Three years ago, “AI coding tool” meant autocomplete. In 2026 it means an agent you can hand a ticket to: Claude Code, OpenAI Codex, and GitHub Copilot will all read your repository, edit files, run your tests, and open a pull request. Watch a demo of any one of them and you could be watching any of the three. The marketing has converged completely.
The products have not. Underneath the identical demos are real structural differences: where the agent runs (your shell, your editor, or a vendor-managed sandbox), how much it is allowed to do before a human looks, which models it can use, and how you pay — flat subscription, metered credits, or raw API tokens. Pick the wrong one and you will not get bad code; you will get constant friction — an agent that cannot reach your database, a bill that spikes the first week you actually use it, or a review workflow your team ignores.
This post compares the three major agents — plus Cursor and Gemini CLI, the notable rest of the field — along the dimensions that actually differ, and ends with concrete recommendations by workflow. One caveat up front: pricing in this space changes every few months, so every number here is as of August 2026 and linked to the primary source.
Three form factors, one identical loop
Every product here is the same thing at the core: a language model running in a loop with tools — read file, edit file, run command, search — until the task is done. If that framing is new, the mechanics are covered in What Are AI Agents, Actually?. What differs is the harness around the loop, and harnesses come in three form factors:
- Terminal agent. A CLI that lives in your shell, with access to your real environment — your git config, your database, your build tools. Maximum capability, maximum blast radius.
- IDE agent. Lives in the editor, edits files you are looking at, shows diffs inline. The natural home of pair-programming-style work.
- Cloud PR agent. Runs on vendor infrastructure in a sandbox, works asynchronously, and delivers its output as a pull request. You review the result, not the process.
Each vendor started in one of these boxes — Claude Code in the terminal, Copilot in the IDE, Codex in the cloud — and by 2026 all three ship all three form factors. The differences that remain are about center of gravity: which surface each product is genuinely good at, and which is a checkbox.
Claude Code: terminal-first, with a cloud escape hatch
Anthropic’s Claude Code is the most terminal-native of the three. It runs as a CLI inside your shell, which means it inherits your entire environment: your dotfiles, your language toolchains, your local Postgres, your deployed-service credentials. That is its superpower and its risk profile in one sentence. It mediates the risk with per-action permission prompts, allowlists, and an OS-level sandboxing mode that restricts filesystem and network access — but the default posture is “powerful tool on your real machine,” not “toy in a jail.”
Around the CLI core, Anthropic has added IDE extensions for VS Code and JetBrains, a GitHub Actions integration, and — newer in 2026 — delegated and scheduled tasks that run in Anthropic-managed cloud sandboxes from the web or desktop app, so long-running jobs continue when your laptop is closed. There is also an Agent SDK that packages the same harness as a library for building your own agents.
Model flexibility is the deliberate trade-off: Claude Code runs Claude models only (the Fable 5 / Opus 4.8 / Sonnet 5 / Haiku 4.5 lineup as of August 2026). Pricing is subscription-first: Claude Code is included in the consumer Pro plan ($17/month billed annually, $20 monthly), with Max tiers at roughly $100–$200 per month for heavier session limits, and Team seats from $20 per user per month. Alternatively you can run it against an API key and pay per token with no session caps — which is also how most teams wire it into CI.
OpenAI Codex: one agent, four surfaces
Codex (the agent, which reused the name of OpenAI’s 2021 code model) is explicitly built as one agent across four surfaces: an open-source CLI, IDE extensions (VS Code, Cursor, Windsurf), a desktop app for macOS and Windows, and Codex Cloud, which runs background tasks against your GitHub repositories in sandboxed containers. The cloud side is where Codex has invested most: you can fan out several tasks in parallel, and automated code review of pull requests is a first-class feature rather than an add-on.
Like Claude Code, Codex is single-vendor on models: it runs the GPT-5.6 family (released July 2026), with the frontier Sol tier, the balanced Terra, and the fast Luna, selectable per task on paid plans. Older GPT-5.4 models are being retired from Codex at the end of August 2026.
The pricing model is bundling: Codex is not sold separately but included with ChatGPT sign-in — even the free and budget tiers get CLI access with modest limits, and ChatGPT Plus at $20/month is the practical entry point for daily use, with higher tiers and Business plans raising the ceiling. Usage is metered in rolling five-hour windows plus weekly caps, and during 2026 OpenAI moved that metering to token-based credits. As with Claude Code, you can bypass the subscription entirely and run the CLI on an API key at per-token rates.
GitHub Copilot: the agent that lives in your pull requests
Copilot is the only product here that spans the full autonomy spectrum, from the autocomplete layer it invented through chat and IDE agent mode up to a fully delegated cloud coding agent. The delegated mode is distinctly GitHub-shaped: you assign an issue to Copilot, delegate from the agents panel, or mention @copilot in a PR comment; the agent then works in an ephemeral environment powered by GitHub Actions and comes back with a draft pull request. It never touches your machine, and it flows through the same branch protections and required reviews as any human contributor — which is exactly the governance story platform teams want.
Copilot is also the most model-flexible of the big three: a model picker spans OpenAI, Anthropic, and Google models, so a team can standardize on the tool without standardizing on a vendor. If you want a framework for that model choice itself, see Claude, GPT, or Gemini: How to Choose an LLM in 2026.
Pricing changed materially on June 1, 2026, when GitHub moved Copilot to usage-based billing: the old “premium requests” gave way to AI Credits consumed at token-based rates per model. Plan prices stayed put — Pro at $10/month with $10 of monthly credits, Pro+ at $39 with $39, Business at $19 per user, Enterprise at $39 — but heavy agent use now draws down a real meter, and the cloud agent additionally consumes Actions minutes. Autocomplete and next-edit suggestions remain unmetered. The practical effect: Copilot is the cheapest way to give an entire org some AI assistance, and no longer a flat-rate buffet for agentic workloads.
Cursor, Gemini CLI, and the rest of the field
Cursor remains the strongest IDE-centric alternative: an AI-native VS Code fork with inline agents, background cloud agents, and multi-vendor model choice. As of August 2026 its plans run from a free Hobby tier through Pro at $20, Pro+ at $60, and Ultra at $200 per month, with usage-based credit pools underneath — the same subscription-plus-meter structure as everyone else.
Gemini CLI is Google’s open-source terminal agent (Apache-2.0), and through 2025 it was the free on-ramp to this whole category thanks to a generous no-cost tier. That era ended: in mid-2026 Google sharply curtailed free individual usage, steering users toward paid Gemini API keys and its newer tooling. It is still a capable, hackable agent if you are happy paying per token — but it is no longer the obvious free default it was a year ago.
Side-by-side: autonomy, environment access, and pricing
| Dimension | Claude Code | OpenAI Codex | GitHub Copilot |
|---|---|---|---|
| Center of gravity | Terminal / local shell | Cloud tasks + CLI | IDE + GitHub PR flow |
| Autonomy range | Pair programming → delegated cloud tasks | Pair programming → parallel cloud tasks + code review | Autocomplete → chat → delegated PR agent |
| Environment access | Your real machine (with permissions/sandbox); optional cloud sandboxes | Local via CLI; sandboxed containers in Codex Cloud | Ephemeral GitHub Actions environment; IDE edits locally |
| Models | Claude only | GPT-5.6 family only | Multi-vendor picker |
| Pricing (Aug 2026) | Bundled with Claude plans from ~$17–20/mo; or per-token API | Bundled with ChatGPT plans from $20/mo; or per-token API | From $10/mo + AI Credits metered per token |
| Team / enterprise fit | Team seats, API billing for CI, Agent SDK | Business workspaces, GitHub/Slack integrations | Org policy, branch protections, admin controls |
Notice what the pricing row has in common: everyone converged on a subscription with a usage meter underneath. The flat-rate era ended because agents, unlike autocomplete, can burn hundreds of millions of tokens per developer per month. If you are sizing the per-token alternative, the mechanics of context, caching, and output pricing are covered in LLM API Pricing, Explained, and this site’s free LLM cost calculator will estimate a monthly API bill across vendors.
Two things that no longer differentiate
First, protocol support. All three agents speak MCP, the Model Context Protocol, for connecting external tools — databases, issue trackers, internal APIs. Since MCP moved to the Linux Foundation’s Agentic AI Foundation in late 2025, tool integrations are portable across agents, so a bespoke connector is no longer a reason to stay locked in. Background in What Is MCP?
Second — with an asterisk — raw capability. On routine, well-scoped tasks the three flagships are close enough that harness fit matters more than model IQ. The asterisk is that frontier-model differences still show up on long, ambiguous, multi-hour tasks, which is precisely where delegated agents operate. That is an argument for keeping the model question open — and for Copilot-style model flexibility if you cannot afford to re-evaluate tools every quarter.
Which agent for which workflow
- You live in the terminal and work local-first — long refactors, infrastructure work, debugging against real services: Claude Code. Nothing else treats your actual shell as the primary surface, and the permission model is built for it.
- You already pay for ChatGPT and want delegation cheap — or you want several parallel background tasks plus automated PR review: Codex. The bundling makes it the lowest incremental cost for individuals, and the cloud task queue is mature.
- Your team lives on GitHub and governance matters — issue-to-PR delegation, branch protections, org-wide policy, predictable per-seat floors: Copilot. It is also the only sane answer for rolling out autocomplete plus agents to hundreds of developers in one procurement motion.
- You want an AI-native editor with model choice: Cursor, accepting the separate subscription.
Two cross-cutting cautions. Security: a coding agent with shell access, web access, and your credentials is a textbook prompt injection surface — treat untrusted repositories and web content accordingly, and prefer sandboxed modes for anything you did not write (see Prompt Injection, Explained). And evaluation: do not choose from demos. Run two agents side-by-side for a week on your real backlog — the friction differences show up within days, and subscriptions are cheap enough that the trial costs less than a bad annual commitment.
The short version
- All three flagships now cover terminal, IDE, and cloud-PR form factors; choose by center of gravity, not feature checklists.
- Claude Code is local-shell-first with Claude models only; Codex is cloud-task-first with GPT-5.6 only; Copilot is GitHub-flow-first with multi-vendor models.
- Every vendor now runs a subscription with a token meter under it. Budget for usage, not just the sticker price — and verify current numbers on the pricing pages linked above, because they change quarterly.
- Integrations are portable via MCP, so the switching cost is lower than it looks. Pick for your workflow today, and re-evaluate when the next model generation lands.