Gemini CLI versus Claude Code needs an access check before a feature comparison in 2026. Google announced that Gemini CLI would stop serving consumer Google AI Pro, Ultra, and free individual access on 18 June 2026. The announcement preserves different enterprise and paid API access paths. A comparison that still treats every Google account as a free Gemini CLI account starts from an outdated assumption. Google's transition announcement.

Current access comes first

The Google announcement directs consumer workflows toward Antigravity CLI while distinguishing Gemini Code Assist Standard and Enterprise customers and paid API access. That makes the account type part of the tool-selection question. The transition should not be summarized as either universal Gemini CLI availability or universal shutdown.

For an existing organization, the useful first record is the approved product, authentication method, billing account, and repository policy. For a new individual setup, Google's current transition guidance deserves attention before an older installation tutorial. This is a recommendation about evaluating access, rather than a claim that every account has the same entitlement.

Claude Code's setup documentation describes its supported environments and authentication arrangements. An available Claude account path also needs verification for the intended workflow. An executable starting successfully is not evidence that the chosen subscription or API account covers every later request. Claude Code setup.

Installation and project setup

Gemini CLI's installation page documents npm installation and the gemini executable. Those instructions explain how to obtain the tool; they should be read alongside the access transition above. A current installation command does not restore an account path that is no longer served. Gemini CLI installation.

npm install -g @google/gemini-cli
gemini

For a fair repository trial, both agent environments should start from the same recorded commit. The task brief should identify the intended behavior, relevant code, validation command, and completion condition. Separate directories prevent one run from inheriting the other run's unreviewed edits.

The first trial should also verify the development environment. If a required compiler or browser is missing from one run, that setup difference needs resolution before interpreting the resulting patch as a model comparison. Installation, account access, and repository preparation are separate parts of the experiment.

Repository instructions

Gemini CLI documents GEMINI.md context files and a hierarchy for loading project guidance. The useful operational check is to inspect the instructions that the session actually loads, including applicable parent or workspace guidance. A file existing in the checkout does not by itself establish how every tool uses it. Gemini context documentation.

A portable task brief can describe commands and boundaries without relying on one tool's instruction-file name. It can then be adapted to the configuration each agent supports. This keeps the acceptance criteria consistent while allowing each environment to receive context through its documented mechanism.

A useful instruction file names the test runner, relevant conventions, and excluded operations. It should avoid broad claims such as every test passing unless there is evidence attached to the relevant commit. Repository instructions are part of the setup to review, not a substitute for validation output.

Permissions and execution boundaries

Gemini CLI documents several sandbox mechanisms and configuration options. The active mechanism depends on the platform and configuration. A trial should therefore record the effective sandbox rather than infer isolation from the product name. Gemini CLI sandboxing.

Claude Code documents permission rules and modes for tool requests. A fair comparison should account for those settings before comparing confirmation frequency or how much work proceeds unattended. Broader granted access is a change in the experiment, not automatic evidence of a better agent. Claude Code permissions.

The recommended trial uses equivalent approved files, dependencies, and services. Production credentials should be outside the test environment. If a task requires a new capability, its reason belongs in the evaluation record so a later reader can understand which conditions produced the result.

A practical comparison table

The table below maps the evaluation to evidence. It does not assign unsupported speed or accuracy scores.

Question	Gemini CLI evidence	Claude Code evidence
Is access available?	Current enterprise or API eligibility	Approved Claude authentication path
Which context applies?	Loaded GEMINI.md guidance	Applicable Claude project guidance
Which actions are allowed?	Effective sandbox and policy	Effective permissions and sandbox
Did the change work?	Repository acceptance checks	The same acceptance checks
What was consumed?	Applicable account usage record	Applicable account usage record
Is the patch acceptable?	Complete diff and reviewer findings	Complete diff and reviewer findings

The evaluation can stop after a small task produces enough evidence for the decision. A large migration is a poor first experiment when account configuration and project setup remain uncertain.

Automation and cost records

Gemini CLI's headless-mode documentation describes noninteractive output formats and exit codes. Automation should distinguish the final response from tool execution and process status. A plausible response alone should not be treated as a successful repository check. Gemini headless reference.

Claude Code's cost guide describes usage reporting and the relationship between API consumption and subscription information. A trial should use the records appropriate to its actual authentication method. Comparing an API invoice with a subscription's remaining usage bar requires an explanation of the different units. Claude Code costs.

A useful report retains unsuccessful attempts, additional service costs, and reviewer time. The goal is to understand the work required for an accepted patch. Excluding failed runs can make a tool look less expensive without changing the total work that occurred.

A decision grounded in repository work

The comparison should end with a narrow conclusion: which configured environment completed the selected task under the approved conditions, and what remained uncertain. That conclusion is easier to reuse than a universal ranking based on tool names alone.

The existing coding agent comparison provides broader context. Current access guidance should accompany any older comparison because Google's consumer transition materially changes the setup decision.

Oyren Codespaces include Claude Code with a Linux terminal and browser VS Code. Agent authentication and model usage are separate from machine runtime. The Claude Code cloud page describes that execution option.