Gemini CLI vs Claude Code: Which Terminal AI Coding Agent Wins in 2026?

Gemini CLI vs Claude Code is the most-searched terminal-agent comparison of 2026, and almost every page ranking for it is out of date. They still promise you 1,000 free Gemini requests a day. That offer ended on 18 June 2026.

On that date, per Google's own developer blog, "Gemini CLI and Gemini Code Assist IDE extensions will stop serving requests for Google AI Pro and Ultra, as well as those using it free of charge." Gemini CLI did not shut down — but it stopped being a tool individual developers can use. It is now reachable only through a Gemini Code Assist Standard or Enterprise license.

That single fact reframes the whole comparison. This is no longer "free Google agent vs paid Anthropic agent." It is "an enterprise-licensed legacy agent vs a $20/month consumer product."

TL;DR — the short answer

• Pick Claude Code if you are an individual developer, a startup, or any team without a Gemini Code Assist license. It is the only one of the two you can actually buy as a person, starting at $17-20/month.
• Pick Gemini CLI only if your organisation already holds Gemini Code Assist Standard or Enterprise, you want published per-day quotas instead of rolling windows, and you need an Apache-2.0 agent you can fork and audit.
• Pick neither if you wanted the free Google terminal agent — that is now Antigravity CLI, a Go rewrite that shares its harness with Antigravity 2.0.
• On raw model scores it is close and unresolved. Gemini 3.1 Pro publishes 80.6% on SWE-bench Verified; Anthropic has stopped publishing SWE-bench for its 5.x models entirely, so a clean head-to-head number does not exist.

What actually happened to Gemini CLI

Google announced the transition at I/O 2026. Three groups lost access on 18 June 2026:

1. Google AI Pro subscribers
2. Google AI Ultra subscribers
3. Free users of Gemini Code Assist for individuals

New GitHub organisation installations stopped the same day. What survived is narrow and explicit: "If your organization uses Gemini CLI or our IDE extensions via a Gemini Code Assist Standard or Enterprise license… your access remains unchanged. We'll continue to support Gemini CLI and Gemini Code Assist with access to the latest Gemini models and other updates."

The repository is emphatically not abandoned. As of 2 October 2026 google-gemini/gemini-cli is unarchived, Apache-2.0, sitting on 107,218 stars, with stable v0.62.0 shipped 29 September 2026 and a nightly build (v0.64.0-nightly.20261002) cut that morning. It also carries 794 open issues.

So Gemini CLI in late 2026 is best understood as a well-maintained enterprise tool with an open-source codebase — not the free hobbyist on-ramp the search results still describe.

The developer reaction was not quiet. The Hacker News thread titled "Gemini CLI will stop working from June 18, 2026" drew 406 points and 209 comments, centring on the loss of an OAuth path for independent developers and on weekly rather than daily quotas in the replacement. In GitHub discussion #27274, contributors objected that Antigravity CLI lacks diff visibility, markdown project memory and vim support, and that volunteer work on an Apache-2.0 project now benefits enterprise licensees first.

Pricing: the comparison is no longer like-for-like

Google publishes Gemini Code Assist seat pricing directly on its Cloud pricing page:

Claude Code is included in every paid Claude plan, per claude.com/pricing:

At the entry level Claude Code Pro ($20) undercuts even annual Gemini Code Assist Standard ($19 with a 12-month commitment, $22.80 without) once you account for the fact that a single developer cannot buy Code Assist Standard the way they buy a Pro subscription.

For budgeting at scale, Anthropic publishes unusually candid figures in its cost management docs: enterprise deployments average around $13 per developer per active day and $150-250 per developer per month, and stay below $30 per active day for 90% of users. Gemini Code Assist's flat seat price is more predictable; Claude Code's is higher but comes with no separate token bill on subscription plans.

Quotas: published numbers vs rolling windows

This is the cleanest real difference between the two products.

Gemini CLI gives you a number. Per the Gemini CLI quota documentation:

• Gemini Code Assist Standard: 1,500 requests/user/day, 120 requests/user/minute
• Gemini Code Assist Enterprise: 2,000 requests/user/day, 120 requests/user/minute
• The retired individual free tier was 1,000 requests/day and 60/minute

Claude Code gives you a window. Usage runs against a rolling five-hour window and a weekly window, and on Team and Enterprise seats that allowance is shared with Claude chat and Cowork — so heavy chat use eats coding capacity. Anthropic publishes no requests-per-day equivalent.

Which is better depends on temperament. A per-day request cap is forecastable and easy to put in a budget spreadsheet. A token-based rolling window adapts to the work: a hundred one-line questions barely register, while one deep refactor of a large codebase can consume a lot. In practice the request-count model penalises many small prompts and the token model penalises long-running context.

Worth noting: requests are not a stable unit either. Anthropic's own docs explain that Agent Teams use roughly 7x the tokens of a standard session, which is why they ship disabled and need CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1.

Models and context

Gemini CLI has no single hardcoded default. It resolves model aliases through fallback chains across gemini-3-pro-preview, gemini-3.1-pro-preview, gemini-2.5-pro, gemini-2.5-flash and auto. That is flexible but has a sharp edge: without preview access, gemini-3.1-pro-preview quietly falls back to gemini-2.5-pro, so you can believe you are running the flagship when you are not. Its Plan Mode model routing (on by default) deliberately uses Pro for planning and Flash for implementation to conserve quota — a genuinely smart bit of engineering.

The flagship, Gemini 3.1 Pro, takes 1M input tokens and 64K output tokens, with a February 2026 knowledge cutoff.

Claude Code runs Anthropic models only, but across a wide ladder — Fable 5.1 and 5, Opus 5.5 down to 4.6, Sonnet 5.5 down to 4.5, and Haiku 4.5. Per the model configuration docs, Sonnet 5.5, Sonnet 5, Fable and Opus 4.7+ carry a native 1M-token window with no usage credits required on any plan; everything else defaults to 200K.

So both reach a million tokens of context. The difference is that Gemini CLI gets there with one model and Claude Code gets there across most of its current lineup.

Benchmarks: an honest non-answer

Every competing article gives you a confident head-to-head score. We cannot, and neither can they.

Google publishes SWE-bench Verified. Anthropic no longer does.

• Gemini 3.1 Pro: 80.6% SWE-bench Verified (single attempt), 68.5% Terminal-Bench 2.0. Its predecessor Gemini 3 Pro scored 76.2% and 56.9%.
• Claude Opus 5.5 (22 Sept 2026): 66.4% Terminal-Bench 4.0, 54.4% FrontierCode v1.1, 81.8% OSWorld 2.1. No SWE-bench Verified score.
• Claude Sonnet 5.5 (28 Sept 2026): 70.6% Terminal-Bench 4.0 — ahead of Opus 5.5 — at $2/$10 per MTok versus Opus's $4/$20.

Two warnings. First, Google reports Terminal-Bench 2.0 and Anthropic reports Terminal-Bench 4.0; 68.5% and 70.6% are different benchmarks and cannot be subtracted from each other. Second, if you find a page claiming a specific SWE-bench Verified percentage for any Claude 5.x model, treat it as fabricated — Anthropic has not published one.

The most interesting verifiable result here is internal to Anthropic: Sonnet 5.5 beats Opus 5.5 on Terminal-Bench 4.0 while costing half as much per token. If you run Claude Code, defaulting to Sonnet is both the cheaper and the better-benchmarking choice for agentic terminal work.

Agent features and ecosystem

Both are mature agents. Both support MCP, both have subagents, both read a project context file (GEMINI.md and CLAUDE.md).

Gemini CLI's distinctive strengths:

• Built-in Google Search grounding — no MCP server needed for live web facts
• Official GitHub Action for automated PR review and issue triage
• Checkpointing to save and resume long sessions
• Sandboxing with trusted-folder controls
• Headless output via --output-format json and --output-format stream-json
• Remote subagents alongside local ones
• Apache 2.0 — you can fork it, audit it, run it air-gapped

Claude Code's distinctive strengths:

• Skills — on-demand instruction bundles that stay out of context until invoked
• Agent Teams — multiple coordinating instances, each with its own context window
• Hooks that preprocess tool output before the model sees it
• Plugins, including code-intelligence plugins giving precise symbol navigation in typed languages
• MCP tool search — tool definitions are deferred by default, so servers cost almost no context until used
• Cost observability that no competitor matches: /usage attribution by skill, subagent, plugin and MCP server, /insights HTML reports, OpenTelemetry export, and a modelPricing managed setting so reported figures match contracted rates
• Cloud deployment through Amazon Bedrock, Google Cloud's Agent Platform and Microsoft Foundry

Claude Code is the richer platform. Gemini CLI is the more auditable one.

Security: neither gets a pass

Agentic CLIs execute code, and 2026 proved the blast radius is real.

Gemini CLI had the worse single incident: CVE-2026-12537, scored CVSS 10.0 — an OS command injection in the container launcher that allowed host code execution on headless CI before the sandbox even started (The Hacker News).

But a separate August 2026 disclosure covered a prompt-injection chain in which a crafted GitHub issue could reach CI workflow secrets through flaws in both Claude Code and Gemini CLI (The Hacker News).

The practical lesson applies to either tool: treat agent CI runners as untrusted, scope their secrets to the minimum, and never let an agent triage third-party issues with production credentials in the environment.

The verdict

For the overwhelming majority of people searching this comparison, Claude Code wins by default — because Gemini CLI is no longer available to them. If you are an individual developer, a freelancer, or a team of five without a Google Cloud contract, there is no Gemini CLI option to weigh. Claude Code Pro at $17-20/month is the decision.

Choose Gemini CLI if your organisation already pays for Gemini Code Assist Standard or Enterprise, you value a published 1,500-2,000 requests/day ceiling over a rolling token window, you need Google Search grounding natively, or your compliance posture requires an Apache-2.0 agent you can audit and self-host. Accept that you are on a maintained legacy path.

Choose Claude Code if you need to buy it as a person, want the deepest agent tooling in the category, want 1M-token context across most of the lineup rather than one preview model, or need per-developer cost attribution for a finance team.

Choose Antigravity CLI if you came here looking for the free Google terminal agent — that is where Google moved it. Our Antigravity CLI vs Claude Code comparison covers that matchup directly.

One closing note on the category: the fact that Google's answer to "we are deprecating the free tier" was "here is a different free agent instead" tells you how aggressively this market is still being bought. Treat every quota number in this article — ours included — as a snapshot of 2 October 2026, not a guarantee.

Pricing, quotas and benchmark scores verified against primary vendor sources on 2 October 2026 and subject to change. All figures link to their source.

Sources

• Google Developers Blog — Transitioning Gemini CLI to Antigravity CLI
• google-gemini/gemini-cli on GitHub
• Gemini CLI — Quotas and pricing
• Google Cloud — Gemini products pricing
• Gemini 3.1 Pro model card
• Claude pricing
• Claude Code — Manage costs effectively
• Claude Code — Model configuration
• Anthropic — Claude Opus 5.5
• Anthropic — Claude Sonnet 5.5
• Hacker News — Gemini CLI will stop working from June 18, 2026
• GitHub discussion #27274 — Gemini CLI transition
• The Hacker News — Google fixes CVSS 10.0 Gemini CLI CI RCE
• The Hacker News — Claude Code and Gemini CLI flaws expose CI secrets