7 Best AI Coding Agents in 2026: Prices, Limits and the Pick

The coding agent stopped being a demo sometime in the last year. Every serious option now edits across files, runs your tests, and keeps going while you watch — and every one of them has found a different way to charge you for it. Some bundle the agent into a chat subscription. Some sell you a pool of usage shared with an unrelated chatbot. Some charge by the credit, where one credit is a cent and a long agent session can eat hundreds.

I read the live pricing page of each of the seven agents below, plus independent coverage of the month's two biggest pricing changes, and compared what a dollar actually buys. The short version: the cheapest paid tier here is $8 a month, the most expensive is $500, and the number that matters most is rarely the monthly price. It is the cap — the weekly limit, the five-hour window, or the credit allowance — that decides whether you can finish a real task without paying twice.

How this list was built

Four rules decided what made the cut. The tool has to be a genuine agent — it plans multi-file work, edits code, and runs commands or tests, not just autocompletes lines. It has to be buyable by an individual right now, at a published price, with no sales call and no waitlist. It has to have a free tier or a genuinely usable entry price. And its pricing page has to be readable today, October 27, 2026, so every figure below can be checked rather than trusted.

Two strong tools were left out on purpose. Devin is covered in our look at Cognition's free Devin plugin (Free, Pro $20, Max $200, Team $80 a month). Replit Agent is covered in our Replit for Business piece. Kimi Code is bundled into the membership tiers in our Kimi K3.1 Pro article. None is a bad choice; they are simply already priced elsewhere on this site.

The seven, at a glance

1. Cursor — the $8 entry point that now comes with a social network

Cursor changed more than any tool on this list in the last month. On October 12, SpaceX's AI arm — now called SpaceXSI — folded Cursor into The X Subscription, one bundle covering Grok, Grok Bot, Cursor, and X Premium. The old standalone Cursor plans, including the $20 plan most developers knew, are closed to new sign-ups; existing subscribers keep theirs for now.

The tiers, from Cursor's live pricing page, are Free at $0, Plus at $8 a month, Premium at $30, Super at $100, and Ultra at $200, with Teams from $30 per user a month. The catch is in the fine print: on Plus, Cursor runs only Grok models. Third-party models such as Claude or GPT are available from Premium up, billed separately at API list prices through top-ups — not part of the included pool. Premium adds four times Plus usage, cloud agents and automations, AI code reviews via Bugbot, and code hosting via Origin. Super and Ultra mostly buy more of the same pool: 15x and 50x Plus usage, with priority access at peak times on Super.

The included pool is shared across Grok and Cursor, which is either a bargain or a trap depending on where you spend it. If you already pay for X Premium, Premium at $30 replaces roughly $90 of separate plans, according to SpaceXSI's own math. If you never touch X, you are paying for a checkmark and a timeline to get an agent that only runs Grok models until you pay more.

Pros: lowest paid entry on this list; cloud agents and code review included from $30; Teams pricing is straightforward. Cons: third-party models cost extra on top of the subscription; the pool is shared with Grok chat, voice, and image tools; standalone Cursor plans are gone for new users.

2. Claude Code — the agent working developers default to

Claude Code is the least surprising entry here and, for most professional developers, the default choice. It runs in the terminal, works across any editor, and is the agent Anthropic builds its own products with. It is not sold separately: you get it through a Claude subscription, and the subscription is the whole pricing story.

Claude's pricing page lists Pro at $17 a month billed annually ($200 up front) or $20 month to month, Max from $100 a month with 5x or 20x Pro's usage, Team standard seats at $20 per seat a month billed annually ($25 monthly), and Enterprise at $20 a seat plus usage at API rates. Max also includes $100 to $200 a month in API credits. The API route is Opus 5.5 at $4 input and $20 output per million tokens, Sonnet 5.5 at $2 and $10, and Haiku 5.5 at $0.10 and $0.50 for prompts up to 100,000 tokens — prompts above that line cost five times as much.

Usage is capped by session and weekly limits on every plan; Max exists to raise that ceiling. There is no published message count: Pro suits a few focused sessions a week, Max 5x is for daily agent work, Max 20x is for teams running agents as a production tool. Our Claude Haiku 5.5 pricing breakdown covers the model family behind those rates.

Pros: a long agent track record; works in any editor via the terminal; API pricing is fully published. Cons: no standalone agent price — you buy the whole Claude subscription; usage caps are session-based and unpublished in message counts; heavy users will need Max, which starts at $100.

3. Codex — OpenAI's agent, bundled into your ChatGPT plan

Codex is the same deal as Claude Code with the names swapped: there is no Codex subscription. You buy a ChatGPT plan, and Codex — the CLI, the IDE extension, the cloud sandbox, and automated pull-request review — comes with it. The plans are Free at $0 with limited access, Go at $8 a month, Plus at $20, Pro at $100 (5x), $200, or $500 a month, Business at $20 per user a month billed annually ($25 monthly), and Enterprise on a custom quote.

The limit that matters is the five-hour window. On Plus, OpenAI publishes roughly 15–160 messages per five hours on GPT-6.1 Sol, depending on task size, with far fewer on the top-tier GPT-6 Astra (5–45) and far more on the lightweight GPT-6 Luna (350–3,000). Weekly limits sit on top and are unpublished. Pro plans currently have no five-hour limit — the weekly allowance is the real cap — and Pro $500 adds Astra Ultrafast mode. You can also point the Codex CLI at an API key instead and pay per token: GPT-6.1 Sol bills at $2 input and $10 output per million tokens.

Pros: the $8 Go tier is the joint-cheapest paid entry here; Pro's removal of the five-hour limit is a real differentiator; the API route gives a clean pay-as-you-go fallback. Cons: Plus users can hit the five-hour window on a single large refactor; the top model, Astra, is the most tightly rationed; weekly caps are unpublished.

4. GitHub Copilot — the credit-metered incumbent

Copilot has the most complicated pricing on this list, and the headline number is not the bill. The plans on GitHub's pricing page are Free at $0, Pro at $10 a month, Pro+ at $39, and Max at $100. Every plan includes a monthly allowance of GitHub AI Credits, where one credit is one cent: Pro includes $15 of credits a month, Pro+ $70, and Max $200. Optional flex allotments — $5, $31, and $100 respectively — buy extra usage on top.

Two details soften that. Code completions and next-edit suggestions do not consume credits at all, and they stay unlimited on every paid plan. The free plan is genuinely usable: 2,000 completions a month, entry-level models, and limited chat and agent usage, including the CLI and agent mode. The catch is the meter: a quick question to a lightweight model costs a fraction of a credit, while a long agent session on a frontier model across many files costs substantially more. When the allowance runs out, you wait for the next cycle, set a dollar budget for paid usage, or switch to a cheaper model — on Business and Enterprise plans, admins decide whether paid usage is allowed at all.

Pros: completions are unlimited and credit-free on paid plans; the free tier is the most generous starter here; agent mode and CLI are included everywhere. Cons: agent work is metered in credits, so a heavy week can outrun the allowance; the credit math is opaque until you watch it drain; Business and Enterprise pricing is not published on the page.

5. Windsurf — the flat $20 option, now called Devin Desktop

Windsurf is the simplest pricing story here, and the name on the box has changed. Since June 2026 the editor has been called Devin Desktop, run by Cognition alongside its Devin cloud agents — older guides that still say Windsurf or Cascade describe the same product under retired names. The plans on Windsurf's pricing page are Free at $0, Pro at $20 a month, Max at $200, Teams at $80 a month plus $40 per full developer seat, and Enterprise on a custom quote.

Pro is the interesting tier: increased quotas, frontier models from OpenAI and Anthropic, free use of SWE-2 — Windsurf's latest model — and cloud agents through Devin Cloud, with extra usage at API pricing. The free tier keeps unlimited inline edits and tab completions, with only a light quota of agent work. Max, at ten times the Pro price, buys significantly higher quotas rather than new features.

Pros: one flat price with no credit meter on the base tier; Devin Cloud access included at $20; unlimited completions even on Free. Cons: the rebrand means documentation is fragmented across two names; Max is a steep jump for what is mostly more quota; Teams pricing starts at $80 before the first seat is counted.

6. JetBrains Junie — bring your own key, pay for what you use

Junie is the only agent on this list that is happy to get out of the pricing business entirely. It runs inside JetBrains IDEs, in the terminal, and from GitHub and GitLab issues, and it works with any model you choose — including your own API key, at provider rates with zero markup. Its subscription is the most transparent of the seven, because the credit is denominated in dollars: one AI Credit equals $1 of AI use at public API rates.

The plans on JetBrains' Junie page are Junie Lite, a free AI layer with no trial and no card, and the AI Pro and AI Ultimate subscriptions. AI Pro costs $10 a month, or $100 a year ($8.33 a month annual), and includes 10 AI Credits per 30 days. AI Ultimate costs $30 a month, or $300 a year, and includes 35 credits. The free AI tier includes 3 credits per 30 days. Top-up credits can be bought at any time and stay valid for 12 months, and your own API keys do not consume credits. Supported models include Claude Opus 5.5 and Sonnet 5.5, Gemini 3.8 Flash, GPT-6 Astra and Sol, and Grok 4.7.

The honest limitation is the credit math: $10 of credits covers light chat and a modest amount of agent work, and a large repository refactor can use it up early in the month. The upside is that you can see exactly what a task cost, and you never pay for a bundle you did not use.

Pros: BYOK with zero markup is the most flexible model access here; credits are denominated in dollars, so the cost of a task is visible; top-ups last 12 months. Cons: the included credit allowance is small for heavy agent work; you need a JetBrains IDE to get the full experience; the best value requires annual billing.

7. Amazon Q Developer — the free tier AWS users already own

Amazon Q Developer is the quiet one. It lives in the IDE and the CLI, runs the latest Claude models, and its free tier is the easiest to start if you already have an AWS account: zero cost, with a capped number of agentic requests per month. AWS does not publish that number on the pricing page — the free tier is a trial, not a plan.

The Pro tier, on AWS's pricing page, is $19 per user a month, billed as a fraction of a user per day, so a mid-month start costs a fraction of $19. It expands the request limits and adds transformation tooling: 4,000 lines of code per user per month for upgrades are included, with usage above that billed at $0.003 per line. That is the only per-line price on this list — the rational pick for a team modernising a large codebase on AWS, and a poor pick for anyone who does not live in the AWS console.

Pros: a real free tier with the latest Claude models; per-day prorated billing; transformation pricing is published to the line. Cons: the free tier's request cap is unpublished; the value is concentrated in AWS-centric work; $19 a seat is the second-highest entry price here.

The pricing traps to check before you pay

Five patterns repeat across these seven pages.

Shared pools. Cursor's included usage is one pool shared with Grok chat, voice, image, and video tools. Heavy use anywhere drains the agent allowance everywhere — treat it as one budget, not two.

Weekly and five-hour caps. Claude Code and Codex both cap usage on windows you cannot see in advance — sessions and weeks for Claude, five hours and weeks for Codex. A monthly price that looks generous can still stop you mid-task. Pro tiers exist largely to lift these caps, which is why the $100 tier is the real price for daily agent work on both.

Credit math. Copilot's credits and Junie's credits look similar and are not. A Copilot credit is a cent, and a task's cost depends on the model and the work. A Junie credit is a dollar at public API rates, so the cost of a task is knowable in advance.

Third-party model surcharges. Cursor's Plus tier runs Grok models only. Claude or GPT inside Cursor are billed separately at API list prices — so the $8 headline is $8 plus API spend for most professional workflows.

Annual versus monthly. Claude Pro is $17 a month on annual billing and $20 month to month. Junie AI Pro is $8.33 on annual billing and $10 monthly. Over a year, the annual discount is worth two months on each.

FAQ

Which AI coding agent is best for a beginner?

Start free. GitHub Copilot's free tier gives 2,000 completions a month plus limited agent use, and Codex Go at $8 a month is the cheapest paid step up. Both let you learn the agent workflow before committing to a $20 or $100 plan.

Can I use an AI coding agent completely for free?

Yes, with limits, on all seven. Every tool on this list has a free tier. The limits are the point: free tiers cap agent requests, model choice, or both. Expect a free tier to cover learning and small tasks, not a production week.

Do I have to pay for the model on top of the subscription?

Usually no, except in two cases. Cursor's Plus tier covers Grok models only — Claude or GPT inside Cursor are billed separately at API rates. Junie lets you bring your own key, in which case you pay the model provider directly and JetBrains adds nothing.

Which agent works with the IDE I already use?

VS Code users have the widest choice: Cursor, Copilot, and Codex all run there natively. JetBrains users get the most from Junie. Claude Code, Codex, and Amazon Q Developer all run from the terminal, so they work with any editor, including none at all.

What happens when I hit my usage limit?

It depends on the meter. Subscription tools make you wait for the window to reset or upgrade to a higher tier. Credit tools let you buy more — Copilot through a dollar budget, Junie through top-ups valid for 12 months. API-key routes have no cap at all, only a bill.

Key takeaways

• Seven capable AI coding agents now start between free and $500 a month, and the cheapest paid tier is $8 — shared by Cursor Plus and Codex Go.
• Cursor's $8 entry is real, but it runs Grok models only; Claude or GPT inside Cursor are billed separately at API rates, and the usage pool is shared with Grok chat and image tools.
• Claude Code and Codex both price the agent inside a wider subscription, and on both, the $100 tier is the real price for daily agent work because it lifts the session and five-hour caps.
• Credit meters are the hidden variable: Copilot charges a cent a credit with task-dependent costs, while Junie charges a dollar a credit at public API rates with no markup.
• Start on a free tier, measure one real week of agent work, and only then pick a paid tier — the cap that stops you mid-task matters more than the monthly price.

As of October 27, 2026, the coding agent market has settled into a pattern: the agent itself is table stakes, and the price is the product. The tool you pick matters less than the meter you agree to — so read the cap before you read the headline.