Best AI Coding Assistants Compared: What $20 Actually Buys

Claude Code is the best all-round pick: it is documented on five surfaces — terminal, VS Code, JetBrains, desktop and web — on one subscription, from $20/month on Claude Pro (checked August 2026). GitHub Copilot Pro is the best value at $10/month, because its 1,500 monthly AI credits are worth $15 at GitHub's published rate and completions are not billed against them. Zed Personal is the best genuinely free option.

Key takeaways

• Best overall: Claude Code — five documented surfaces, from $20/month on Claude Pro
• Best value: GitHub Copilot Pro at $10/month, whose 1,500 AI credits are worth $15 at the published $0.01 rate
• Best free: Zed Personal — free forever, unlimited on your own API keys
• No two vendors meter the same unit: credits, dollars, messages, five-hour windows, or hours of orbs
• Windsurf is no longer a separate product — windsurf.com now redirects to Devin Desktop

How the tools were picked#

Every number here came from a primary page: a vendor pricing page, its billing docs, or its manual. Those URLs are listed at the bottom. No source, no claim.

Three rules did the filtering:

1. No published price, no entry. Contact-sales-only products were excluded from the ranking.
2. No numbers from memory or from other blogs. Where a vendor's own pages left a figure unstated or gave it in only one place, the gap is named rather than filled.
3. No benchmarks. Not one SWE-bench, HumanEval or Terminal-Bench score could be verified from a primary source, so none appear here.

These rankings therefore score published terms — pricing transparency, platform coverage, how clearly limits are stated — not code quality. See the testing method, and how to evaluate AI tools before buying for the trial process to run first.

What $20 actually buys#

Most comparisons line these tools up by monthly price, as if a "$20 plan" meant the same thing everywhere. It does not.

Plans at or near $20, verified against each vendor pricing page in August 2026.

Tool	Plan at or under $20	Allowance measured in	Free tier
Claude Code	Claude Pro, $20/mo	Usage windows: rolling five-hour and weekly	No — listed from Pro upward
GitHub Copilot	Pro, $10/mo	1,500 AI credits; completions not billed at all	Yes — up to 2,000 completions/mo
Cursor	Pro, $20/mo	$20 of other-model usage, plus unquantified first-party usage	Yes — Hobby, limited agent requests
OpenAI Codex	ChatGPT Plus, $20/mo	Local messages and cloud chats sharing one five-hour window	Yes — Codex on the Free plan
Kiro	Pro, $20/user/mo	Credits, at a burn rate that changes by model	Yes — 50 credits
Amp	Megawatt, $20/mo	750 hours of small orbs plus $20 of agent usage	None on the pricing page
Devin Desktop	Pro, $20/mo	Messages priced by model, task size and reasoning	Yes — light agent quota
Zed	Pro, $10/mo	$5 of monthly tokens	Yes — 2,000 accepted edit predictions
Cline	No paid individual plan	Whatever your own provider charges	Yes — the Open Source plan
Aider	No subscription at all	Whatever your own provider charges	Yes — the tool itself
Tabnine	No plan at this price	Per seat, from $39/user/mo billed annually	None listed

Ten different units across eleven tools. That is the actual finding, and it is why ranking these by monthly price tells you almost nothing.

No two tools bill the same unit#

A GitHub credit, a Kiro credit, a Codex message and an Amp orb-hour are not comparable quantities, and no vendor publishes a conversion between them. Translate each into the question you care about: what happens on a heavy Thursday?

• GitHub Copilot takes the most frequent action out of metering entirely. Its docs state that code completions and next edit suggestions "are not billed in AI credits" and stay unlimited on all paid plans. Agent and chat work draws credits at $0.01 each.
• Claude Code meters time windows, not units. On Team and Enterprise plans each seat draws an allowance resetting on a rolling five-hour window and a weekly window — and that allowance covers Claude chat and Cowork too, so a long afternoon in chat spends your coding budget.
• OpenAI Codex publishes limits per five hours rather than per month, and local messages share that window with cloud chats and other agentic features.
• Cursor meters in dollars, which sounds clearest until you notice only third-party usage carries a number. Its own models sit behind an unquantified "generous included usage" line.
• Kiro publishes its burn-rate variation: Claude Sonnet 4.6 costs roughly 1.3x more credits than its Auto agent for the same work.
• Amp sells hours of orbs alongside a dollar figure of agent usage. Its manual states there is zero markup on the providers' API pricing for individuals and non-enterprise workspaces.

Where each tool actually runs#

Price tables hide platform coverage, and only some vendors document it on the page they ask you to buy from.

Surfaces named in each vendor's own documentation, checked August 2026.

Tool	Documented surfaces
Claude Code	Terminal CLI, VS Code, JetBrains, desktop app, web, Claude mobile apps
GitHub Copilot	VS Code, Visual Studio, JetBrains, Xcode, Neovim, Eclipse, SQL Server Management Studio, Zed, Raycast
Kiro	Kiro IDE, Kiro CLI, Kiro web, ACP-compatible IDEs, CI/CD pipelines
Amp	CLI, VS Code and VS Code-based editors, Neovim, Zed, JetBrains (deprecated)
Cline	VS Code and CLI on the free plan; JetBrains listed under enterprise only
Aider	Terminal, via pip, pipx, uv, Docker, GitHub Codespaces and Replit
Devin Desktop	A full IDE with syntax highlighting, autocomplete and debugging tools
Cursor	Not named anywhere on its pricing page
Zed	Not named anywhere on its pricing page
Tabnine	Described as all major IDEs, without a list

Claude Code — best for complex, multi-file work#

The VS Code extension installs into Cursor as well as VS Code, and the JetBrains plugin covers IntelliJ IDEA, PyCharm and WebStorm — though it needs the CLI installed separately first. Cost documentation is unusually concrete: for deployments billed per token, the docs put the average at "around $13 per developer per active day and $150-250 per developer per month," with 90% of users under $30 per active day.

The specific catch: on native Windows, Claude Code falls back to PowerShell as its shell tool unless you install Git for Windows, which the docs recommend so it can use Bash instead.

Pick this if your work crosses many files and several machines. Skip it if you only want inline completions inside one editor.

GitHub Copilot — best value, and the only unlimited completions#

The arbitrage improves once you read what credits do not cover: the docs state that code completions and next edit suggestions "remain unlimited for all paid Copilot plans." The action you take hundreds of times a day sits outside the meter. Paid tiers run $10 (Pro, 1,500 credits worth $15), $39 (Pro+, 7,000 credits worth $70) and $100 (Max, $200 of credits) per month.

The specific catch: on Copilot Free you cannot choose your model. The docs describe it as limited chat and agent usage "with models available through auto model selection only," making model choice a paid feature rather than a rationed one.

Pick this if completions are 80% of what you want. Skip it if you run long agentic sessions, which is exactly where credits drain.

Cursor — the all-in-one editor with the least transparent limits#

Pro is $20 and includes $20 of third-party model usage. Pro Plus at $60 includes $70, and Ultra at $200 includes $400, so the ratio improves as you climb, and overage continues at the same API rates rather than hard-stopping. The problem is the other side of the ledger: every individual plan also lists "Generous included usage" for Cursor's own models, and that phrase is never converted into requests or dollars.

The specific catch: the India-only Start plan at ₹649/month covers only Cursor's own model pool, excludes third-party models, runs in non-fast mode, and pins Grok 4.5 to a fixed medium effort level. A real discount with a real ceiling, not a cheaper Pro.

Pick this if you want one application instead of an editor plus a plugin. Skip it if you must forecast spend to the dollar.

OpenAI Codex — best if you already pay for ChatGPT#

Plans run Free at $0, Go at $8, Plus at $20, Pro from $100, and Business at $20 per user per month, with Pro offering 5x or 20x higher rate limits.

The specific catch: because local and cloud usage draw on the same window, a morning of cloud tasks can leave you rate-limited in your editor after lunch, with no way to buy out of that window.

Pick this if you already pay for ChatGPT Plus. Skip it if you need a predictable monthly ceiling for a finance team.

Zed — the best free tier if you bring your own keys#

That "accepted edit predictions" unit appears nowhere else here — a fair one, since you are metered on suggestions you took, not on suggestions the editor offered. Pro is $10 a month with $5 of tokens included and a free trial carrying $20 of token credits; Business is $30 per seat per month, with unlimited edit predictions and no bundled token allotment.

The specific catch: the free Personal plan carries no token allowance — only 2,000 accepted edit predictions, then whatever your own API keys pay for. Free here means a free editor plus your own provider bill: excellent if you have keys, useless if you do not. Pro is also the only plan on this page whose included model value ($5) is under half its price ($10).

Pick this if you already pay providers directly. Skip it if you want one invoice covering everything.

Cline — the open-source agent that runs on your own keys#

Bring-your-own-key works across the providers the page names, among them OpenAI, Anthropic, Gemini, OpenRouter, AWS Bedrock and GCP Vertex. The client-side architecture keeps inference on your own credentials, which suits a team working through the AI tool security checklist.

The specific catch: JetBrains support is listed only under the enterprise version, which has no published price. If your team lives in IntelliJ or PyCharm, the free open-source build does not cover you.

Aider — the terminal agent with no subscription at all#

Providers named in the install docs include OpenAI, Anthropic, Gemini, GROQ, DeepSeek, Ollama and Azure. Ollama is the interesting entry: the shortest documented path here to running an assistant against a locally hosted model.

The specific catch: the supported Python range depends on how you install it — 3.8 to 3.13 via the quick-start installer, but only 3.9 to 3.12 via pip or pipx — and every request is billed by your provider from the first keystroke, with no included allowance to absorb an experiment that goes wrong.

Kiro — the clearest credit accounting of any paid tool#

Individual plans run $20, $40, $100 and $200 per user per month. The free tier gives 50 credits via social login or an AWS Builder ID, with Claude Sonnet 4.5 and open-weight models including Qwen3 Coder Next, DeepSeek v3.2 and MiniMax 2.1.

The specific catch: add-on credits cost $0.04 each, and usage limits reset at the start of each billing month with no rollover for unused credits. That $0.04 is not like-for-like with GitHub's $0.01 credit — the two measure different things — but overage is where Kiro stops being a $20 tool, and 50 free credits is a look at the product, not a week of work.

Amp — zero-markup pass-through pricing for teams#

Megawatt at $20 includes 750 hours of small orbs and $20 of agent usage. Gigawatt at $200 includes 1,000 hours of large orbs and $200 of agent usage.

The specific catch: Megawatt gives you low and medium modes only. High and ultra start at the $200 Gigawatt plan, so the cheap tier is capped by capability rather than volume. Enterprise usage is 50% more expensive than individual and team plans, upgrading a workspace to Enterprise takes a one-time $1,000 purchase, and unused credits expire after one year of account inactivity.

Devin Desktop — the tool that used to be Windsurf#

Plans are Free at $0, Pro at $20, Max at $200, Teams at $80 a month base plus $40 per full developer seat, and custom Enterprise pricing. The free plan gives a light agent quota with limited model availability, plus unlimited inline edits and unlimited Tab completions.

The specific catch: nothing on the paid plans is countable in advance. Cost per message varies with the model, task size, complexity and reasoning required, allowances refresh on daily and weekly cycles, and overage is consumed at API pricing. Teams also bills an $80 base before a single seat.

Tabnine — built for air-gapped and regulated environments#

The Agentic Platform tier is $59 per user per month, also annual. If air-gapped deployment is a hard requirement, this is the shortest route on published pricing.

The specific catch: the pricing page lists no free tier, and both plans are quoted on annual terms, so there is no documented way to try it for one month. It also claims compatibility with "all major IDEs" without naming any, so you must leave the pricing page to check yours.

How to choose: a decision tree#

Answer these in order and stop at your first yes.

1. Does your security policy forbid sending code off your network?

   Tabnine, at $39 per user per month — the only pricing page here publishing on-premises and air-gapped deployment. If self-managed keys suffice, Aider pointed at Ollama is cheaper.

2. Is 80% of what you want inline completions in an editor you already use?

   GitHub Copilot Pro at $10 a month. Completions are not billed against credits at all.

3. Do you already have API keys and want the editor out of the billing path?

   Zed Personal, free forever and unlimited on your own keys — or Cline's Open Source plan if you want an agent rather than an editor.

4. Do you already pay for ChatGPT Plus?

   Use Codex before buying anything else. It is included, and Go at $8 a month is the cheapest paid tier here.

5. Does your work span many files and more than one machine?

   Claude Code, from $20 a month on Claude Pro. Five documented surfaces on one subscription is the differentiator.

6. None of the above?

   Start on Claude Code for a week of real work. If you end it wanting faster completions rather than deeper reasoning, move to Copilot Pro and save $10 a month.

Who this roundup is not for#

If you want a performance ranking, this is not it. No benchmark scores appear here, because none could be verified from a primary source, and a comparison built on unsourced benchmarks is worse than none.

If you are running a procurement process, treat this as a shortlist, not evidence. Every price here is a public list price checked on one day, and air-gapped and enterprise tiers are negotiated rather than published.

If you want the single fastest model, no pricing page can tell you that. Buy on billing terms and platform coverage, which are stable; the guide to prompts that survive model updates is the more durable investment.

What was left out, and why#

Naming the gaps beats quietly filling them with numbers from someone else's blog.

• JetBrains AI and Junie. Pricing for either could not be read from a primary source, so no tier names and no prices for them appear here.
• Gemini Code Assist. Usable individual pricing could not be loaded. The Google Cloud quota documentation that did load is a different thing — per-user caps inside a Google Cloud project, at 6,000 code requests per day, 960 for chat and related features, and 1,500 (Standard) or 2,000 (Enterprise) requests per user per day through agent mode or the Gemini CLI. That says nothing about an individual free tier.
• Amazon Q Developer, Replit, Augment Code, Qodo and Sourcegraph Cody. Pricing pages not fetched for this edition.
• Windsurf as a standalone entry. It redirects to Devin Desktop, so listing it separately would republish a factual error.
• Every benchmark, star count, user count and funding figure. None survived verification against a primary source.

Prices here are re-checked at least every 90 days, and the verification date sits at the top. The editorial policy covers what happens when a vendor changes terms between checks.

The verdict#

Buy Claude Code if you want one answer. Five documented surfaces on a single $20 subscription is the strongest coverage anyone publishes, and its cost docs name a per-developer daily figure instead of hiding behind a range. Shared usage windows are the real trade-off, and you should expect to hit them.

Buy GitHub Copilot Pro if you want the best value. Ten dollars a month for 1,500 credits worth $15, with completions never metered, is the clearest arithmetic on this page. It is weaker on long agentic runs, and locking model choice behind payment on the free tier is a genuine annoyance.

Start with Zed Personal or Cline if you have API keys and no budget. Both are free in the durable sense rather than the trial sense — and if that setup survives a month of real work, you have saved the $20 the rest of this page is arguing about. More comparisons live in the AI tools section.