enterprise license or metered API-key billing. For a solo developer, the effective price of Gemini CLI in July 2026 is "become an enterprise."

Sandboxing and permissions: the deepest technical divide

This is where the two survivors genuinely differ in architecture, not just branding.

Codex CLI: policy matrix, default-deny network

Codex exposes a clean two-axis model. Three sandbox policies — read-only, workspace-write (the "Auto" preset), and danger-full-access ("No sandbox; no approvals (not recommended)") — crossed with three approval policies: untrusted / on-request / never. Enforcement is OS-native: macOS "uses Seatbelt policies and runs commands using sandbox-exec"; Linux "uses bwrap plus seccomp by default"; Windows runs via WSL2 or a native sandbox implementation — Codex is the only agent of the three with a native Windows sandbox.

The standout default is the network posture. The Codex docs state it plainly: "By default, the agent runs with network access turned off." Enabling it is an explicit opt-in — [sandbox_workspace_write] network_access = true in config.toml. An agent that cannot reach the network cannot exfiltrate your environment variables to a pastebin, full stop. The full-bypass flag is named with appropriate menace — --dangerously-bypass-approvals-and-sandbox, aliased --yolo.

Claude Code: three layers, per-domain prompts, and a documented weakness

Claude Code's docs describe three distinct control layers: permission rules (evaluated before a tool runs), permission modes (auto mode routes actions through a separate classifier; --dangerously-skip-permissions is blocked outright when running as root/sudo), and the OS-enforced sandbox that "holds regardless of what the model chose to run." That last phrase is the correct threat model for 2026: you are not defending against a malicious model so much as a manipulated one — prompt injection via a poisoned README survives any amount of model-level alignment, but not a seccomp filter.

Mechanically: macOS Seatbelt, Linux/WSL2 bubblewrap + socat, with an optional seccomp filter (installed via npm install -g @anthropic-ai/sandbox-runtime) that blocks Unix domain sockets. Native Windows is not supported — WSL2 is the path, a real gap against Codex for Windows-first shops. Default filesystem policy: writes confined to the working directory and session temp dir; reads allowed almost everywhere — including, notably, credential files like ~/.aws/credentials unless you deny them via sandbox.credentials. Network: zero pre-allowed domains; every new domain triggers an approval prompt, routed through a proxy running outside the sandbox.

Then comes the part that deserves genuine credit: Anthropic documents its own sandbox's weakness. The proxy "does not terminate or perform TLS inspection," so "code running inside the sandbox can potentially use domain fronting or similar techniques to reach hosts outside the allowlist." Organizations needing stronger guarantees are told to run their own TLS-terminating proxy via sandbox.network.httpProxyPort/socksProxyPort. Vendors rarely print their bypass vectors in their own docs; this one did, and it changes the comparison. On paper, Claude's per-domain allowlist looks more flexible than Codex's binary network toggle. In adversarial terms, Codex's default-deny is the stronger guarantee — you cannot domain-front through a network stack that does not exist — while Claude's allowlist-plus-proxy is more usable but weaker without the custom-proxy hardening step.

Permissions comparison in one framework

Question Claude Code Codex CLI
Can the agent write outside the repo by default? No (cwd + session temp only) No (workspace-write)
Can it read your cloud credentials by default? Yes, unless denied via sandbox.credentials Read-only policy available; workspace-write reads follow sandbox
Network by default? Prompt per new domain Fully off
Documented bypass vector? Yes — domain fronting past the non-TLS-terminating proxy Nothing equivalent documented
Escape hatch dangerouslyDisableSandbox (ignorable via managed settings) --yolo
Native Windows enforcement No Yes

MCP and extensibility

Claude Code manages MCP servers with claude mcp add --transport http|stdio (SSE is deprecated; WebSocket servers are added via claude mcp add-json with type "ws"). If you have not wired a server into Claude Code before, the mechanics are covered step by step in How to set up an MCP server in Claude Code. Beyond MCP, the closed core has a deliberately open rim: the distribution repo's plugins/ directory anchors a plugin ecosystem, and skills, hooks, and subagents round out the extension surface.

Codex CLI configures through config.toml and ships /model switching, image inputs, subagent parallelization, built-in web search, cloud task integration, and an exec mode for scripting the agent non-interactively — plus, being Apache-2.0, the ultimate extensibility mechanism: fork it. The multi-provider support (any Chat Completions or Responses API backend) extends that logic to the model layer.

Gemini CLI configured MCP servers in ~/.gemini/settings.json and carried project context in GEMINI.md files. Google says Agent Skills, Hooks, Subagents, and Extensions all carry over to Antigravity CLI as plugins — the migration story for extensions is actually decent; the trust story, per the community reception documented above, is not.

The 2026 pattern: MCP won. All three converged on it as the integration standard, so the differentiator is no longer "does it support MCP" but transport breadth, config ergonomics, and whether the extension surface survives the vendor's next strategy pivot.

Enterprise readiness

Claude Code has the most granular lockdown story. Managed settings can make the sandbox a hard gate: {"sandbox": {"enabled": true, "failIfUnavailable": true, "allowUnsandboxedCommands": false}} means the agent will not start without sandboxing and the dangerouslyDisableSandbox escape hatch is silently ignored. allowManagedDomainsOnly and allowManagedReadPathsOnly stop developers widening network or read policy locally, and the sandbox always denies writes to Claude Code's own settings.json files — a command can't rewrite its own policy. The Enterprise plan layers SSO, SCIM, audit logs, a compliance API, and a HIPAA-ready option on top.

One under-reported nuance for cloud-hosted deployments: third-party clouds lag the API lineup — but the gap narrowed after our original check. As of Claude Code v2.1.207 (current npm is past that version), opus resolves to Opus 4.8 on Bedrock, Vertex (Google Cloud's Agent Platform), and Claude Platform on AWS — full parity with the Anthropic API — with only Microsoft Foundry still pinned to Opus 4.6. sonnet still lags everywhere except the Anthropic API, though: Sonnet 4.6 on Claude Platform on AWS, Sonnet 4.5 on Bedrock, Vertex, and Foundry, versus Sonnet 5 on the Anthropic API. (Before v2.1.207, opus itself lagged too — Opus 4.7 on Claude Platform on AWS, Opus 4.6 on Bedrock and Vertex — so this was a live-moving target even within the two weeks since our original pass.) Newer models needing the latest version still require explicit full model names or ANTHROPIC_DEFAULT_*_MODEL environment variables. If your enterprise mandate is "everything through Bedrock," budget for Sonnet running one to two generations behind — Opus, as of today, is not — or for the config work to pin forward. This is also the concrete cost of alias indirection flagged in the models section: the same one-word config still resolves to different models depending on where it runs, just fewer of them than two weeks ago.

Codex CLI counters with Business at $20/user/month annual ($25 monthly) at Plus-level limits, and Enterprise/Edu usage that "scales with credits" with no fixed rate limits — plus the native Windows sandbox and the auditability of an Apache-2.0 codebase your security team can actually read.

Google's enterprise story is, ironically, the only Gemini CLI story left: Code Assist Standard/Enterprise orgs keep a maintained, open-source CLI with model releases and security fixes. For everyone else, the transition is the enterprise-readiness datapoint — platform continuity risk is now empirically nonzero for this category.

Decision framework

Your situation Pick Why
Budget-constrained solo dev Codex Free or Go ($8/mo) Only sub-$20 on-ramp among the three; Anthropic's ladder starts at $20
Windows-native team, no WSL Codex CLI Only native Windows sandbox; Claude Code's sandbox requires WSL2
Long-horizon, multi-session refactors Claude Code Max + /model fable Fable 5 is positioned for "tasks larger than a single sitting"; Sonnet 5 adds native 1M context
Hard egress-control requirement Codex CLI, or Claude Code plus your own TLS-terminating proxy Default-deny network cannot be domain-fronted; Claude's allowlist needs the httpProxyPort hardening step to close its documented gap
Must audit or fork the agent's source Codex CLI The only one of the three whose shipping product is Apache-2.0
Regulated enterprise (SSO/SCIM/audit/HIPAA) Claude Code Enterprise SSO, SCIM, audit logs, compliance API, HIPAA-ready option, plus hard sandbox gates via managed settings
"Everything through Bedrock/Vertex" mandate Claude Code, with eyes open Works — Opus is now at parity (4.8) on Bedrock/Vertex, but Sonnet still lags one to two generations (4.5 vs. Sonnet 5); budget config work to pin full model names forward
Existing Code Assist Standard/Enterprise org Gemini CLI Still maintained for licensed orgs, with model releases and security fixes
Displaced individual Gemini CLI user Re-evaluate all three Antigravity is closed source; if the Apache-2.0 license factored into your original choice, Codex CLI is the closest philosophical match

Which CLI to actually pick

The 2026 terminal-agent market resolved into a two-horse race with a cautionary tale attached. Claude Code is the premium, closed-core, deeply documented option — its sandbox honesty and enterprise lockdown surface are best-in-class, and Fable 5 gives it a unique long-horizon story, but it costs at least $20/month, shares its usage pool with chat, and has no native Windows enforcement. Codex CLI is the open, cheap, default-deny option — auditable Rust, an $8 on-ramp, the strongest default network guarantee, and a multi-provider hedge, at the price of a blunter permission model. And Gemini CLI is the reminder that the license on the client was never the guarantee that actually mattered. Choose by your binding constraint — the decision table above is the whole article in nine rows.

How the pricing and feature claims were checked (2026-07-02, corrected 2026-07-16)

All fast-moving facts were checked against primary sources on July 2, 2026. This article was re-checked against current primary sources on July 16, 2026, and updated in two places where the underlying facts had already moved: OpenAI shipped a new Codex flagship generation, GPT-5.6 (Sol/Terra/Luna), on July 9, 2026, superseding the GPT-5.5/5.4/5.4-mini/5.3-codex-spark lineup and giving the Pro 20x tier a published $200/month price; and Claude Code's opus alias now resolves to Opus 4.8 on Bedrock, Vertex, and Claude Platform on AWS (not just the Anthropic API) as of Claude Code v2.1.207, closing part of the third-party-cloud model lag described below. Everything else was reconfirmed unchanged.

  • npm versions and licenses: reproduced via the three npm view commands shown above (2.1.198 / 'SEE LICENSE IN README.md'; 0.142.5 / Apache-2.0; 0.49.0 / Apache-2.0).
  • GitHub star counts: read from the three repository pages on 2026-07-02 — anthropics/claude-code 135,357; openai/codex 94,947; google-gemini/gemini-cli 105,699.
  • Rust share: computed from GitHub's languages API for openai/codex — 38,620,595 Rust bytes of 40,053,610 total = 96.4%.
  • Claude pricing (all tiers, annual pricing, Team premium usage note, Enterprise features): claude.com/pricing.
  • Shared usage-pool quote: Anthropic Help Center article on using Claude Code with Pro/Max plans.
  • Cost telemetry ($13/developer/active day; $150–250/month; 90% under $30/day): Anthropic's Claude Code cost documentation.
  • Claude model aliases, minimum CLI versions, Fable 5 description, third-party-cloud model lag: Claude Code model configuration docs (docs.claude.com/en/docs/claude-code/model-config, now at code.claude.com/docs/en/model-config), re-verified 2026-07-16 — as of Claude Code v2.1.207, opus now resolves to Opus 4.8 on Bedrock, Vertex, and Claude Platform on AWS, not just the Anthropic API; only Microsoft Foundry still lags on Opus.
  • Claude sandbox architecture, credential-read default, domain-fronting disclosure, managed-settings keys: Claude Code sandboxing docs.
  • Codex pricing, message windows, credit rates, per-message estimate: developers.openai.com/codex/pricing (now redirects to learn.chatgpt.com/docs/pricing), re-verified 2026-07-16 — Pro 20x now has a published $200/mo flat price, and message-window figures now reference the GPT-5.6 family.
  • Codex model lineup and deprecation scoping: Codex models page (now at learn.chatgpt.com/docs/models), re-verified 2026-07-16 — the GPT-5.6 (Sol/Terra/Luna) family launched July 9, 2026 and is now the recommended lineup, with GPT-5.5 as previous-generation. The deprecation of gpt-5.2 and gpt-5.3-codex is stated for "when signing in with ChatGPT" and does not cover API-key usage.
  • Codex sandbox matrix, default-off network wording, config key: Codex sandboxing docs and config reference.
  • Gemini cutoff date and affected tiers: Google's Code Assist consumer-account deprecation notice. Community reception and vote tally (291 down / 6 up, read on 2026-07-02): gemini-cli discussion #27274.
  • Not independently verifiable: Antigravity CLI internals (it is closed source) — its language, plugin carryover, and May 19, 2026 availability date are as stated in Google's announcements.

Sources

Top comments (0)