AI Stacks · Pricing checked November 21, 2026

AI Coding Agents Compared: Cursor vs Claude Code vs Codex vs Copilot

Every coding agent plan on one page, priced from each vendor's own site on November 21, 2026, with what actually went wrong when we built with Cursor, Claude Code, Codex and Antigravity on the show.

Watch: Episode 30, Claude Opus 4.6, Codex 5.3, and the Rise of Multi-Agent AI from 06:47. Sam moves a live app from Claude Code to the Codex app and compares speed and bug-finding.
Subscribe on YouTubeWe build something real with AI every week

AI price-change tracker and build-cost estimator: every dated price change from the vendors' own pages, starting with Gemini 4 Carbon doubling on December 1, 2026.

The short answer

Pick Claude Code when you want fast, fun iteration and lots of parallel agents in the terminal. Pick OpenAI Codex on high reasoning when you care more about fewer bugs than speed. Pick Cursor if you want a polished IDE with a built-in browser, but note that its plans now bundle Grok and bill third-party models like Claude separately. Antigravity has the most generous free tier but was the least polished when we used it. GitHub Copilot is the cheapest paid on-ramp, at $10 a month.

We've switched coding tools more than any other part of our stack. Jordan's team went from Cursor to Claude Code to Codex in about twelve weeks in 2025, and it's still moving. Here's the dated pricing, then what actually happened when we built with each one.

Pricing and limits (checked November 21, 2026)

Every number below comes from the vendor's own pricing or docs page, linked under each table. The usage allowances are mostly stated as multiples or estimates rather than hard numbers, so compare them as tiers, not as exact quotas.

Cursor

Plan / modelPriceIncludedDocumented limitsOverage / pay-as-you-go
Free$0Limited agent and Chat requestsCursor's own models only—
Plus$8/mo ($80/yr)Base weekly usage poolGrok models only; no third-party models or BYO API keysPrepaid credit at per-model rates
Premium$30/mo ($300/yr)4x Plus usage, refreshes weeklyAdds cloud agents, Bugbot, third-party modelsThird-party models at API list price, billed separately
Super$100/mo ($1,000/yr)15x Plus usageSame as PremiumSame as Premium
Ultra$200/mo ($2,000/yr)50x Plus usageSame as PremiumSame as Premium
Teams$30, $100 or $200 per seat/moMatches the seat tierSpending caps availableAPI list price + $0.25 per 1M tokens on eligible third-party models

Source: Cursor pricing, checked November 21, 2026. Cursor now bundles Grok models and X features into its plans; included usage is stated only as multiples of Plus.

Claude Code (Claude plans)

Plan / modelPriceIncludedDocumented limitsOverage / pay-as-you-go
Free$0Limited usage every 5 hours and every weekNo Claude Code—
Pro$20/mo ($17/mo billed annually)Includes Claude Code5-hour and weekly limitsUsage credits
Max$100/mo5x Pro usage; Claude Code5-hour and weekly limitsUsage credits
Ultra$200/mo10x Pro weekly usage; Claude CodeNo 5-hour limit for verified accountsUsage credits
Team$25 standard / $125 premium per seat/mo ($20 / $100 annual)Claude Code and Claude CoworkPer seat—
API keyPer token (e.g. Opus 5.5 $4 in / $20 out per 1M)Claude Code billed at API ratesAPI rate limits—

Source: Claude Code (Claude plans) pricing, checked November 21, 2026. API rate from platform.claude.com/docs/en/about-claude/pricing, same date.

OpenAI Codex (ChatGPT plans)

Plan / modelPriceIncludedDocumented limitsOverage / pay-as-you-go
Free$0Codex on quick coding tasks (rolling out)——
Go$8/moCodex for lightweight coding tasks——
Plus$20/moWeb, CLI, IDE, iOS; OpenAI estimates messages per 5 hours by modelWeekly limits may also applyPurchased credits
ProFrom $100/mo ($100, $200 or $500 tiers)Higher limitsNo five-hour limit currentlyCredits
Business$25/user/mo ($20 annual, 2+ users)Same 5-hour estimates as PlusWeekly limits may applyCredits
API keyUsage-basedBilled at API pricing, not the ChatGPT allowanceAPI rate limitsPer token

Source: OpenAI Codex (ChatGPT plans) pricing, checked November 21, 2026.

GitHub Copilot

Plan / modelPriceIncludedDocumented limitsOverage / pay-as-you-go
Free$0Limited chat and agent use2,000 completions/month—
Pro$10/mo1,500 AI credits/moUnlimited completions$0.01 per extra credit
Pro+$39/mo7,000 AI credits/moUnlimited completions$0.01 per extra credit
Max$100/mo20,000 AI credits/moUnlimited completions$0.01 per extra credit
Business$19/seat/mo1,900 AI credits per user/moAdmins set usage limitsAdmin-controlled
Enterprise$39/seat/mo3,900 AI credits per user/moAdmins set usage limitsAdmin-controlled

Source: GitHub Copilot pricing, checked November 21, 2026.

Google Antigravity

Plan / modelPriceIncludedDocumented limitsOverage / pay-as-you-go
For Individuals$0/monthBaseline agent quota, refreshed weeklyWeekly baseline only—
Google AI Pro / UltraNot shown on the Antigravity pageLarger agent quota refreshed every 5 hoursVary by planPaid AI credit overages
Business (Google Cloud)From $30/seat/mo, or $0-seat pay-as-you-goSeat credits for first-party Gemini modelsSpend caps, admin controlsConsumption-based

Source: Google Antigravity pricing, checked November 21, 2026.

Windsurf (Devin pricing)

Plan / modelPriceIncludedDocumented limitsOverage / pay-as-you-go
Free$0Light agent quota; limited models1 member—
Pro$20/moHigher quotas, frontier modelsAllowance refreshes daily and weeklyExtra usage at API pricing
Max$200/moSignificantly higher quotas1 memberExtra usage at API pricing
Teams$80/moMultiplayer sessions, SSO, central billingUp to 100 users—

Source: Windsurf (Devin pricing) pricing, checked November 21, 2026. windsurf.com/pricing now redirects to devin.ai/pricing.

What broke when we built with it

Claude Code: confident about buggy code

Claude Code was where Sam wrote his first app. He built an applicant tracking system front end in about an hour, though it took "a few back and forth" prompts to get past early problems (Episode 9). He also hit a human limit on parallel agents:

for me, personally, my limit was around three agents, making sure I was, like, you know, engaging appropriately with each agent

— Sam Nadler, Episode 9, at 04:31

Jordan's complaint, and the reason he moved most of his work to Codex, was about quality:

One thing I would notice like repetitively with with [Claude Code] is that it would make mistakes, but be confident that there were no mistakes.

— Jordan Metzner, Episode 35, at 17:23

When he tried Claude's agent teams, he still had to "go back and clean up" major issues (Episode 31). Sam's Offsiteio estimator, now in production, started in Claude Code and moved to Codex "after some different neat bugs and updates" (Episode 38).

Codex: slow runs and a token bill nobody could trace

Codex had the opposite problem. Sam moved his AI personal-trainer app into the Codex desktop app on extra-high reasoning:

just this really deep reasoning, finding a ton of bugs. You know. It usually takes, like, five to eight, maybe even ten minutes, so it's not as instantaneous

— Sam Nadler, Episode 30, at 07:20

His workaround was to use a lower-reasoning model for quick iteration and save deep reasoning for polish. The bigger problem was cost visibility. In 2025 Jordan said the team was "maxing out the codex usage because it's cheap as hell" (Episode 14). By 2026 that had flipped:

the [Codex] computer use feature over the past few weeks, like, definitely killed us on spend on tokens. And it took us like two and a half weeks to just figure out what it was that was causing all the spend.

— Jordan Metzner, Episode 45, at 21:13

it did take us about a week and a half on a few occasions to really trace where the spend was coming from. It wasn't very transparent on their dashboards.

— Sam Nadler, Episode 45, at 21:46

Cursor: surprise bills and rate limits

Cursor was Jordan's go-to IDE for most of 2025 ("definitely the most developed", Episode 24), but the cost of running Anthropic models through it showed up quickly: "I see my cursor credit card getting hit all the time" (at 33:05). Later:

we saw recently how one developer used a new feature in cursor and left us with like a $2,000 bill.

— Jordan Metzner, Episode 36, at 13:54

During Gemini 3's launch week Jordan couldn't get the new model to run in Cursor or in Antigravity: "I'm being throttled everywhere" (at 18:13). Looking back a year, he remembered Cursor scattering numbered markdown notes that "wouldn't even be in sequential numbers" (at 16:39). By 2026 he said the most advanced teams had "moved on from Cursor" (at 12:49).

Antigravity: free, but rough

Jordan used Antigravity's free tier "until they kick me off", and found it "not as polished as a product" (at 31:37). A later update swapped the IDE for an agent interface:

It's an entire new interface copying, you know, the the [Codex] and [Claude Code] type interface. So I kind of lost the IDE part of it, and I don't know if I can get it back.

— Jordan Metzner, Episode 43, at 22:59

All of them: big codebases

Talking with Bito's CEO, Jordan said none of these tools could index a large monorepo or microservice codebase well: "None of them have the ability to do that" (at 08:05).

We have no build of our own with GitHub Copilot, Windsurf or Kiro on the show, so they appear in the price table only.

What our guests reported

Guests run these tools at company scale. Iliya Valchanov's team moved everyone from Cursor to Claude Code and cancelled Cursor (Episode 35). A healthcare guest called Claude Code "already very expensive" next to everything else (at 34:28). Espresso AI's Ben Lerner said his team pays "a lot less" for Claude Code than a third of an engineer's cost per engineer, for a 10 to 30% productivity gain, but that 25 background agents doing serious back-end work is something "we're just not there yet" on (Episode 27).

How to choose

  • You want to try agents for free: Antigravity for individuals, Codex on ChatGPT Free, Cursor Free or GitHub Copilot Free.
  • You already pay for ChatGPT or Claude: Codex is in ChatGPT Plus ($20) and Claude Code is in Claude Pro ($20), so start there before you add another subscription.
  • You run agents all day: compare Claude Max/Ultra ($100/$200), ChatGPT Pro (from $100) and Cursor Super/Ultra ($100/$200). Remember that Cursor bills Claude and other third-party models on top.
  • You manage a team budget: set spend caps on day one. Both of our real cost surprises (a $2,000 Cursor bill and a Codex computer-use spike) came from single features nobody was watching.

FAQ

How much does Claude Code cost?

As of November 21, 2026, Claude Code is included in Claude Pro ($20 a month, or $17 a month billed annually), Max ($100 a month, 5x Pro usage) and Ultra ($200 a month, 10x Pro weekly usage). It is not included on the free plan. You can also run it with an API key and pay per token, for example $4 per million input tokens and $20 per million output tokens on Claude Opus 5.5.

What does Cursor cost now?

Cursor's pricing page lists Free, Plus ($8 a month), Premium ($30), Super ($100), Ultra ($200) and Teams (from $30 per seat). Plus only includes Grok models. On Premium and up, third-party models such as Claude are billed separately at API list price from prepaid credit.

Which ChatGPT plans include OpenAI Codex?

OpenAI's Codex pricing page lists Codex on Free (quick tasks, rolling out), Go ($8 a month), Plus ($20), Pro (from $100) and Business ($25 per user monthly). OpenAI gives estimated message counts per five hours rather than hard caps, and weekly limits may also apply.

Is Claude Code or Codex better?

On the show, Claude Code felt faster and more fun to iterate in, while Codex on extra-high reasoning took five to ten minutes per run but found more bugs and wrote cleaner code. Jordan's daily driver became Codex 5.5 extra high. Sam often starts a build in Claude Code and finishes it in Codex.

What is the cheapest AI coding agent?

Several have free tiers: Cursor Free, GitHub Copilot Free (2,000 completions a month), Codex on ChatGPT Free, Antigravity for individuals and Windsurf Free. The cheapest paid plans on November 21, 2026 were Cursor Plus and ChatGPT Go at $8 a month, then GitHub Copilot Pro at $10.

Sources