AIArticle All three agents converged on the same features in 2026 — the real fork is how each one meters your usage.
Mariana Souza Every few months the same question resurfaces: Copilot, Codex, or Claude Code? And every answer that leads with benchmark deltas is stale by the time you read it, because the underlying models leapfrog each other quarterly. The durable difference in mid-2026 isn't which agent is smartest this month. It's that all three products converged on the same feature set — and then diverged hard on how they charge you for it.
That second part is the comparison that actually matters, and it's the one most head-to-heads skip.
The products converged #
Rewind to 2024 and these were three different species. Copilot was autocomplete. Codex was a retired 2021 model name. Claude Code didn't exist. Today they're near-isomorphic: each has a terminal agent, IDE integration, asynchronous cloud execution, and automated code review. GitHub Copilot grew an agent mode, a CLI, and cloud agents on top of its completions engine. Codex — now OpenAI's coding agent, not the old model — runs locally, in your editor, or as sandboxed cloud tasks. Claude Code started as a terminal agent and expanded into desktop, web, and IDE extensions.
Feature-matrix comparisons are therefore mostly noise. You can get multi-file agentic edits, test-running loops, and PR review from all three. What you can't get from all three is the same billing model — and in 2026, both OpenAI and GitHub blew theirs up within nine weeks of each other.
Then the meters diverged #
On April 2, 2026, OpenAI moved Codex from per-message billing to token-based credits across Plus, Pro, and Business plans. On June 1, GitHub followed: Premium Request Units died, replaced by usage-based AI Credits where one credit equals one cent and chat, agents, CLI, and code review are metered at model API rates. Anthropic, meanwhile, stayed on flat subscriptions with rolling five-hour usage windows.
That leaves three distinct philosophies at roughly matching price points:
| Entry | Mid | Heavy | |
|---|---|---|---|
| Copilot | Pro $10/mo + $15 in credits | Pro+ $39 + $70 credits | Max $100 + $200 credits |
| Codex | included in ChatGPT Free/Go ($8)/Plus ($20) | Pro $100 (5x limits) | Pro $200 (20x) |
| Claude Code | Pro $20/mo ($17 annual) | Max $100 (5x Pro) | Max $200 (20x) |
Copilot is now pay-per-token with a plan floor. Completions stay unlimited on paid plans, but agentic work draws down a credit balance at API rates — and the community reaction to that shift was blunt; Visual Studio Magazine's roundup of developer feedback summarized it as "you will get less, but pay the same price." The $10 Pro plan's $15 of credits evaporates fast once you lean on agent mode for repo-wide tasks. It's a teaser rate for what Copilot has actually become.
Codex is a shared pool: your agent usage and your ChatGPT usage come out of the same bucket, with per-model credit rates (OpenAI's current GPT-5.6 tiers meter anywhere from roughly 25 to 750 credits per million tokens depending on variant and direction). The upside is a genuinely free on-ramp — Codex is included even on ChatGPT's Free and $8 Go tiers. The downside is unpredictability: OpenAI's own docs put a Plus subscriber at 20–110 local messages per five-hour window, a 5x spread, because a gnarly multi-file task can burn an order of magnitude more tokens than a small edit.
Claude Code is the last flat-rate holdout: $20 Pro, $100 or $200 Max, metered by five-hour windows rather than a running dollar balance. For individuals who run long agent sessions daily, that predictability is the product. The catch shows up at the org level — Claude Code on a Team plan requires a Premium seat at $100/seat/month billed annually ($125 monthly), which makes a ten-person rollout a real line item, not an impulse buy.
How to actually pick #
Ignore this month's leaderboard and ask two questions: where does your work already live, and which failure mode of billing annoys you more — surprise overage or hitting a wall?
If your org runs on GitHub Enterprise, Copilot's gravity is hard to fight. It's still the distribution king — JetBrains' January 2026 developer survey put it at 76% awareness, far ahead of everything else — and the review/agent integration into PRs is the smoothest if GitHub is your platform. Just budget honestly: model your agent usage at API rates, because that's what you're paying now. The $39 Pro+ tier with $70 in credits is the realistic floor for anyone using agent mode seriously.
If you already pay for ChatGPT, try Codex first — it costs you nothing incremental, and its sandboxed cloud tasks are the best fire-and-forget experience of the three for parallelizable work like dependency bumps and mechanical refactors. But if you find yourself rationing messages mid-afternoon on Plus, that's the shared-pool problem, and the fix is a $100 jump to Pro.
If your workflow is terminal-first and you actually max out your sessions — long agentic runs, multiple worktrees, test-fix loops — Claude Code's flat windows deliver the most agent-hours per dollar at the $100/$200 tiers, and it's the tool with the strongest momentum: the same JetBrains survey had its awareness jumping from 31% to 57% in about seven months, tied with Cursor as co-leader in the agentic-tool category at 18% adoption. It's also the priciest way to dabble; if you only want autocomplete, don't buy it.
The verdict #
There's no single winner, but there is a wrong way to choose: picking on model quality. All three vendors swap in new models continuously — Copilot itself now serves Anthropic and OpenAI models side by side, which tells you models are becoming the commodity layer. What you're actually committing to is a billing philosophy and a platform. Usage-metered tools (Copilot, Codex) are cheaper if your agent use is bursty; flat windows (Claude Code) win if it's constant. Everyone converged on $100 as the serious-use price point, which is the real headline of 2026: the $10 AI coding tool is gone, and the vendors have stopped pretending otherwise.
Sources & further reading #
[GitHub Copilot Plans and Pricing](https://github.com/features/copilot/plans)— github.com -
[GitHub Copilot is moving to usage-based billing](https://github.blog/news-insights/company-news/github-copilot-is-moving-to-usage-based-billing/)— github.blog -
[Codex Pricing](https://developers.openai.com/codex/pricing/)— developers.openai.com -
[Claude Pricing](https://www.claude.com/pricing)— claude.com -
Devs Sound Off on Usage-Based Copilot Pricing Change— visualstudiomagazine.com -
[AI Coding Tool Adoption 2026: Developer Survey Results](https://www.digitalapplied.com/blog/ai-coding-tool-adoption-2026-developer-survey)— digitalapplied.com
[Mariana Souza](https://sourcefeed.dev/u/mariana_souza)· Senior Editor
Mariana covers the fast-moving world of machine learning and generative AI, with a particular focus on how these technologies are reshaping development workflows. When she isn't stress-testing the latest foundation models, she's usually at a local hackathon.
Discussion 0 #
No comments yet
Be the first to weigh in.