Cursor vs Windsurf vs GitHub Copilot (2026): What One Month of Coding Actually Costs

Disclosure: this post contains affiliate links. If you buy through them, I may earn a commission at no extra cost to you. Every price below was read from the vendor's own page on 21 November 2026 and is quoted in USD before tax.

Every coding tool I pay for changed its meter in 2026. Cursor folded itself into The X Subscription, where one subscription covers Grok, Cursor and X. Windsurf stopped being Windsurf on 2 June 2026 and now ships as Devin Desktop under Cognition, with a quota measured in tokens instead of prompt credits. GitHub Copilot moved everyone onto GitHub AI Credits, where one credit is exactly one cent.

So the question "which coding agent is worth $20?" no longer has an answer, because none of these three is billed in the same unit. One sells a weekly usage pool. One sells a daily and weekly token quota. One sells dollars-as-credits. This post prices the same month of work through all three, using each vendor's published rates, and shows where the bill actually comes from.

Short version: at 2026 token prices, eighty agent tasks cost between $1 and $32 of model usage. The plan you buy is mostly paying for features, not tokens — unless you run frontier models through a vendor that doesn't include them, and then the tokens are the whole bill.

What actually changed in 2026

Three structural changes matter more than any feature list.

Windsurf is now Devin Desktop. Cognition shipped the rename as a regular over-the-air update on 2 June 2026: same IDE, same extensions, same keybindings, with the Agent Command Center as the main view. Cognition's FAQ states that plans and pricing did not change, and existing subscribers kept their old prices — $15/month Pro and $30/seat Teams are grandfathered indefinitely, while new buyers see $20 Pro and $200 Max. In March 2026 the same product had already replaced prompt credits with a quota system measured by token cost, so "1 prompt = 1 credit" is gone. If you are comparing blog posts from last year, half of them describe a product that no longer has that name or that meter.

Cursor is a plan inside a bigger subscription. New customers no longer buy "Cursor Pro". They buy The X Subscription: Free, Plus at $8, Premium at $30, Super at $100, Ultra at $200, and Teams seats at $30, $100 or $200 per seat. Paid plans include one weekly pool of usage for Cursor's own models — Grok 4.7, Grok 4.6, Grok 4.5 and Composer 2.5 — shared across Cursor, the Grok apps and the Grok Bot. Third-party models are not in that pool at all. Legacy Cursor plans (Pro at $20, Pro+ at $60, Teams Standard at $40/seat) still work for people who had them.

Copilot converted its allowance into money. GitHub AI Credits are $0.01 each, and every plan states its allowance as a credit count: Pro gives 1,500, Pro+ gives 7,000, Max gives 20,000, Business gives 1,900 per seat, Enterprise gives 3,900 per seat. Code completions and next-edit suggestions are not billed in credits and stay unlimited on paid plans. That makes Copilot the only one of the three whose pricing page tells you what your plan buys in dollars of model usage.

A solo builder desk at night with a code diff on two monitors

Eighty agent tasks a month is a real day-job workload. At 2026 list prices, the tokens behind it are the cheap part.

The workload I priced — and why these numbers are comparable

Vendors meter differently, so the only honest common denominator is tokens. Cognition's own documentation publishes a worked example: a single mid-size refactor conversation in Devin Desktop, start to finish, consumes about 200,000 tokens across planning, context gathering and three tool calls. I used a deliberately ordinary task profile just under that:

  • One agent task = 160,000 input tokens (40% fresh, 60% served from cache) + 12,000 output tokens.
  • Month A — side project: 40 tasks (2 a day, 20 days).
  • Month B — full-time solo builder: 80 tasks (4 a day).
  • Month C — agent-heavy: 240 tasks (12 a day, parallel agents).

Cache ratio is the variable that moves these numbers most. Agents re-read the same files all day, so a setup that keeps one model and a stable context cheapens every later request; hopping between models throws the cache away and pushes the fresh-input share up. I used 60% cache reads because that is what a steady single-model session looks like, and flagged it rather than hiding it.

Rates are the vendors' own published list prices on 21 November 2026. Cursor publishes token rates for its own models and for third-party models; Devin publishes per-model token rates for extra usage; Copilot publishes per-model rates and states that the resulting cost is converted into credits at one cent each. The arithmetic is identical across all three, which is the point.

The plans, as they stood on 21 November 2026

Read this table as "what you are buying", not "what it costs per token". Note the third-party column: it decides the bill for anyone who wants Claude or GPT rather than the vendor's house model.

VendorPlanMonthlyAnnualIncluded usageThird-party models
Cursor
(The X Subscription)
Free$0—Limited agent and chat requestsNot available
Plus$8$80Base weekly pool; Grok 4.5–4.7 + Composer 2.5 only, fixed settingsNot available
Premium$30$3004× Plus pool; adds Cloud Agents, Automations, Bugbot, code hostingAt API list price, from prepaid credit
Super$100$1,00015× Plus pool; priority model access at peakAt API list price, from prepaid credit
Ultra / Teams$200 / $30, $100 or $200 per seat$2,00050× Plus pool; Teams seats match the tierAPI list price (+ $0.25/M Cursor Token Rate on Teams)
Devin
(ex-Windsurf)
Free$0—Daily + weekly quota; SWE-2 Flash Medium; unlimited Tab and inline editsLimited model list
Pro$20Not publishedDaily + weekly quota; full model access; Devin Cloud agentsExtra usage at API list price
Max$200Not publishedLarger weekly quota, no daily cap; Devin BotsExtra usage at API list price
Teams$80 flatNot publishedShared prepaid credit pool; up to 100 users; SSO; 2 free cloud sessions per memberExtra usage at API list price
GitHub
Copilot
Free$0—2,000 completions; limited chat and agent; auto model onlyNot available
Pro / Pro+ / Max$10 / $39 / $100None published1,500 / 7,000 / 20,000 credits ($15 / $70 / $200 of usage)Yes, including Claude Code and Codex agents
Business$19 per seatNone published1,900 credits per seat ($19)Yes, under admin policy
Enterprise$39 per seatNone published3,900 credits per seat ($39)Yes, plus priority access

Two things jump out of that table before any math. Cursor's cheapest paid plan cannot use a single third-party model — an $8 Plus seat is a Grok-and-Composer plan. And only Cursor publishes an annual price; Copilot and Devin list monthly only, so the 16.7% annual discount Cursor gives ($300 instead of $360 for Premium) has no counterpart elsewhere.

The same month of work, priced three ways

Here is the core table. Every row is the identical workload — same 172,000 tokens per task — billed at that vendor's published rate for that model. Nothing here includes plan prices yet; this is what the tokens themselves cost.

Plan / modelPer task40 tasks80 tasks240 tasksNote
Devin — SWE-2 Flash (list, free to 4 Dec)$0.013$0.53$1.07$3.21cheapest by 6×
Cursor — Composer 2.5$0.081$3.25$6.50$19.49house model
Devin — Gemini 3.8 Flash (intro rate)$0.100$4.01$8.02$24.05intro to 31 Dec
Copilot — GPT-6.1 Terra$0.129$5.15$10.30$30.91versatile tier
Cursor — Grok 4.7$0.248$9.92$19.84$59.52house frontier
Copilot — Claude Sonnet 5$0.267$10.69$21.38$64.13workhorse
Copilot — Gemini 4 Neon$0.386$15.46$30.91$92.74premium
Devin — SWE-2 (list rate)$0.401$16.03$32.06$96.19free until 30 Nov

The spread is 31× between the cheapest and most expensive way to run the same eighty tasks. That is not a promotion artefact; it is the difference between a small in-house model and a frontier one, plus one vendor currently giving its own model away.

Note the output column: output tokens are where the money is, and their prices run from $0.50 to $15 per million across these rows. Input is nearly free by comparison, which is why context discipline pays more than prompt golf.

What a plan's allowance actually buys

This is where the three vendors stop being comparable. Copilot converts credits to dollars on the pricing page. Cursor and Devin publish quota as "base", "4× Plus", "increased" or "significantly higher" — no dollar figure and no token figure, so you cannot compute coverage before you buy.

PlanPrice / monthUsage includedTasks on GPT-6.1 TerraTasks on Claude Sonnet 5
Copilot Pro$10$15.00 (1,500 credits)11656
Copilot Pro+$39$70.00 (7,000 credits)543262
Copilot Max$100$200.00 (20,000 credits)1,553748
Copilot Business (seat)$19$19.00 (1,900 credits)14771
Copilot Enterprise (seat)$39$39.00 (3,900 credits)303146
Cursor Plus / Premium / Super / Ultra$8 / $30 / $100 / $200Not published in dollarsunknownunknown (Premium+)
Devin Pro / Max$20 / $200Not published in dollarsunknownunknown

Now the practical question: for each month size, what is the cheapest plan that covers the work?

40 tasks a month. On Sonnet 5 that is $10.69 of tokens. Copilot Pro at $10 includes $15, so it covers the month with about $4 of headroom. Cursor Premium costs $30 and Devin Pro costs $20 for the same work, and neither will tell you whether the pool covered it until you look at the usage meter.

80 tasks a month. Sonnet 5 costs $21.38 — more than Copilot Pro's $15 allowance, so you either upgrade to Pro+ at $39 or keep Pro and set a usage budget: $10 + $6.38 of overage is $16.38, still less than Pro+. If you build on house models instead, the same month is $6.50 of Composer 2.5 or $10.30 of GPT-6.1 Terra, both comfortably inside cheap plans.

240 tasks a month. Sonnet 5 costs $64.13. Copilot Pro+ includes $70, so $39 covers it. Cursor Super is $100 and Devin Max is $200 for the same month of tokens — worth it only if you need cloud agents, review bots, parallelism or the priority lane, not for the tokens.

The reversal is real and worth sitting with: the cheapest way to run a heavy agent month on a frontier model is a $39 plan whose included usage is stated in money, and the most expensive way is buying a bigger plan from a vendor that states it in multipliers.

Chart comparing cost per agent task and monthly cost across Cursor, Devin and GitHub Copilot

Left: cost of one agent task by model. Right: what each Copilot plan converts to in usage dollars. Bottom: the same month at 40, 80 and 240 tasks.

Five line items that are not on the pricing page

1. Devin's free flagship model ends 30 November 2026. SWE-2 is free on self-serve plans in Devin Desktop and Devin CLI through 30 November; list pricing of $3.00/1M input and $15.00/1M output applies afterwards. On the profile above, that turns an 80-task month from about $0 into $32.06. SWE-2 Flash is free until 4 December at 18:00 UTC, then $0.10/1M input and $0.50/1M output — still cheap, but no longer free.

2. Copilot's discounts are dated. Paid plans get 30% off model costs when using Copilot Fusion through 30 November 2026, and 10% off with auto model selection. If you are benchmarking this month, you are measuring a temporarily discounted bill.

3. Introductory model pricing expires under you. Gemini 4 Carbon is at $2/$10 per million until 1 December 2026 and doubles to $4/$20 afterwards. Gemini 3.7 and 3.8 Flash sit at $0.75/$3.75 as introductory pricing through 31 December 2026, then $1.50/$7.50. Devin's own docs list these dates; the pricing pages summarise them.

4. Completions are free; agents are not. Copilot bills chat, CLI, cloud agent, Spaces and third-party agents in credits, but code completions and next-edit suggestions are unlimited on every paid plan. Devin gives unlimited Tab completions and inline edits even on Free. So the "is autocomplete worth $10?" question is answered: on all three, the completion layer is not what you are paying for. The agent layer is.

5. Teams is priced three different ways. Copilot charges $19 or $39 per seat and pools credits at the billing entity. Cursor charges $30, $100 or $200 per seat, and adds a $0.25 per million token Cursor Token Rate on eligible third-party requests. Devin charges a flat $80 a month for a team of up to 100 users, funded as shared prepaid credit, SSO included. For a three-person studio, that is $57 (Copilot Business) versus $90 (three Cursor Premium seats) versus $80 flat — and only one of those gets cheaper per person as the team grows.

What I would actually pay for, as one person

If I were starting today with no legacy plan, I would not buy the $200 tier of anything. The token math says the work I do in a month is worth $6 to $65 at list prices, and the expensive tiers are selling concurrency and review automation, not tokens.

My starting stack would be one $10–$39 seat that states its allowance in dollars, plus one cheap house model for the boring tasks. Copilot Pro at $10 with a $10 additional-usage budget gives $25 of monthly model usage with a hard ceiling — enough for 80 Sonnet-class tasks or 190 GPT-6.1 Terra tasks — and completions stay free. If I wanted an agent command center and local-plus-cloud handoff, I would pay Devin Pro's $20 for the editor and keep the heavy lifting on SWE-2 Flash while it is free, with a note in the calendar for 4 December.

Cursor is the one I would buy for a specific reason rather than by default. Composer 2.5 at $0.50/$2.50 per million is the cheapest house model of the three by a wide margin, and Cloud Agents plus Bugbot on Premium are genuinely useful if you ship continuously. But if your plan says "I want Claude Sonnet in my editor", Cursor Plus will not do it, and Premium bills those tokens outside the pool at list price. That is a $30 subscription plus a metered bill, and you should budget both.

One more discipline that beats plan-shopping: keep one model per task type. Cache reads are 10–20% of input price on every vendor here, and switching models mid-session throws that away. Scoping a task tightly and keeping the session short cuts more off the bill than any discount table.

After the code ships: the audio and email layer

Most solo builders I know pay for a coding agent and then hand-write the changelog, the newsletter and the demo voiceover. Those are the same content in three shapes. Once a build is done, the release notes can be read aloud as a short audio update — useful for a podcast feed or a client walkthrough — and gated behind an email capture so the people who care about the next release hear about it.

That is the two-tool layer I use: ElevenLabs for narration (Starter is $6 a month for 30,000 credits, $60 if billed annually) and systeme.io for the funnel and list (free to 2,000 contacts, three funnels, unlimited sends, 0% transaction fee on every plan). Compared with the coding bill above, both are rounding errors — and unlike a $200 agent tier, they produce an asset you keep. I walked through the whole pipeline in One Article, Two New Assets, and the funnel comparison is in Systeme.io vs MailerLite vs Kit.

If your automation layer is where the money leaks, the same method applied to workflow tools is in Zapier vs Make vs n8n; if you are pricing the assistant you talk to rather than the one that writes code, see Claude vs ChatGPT vs Gemini for work. And if the agent is spending its day in your notes and docs, the second-brain cost comparison and the meeting-notes comparison use this same token-and-minute arithmetic.

FAQ

Is Windsurf gone?

The name is. On 2 June 2026 Windsurf became Devin Desktop under Cognition — same IDE, same extensions, same keybindings, delivered as an over-the-air update. Cognition states that plans and pricing did not change for existing subscribers, and legacy $15 Pro and $30/seat Teams plans are grandfathered. New buyers see the Devin plans: Free, Pro $20, Max $200, Teams $80 flat, Enterprise custom.

Which of the three is cheapest for one person?

It depends on the model, and the table above is the answer: $1.07 a month for 80 tasks on SWE-2 Flash, $6.50 on Composer 2.5, $10.30 on GPT-6.1 Terra, $21.38 on Sonnet 5, $32.06 on SWE-2 at list price. The cheapest plan that covers it is usually Copilot Pro at $10 with an overage budget, because it is the only one where you can compute coverage before buying.

Do autocomplete suggestions use my credits?

No. Copilot bills chat, CLI, cloud agent, Spaces and third-party agents in credits, while code completions and next-edit suggestions are unlimited on paid plans and capped at 2,000 a month on Free. Devin gives unlimited Tab completions and inline edits on every plan, Free included. Completions are the free layer everywhere; agents are the metered one.

Can I use Claude or GPT inside Cursor's $8 plan?

No. Plus includes only Cursor's own models — Grok 4.7, 4.6, 4.5 and Composer 2.5 — at fixed settings, without Auto, Cloud Agents, Automations, Bugbot, third-party models or your own API key. Third-party models start at Premium, and they are billed at API list price from prepaid credit rather than drawing on the included pool.

What happens when I hit the limit?

Cursor: the weekly pool refreshes, or you keep working from prepaid credit, which stays available for 365 days. Devin: on-demand credits, which roll over and never expire; Pro and Max refresh a daily and weekly quota, Max weekly only. Copilot: extra usage at $0.01 per credit against a dollar budget you set, with alerts at 75%, 90% and 100%, or upgrade — the upgrade charges only the plan difference and applies the bigger allowance immediately.

Is annual billing cheaper?

Only Cursor publishes an annual option: $80 Plus, $300 Premium, $1,000 Super, $2,000 Ultra, which is 16.7% off — two months free. As of 21 November 2026, neither Copilot nor Devin lists an annual price for individual plans.

What is the cheapest way to run 240 agent tasks a month?

Copilot Pro+ at $39 includes $70 of usage, which covers 262 Sonnet-5 tasks at list price — the cheapest frontier-model route in this comparison. If a Flash-tier model is good enough, Devin's SWE-2 Flash is free until 4 December 2026 and about $3.21 a month afterwards. If you want the cheapest house model with a full IDE, Cursor's Composer 2.5 prices the same month at $19.49 of tokens, which any paid Cursor tier almost certainly covers.

Bottom line

None of these three tools is expensive because of its tokens. Eighty agent tasks are $6 to $32 of model usage at 21 November 2026 rates. They are expensive because the plan tiers bundle concurrency, cloud agents, review automation and premium-model access, and because two of the three vendors refuse to tell you what the included quota is worth until the month closes.

So buy in this order: first decide which model you actually want to work in, because that sets the token bill; then check whether the vendor includes that model in the plan or bills it separately; then pick the smallest tier whose allowance, in dollars, covers your month — and set a hard budget on the metered part. The vendor that lets you do that arithmetic before you pay is, this month at least, the one charging $10.

Sources

  • Cursor pricing and plan limits — cursor.com/pricing, cursor.com/docs/models-and-pricing, cursor.com/help/account-and-billing/pricing (read 21 Nov 2026)
  • Devin plans, quotas and model rates — devin.ai/pricing, docs.devin.ai/admin/billing/self-serve, docs.devin.ai/desktop/accounts/quota, docs.devin.ai/desktop/models, docs.devin.ai/desktop/devin-desktop-faq (read 21 Nov 2026)
  • GitHub Copilot plans, AI credits and per-model rates — github.com/features/copilot/plans, docs.github.com/en/copilot/get-started/plans, docs.github.com/en/copilot/concepts/billing/usage-based-billing-for-individuals, docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing (read 21 Nov 2026)
  • Task profile: 160k input / 12k output per agent task, anchored on Cognition's published 200k-token refactor example; 60% cache-read share is my assumption and stated as such.

Comments

Popular posts from this blog

Claude vs ChatGPT vs Gemini for Work: Which AI Assistant Saves You More Time in 2026?

Same Video, Five Languages: What AI Dubbing Really Costs in 2026 (ElevenLabs vs Rask AI vs HeyGen)

Systeme.io vs MailerLite vs Kit: Best Free Funnel & Email Stack for Solo Creators (2026)