AI coding agent pricing only becomes comparable after you separate three layers: the subscription or seat, the usage included with it, and the price or behavior after that allowance runs out. A $20 plan can mean shared product usage, a daily and weekly quota, two model pools, or a monthly credit allowance. Those are different commercial contracts, not equal quantities of completed coding work.
This comparison covers OpenAI Codex, Claude Code, Devin, Cursor, and GitHub Copilot. It is based on the vendors' official pricing and billing documentation checked on September 10, 2026. It is not a firsthand same-task benchmark, and it does not rank code quality, autonomy, speed, or value for money.
Editorial review: Agent.Space Editorial checked the official sources linked in the price table and vendor-specific guides. The dated snapshot is a documentation review, not an invoice or independently reproduced performance test.
The short answer: compare three price layers
Use this order before choosing a plan:
- Subscription or seat: the recurring amount required to enter the product, including annual commitments or team minimums.
- Included allowance: what usage the plan makes available, how it resets, which surfaces share it, and whether the model changes consumption.
- Overage: what happens after the allowance is exhausted—stop, wait for a reset, buy credits, enable on-demand use, or pay an API rate.
Then add one operational question: what product are you actually buying? An editor with an agent, a local terminal agent, a vendor-hosted cloud worker, and a GitHub-native agent can all produce code, but they do not replace the same part of a workflow.
The table below is therefore a billing map, not a ranking. “Included” never means a fixed number of accepted pull requests.
Pricing snapshot checked September 10, 2026
Prices below are the public list prices shown on the linked official pages. Taxes, currencies, regional availability, annual commitments, contracts, model choice, and legacy billing can change what a particular account pays.
The important comparison is not the number in the second column alone. It is the complete row, plus the exact model and product surface you intend to use. For a closer look at what included allowances might buy at API prices, see the token-value analysis of four flagship AI subscriptions, which separates the author's full-quota estimates from the value of actual usage.
For a deeper explanation of OpenAI's multiple plan, credit, and API paths, use the separate OpenAI Codex pricing guide. The corresponding Claude Code billing guide keeps Anthropic's personal, team, Usage Credit, and API routes separate.
Do not merge every paid continuation option into a single “credit” column. OpenAI's paid weekly reset immediately pulls the normal weekly Work and Codex allowance forward and changes the next weekly schedule; it is not stored credit. Anthropic's usage bundles add Usage Credits that are consumed only after included plan usage. Both depend on the live account, and neither is interchangeable with an API project's balance.
Why the same monthly price does not buy the same thing
Codex Plus, Claude Pro, Devin Pro, and Cursor Pro all show a $20 monthly price in this snapshot. That does not make them four equivalent bundles.
- Codex ties coding-agent use to a broader ChatGPT plan and a shared Codex/ChatGPT Work allowance. It also has a distinct API-key route with a narrower set of surfaces.
- Claude Code draws from the applicable Claude plan when authenticated with a subscription identity; Console API authentication creates a different bill.
- Devin Pro defines a daily and weekly quota that covers Devin sessions, its terminal route, and Devin Desktop, with on-demand credits after the quota.
- Cursor Pro combines an AI-oriented editor with Agent and Cloud Agent access, then separates included use into two model pools.
- GitHub Copilot combines IDE, CLI, GitHub, and agent features under an AI Credit system, while excluding paid-plan completions and next edits from that meter.
Even the word “task” is unstable across products. A cloud session that checks out a repository and opens a pull request is not the same billing object as one IDE chat, one terminal turn, one model request, or one accepted change. Vendors can also change how model multipliers, tools, caching, or background work consume an allowance.
This is why a price table should preserve each vendor's native unit instead of inventing a conversion such as “one Devin session equals ten Codex messages.” No official source establishes that equivalence.
Layer 1: normalize the subscription or seat
Start a budget sheet with the fixed commitment. Record these fields rather than copying only the headline price:
The Devin Teams minimum shows why this matters. The current documentation says every Teams account costs at least $80 per month. A full seat costs $40 and counts toward the floor: zero full seats leaves $80 as shared prepaid credits; one seat leaves $40; two seats satisfy the minimum; three seats cost $120. Writing “$80 plus $40 per seat” without this arithmetic would overstate some configurations.
Annual prices need the same care. Claude Pro's $17 figure is a monthly equivalent for a $200 upfront annual subscription, not a month-to-month price. OpenAI and Anthropic team prices also differ between annual and monthly billing. Keep cash timing and commitment length visible in the comparison.
Layer 2 and 3: record allowance and overage separately
After the fixed cost, create one row for the included allowance and a second row for excess use.
For the allowance, record:
- the vendor's exact unit: messages, quota, credits, model pool, tokens, or another meter;
- every reset window, such as rolling, daily, weekly, or monthly;
- which products and people share the same pool;
- whether model choice, task length, cloud execution, review, or automation consumes it differently;
- what the account UI currently reports as remaining usage.
For overage, record:
- whether extra use is available on the plan;
- its unit price or linked rate card;
- whether auto-reload is enabled and who controls the ceiling;
- whether purchased credits roll over or expire;
- what happens when both the included allowance and paid balance are exhausted.
Write unknown when a vendor does not publish a stable task quantity. Do not replace an unknown with a third-party estimate and present it as a plan guarantee.
If the overage route is token-based, calculate it from every model call, cache event, output, tool charge, and retry in the accepted task. The separate guide to calculating coding-agent API cost provides that formula. Do not mix an API token estimate into a subscription table unless the plan explicitly bills that route at API rates.
Three buying scenarios
You already pay for an eligible product
Begin with marginal cost. If an existing ChatGPT, Claude, Cursor, or Copilot plan includes the exact agent surface and enough capacity, the next bounded task may not require another subscription. Check the active identity, live allowance, and overage setting before assuming the marginal cost is zero.
This is an adoption advantage, not evidence that the included agent will produce the best result for your repository.
You are a high-usage individual
Compare the higher tier with real overage from your current tier. A larger fixed plan is easier to budget, but it can waste money if usage is uneven. On-demand credits can fit bursts, but an unnoticed auto-reload can make the total unpredictable.
Use at least one full billing cycle of account data when possible. Record accepted outcomes, retries, and time waiting for limits—not just prompts sent.
You are buying for a team
Seat price is only the first line. Add:
- how many people need full daily access versus occasional access;
- whether allowances are individual or pooled;
- repository, SSO, audit, policy, and data-control requirements;
- review, automation, and cloud-task consumption;
- unused seats, shared-credit risk, and overage approval.
A $19 or $20 seat and an $80 account minimum answer different procurement questions. Model a small, normal, and burst month before choosing the contract.
What this comparison cannot tell you
Official pricing can tell you how vendors describe plans and meters. It cannot tell you which agent will finish your task with the least total cost.
That requires a controlled workload: the same repository state, a comparable task, explicit permissions, deterministic acceptance checks, and complete accounting for retries, repairs, reviewer time, and failed work. Product surfaces may make a perfectly identical test impossible, so document the remaining differences instead of hiding them.
This article therefore does not claim:
- that the cheapest subscription produces the cheapest accepted result;
- that a higher allowance improves code quality;
- that a credit, quota unit, message, or session has a stable cross-vendor conversion;
- that any product is universally better for solo developers or teams;
- that all five products are available inside Agent.Space.
When quality, autonomy, or speed decides the purchase, run that separate benchmark. Do not infer a winner from the billing table.
The takeaway
AI coding agent pricing is comparable only after you preserve the differences. Record the subscription or seat, the included allowance, and the overage path as separate layers. Add the product surface and billing owner, then model your own normal and burst usage.
In this September 10, 2026 snapshot, the public entry prices range from Codex Go at $8 and Copilot Pro at $10 to several $20 individual plans, but those figures buy different workflows and meters. Team pricing is even less interchangeable because per-seat plans, account minimums, shared pools, and governance features solve different problems.
If your comparison leads to an Agent and model combination currently supported by Agent.Space, start one bounded Agent.Space task with a clear acceptance check. Treat the result as your evidence; do not treat the sticker price as a benchmark.
FAQ
Which AI coding agent has the cheapest paid plan?
Among the dollar-denominated entries shown on the official pages in this snapshot, Codex Go is $8 per month and GitHub Copilot Pro is $10. Cursor also lists an India-only Start plan at ₹649. These are not equivalent bundles, and availability, tax, included use, and surfaces still need to be checked on the buyer's account.
Are all $20 AI coding plans equivalent?
No. Codex Plus, Claude Pro, Devin Pro, and Cursor Pro each use different allowances, products, models, reset windows, and overage rules. Compare the complete billing structure, not only the recurring price.
What is the lowest team price in this comparison?
GitHub lists Copilot Business at $19 per granted seat per month, but its plans page currently directs new customers to contact sales while new self-serve purchases remain temporarily paused. OpenAI Business and Claude Team Standard list $20 per user or seat per month with annual billing, while Cursor Teams Standard is $40 and Devin Teams has an $80 account minimum. Those plans include different controls and usage systems, so the lowest seat number is not a complete team-cost answer.
How should I compare usage credits and quotas?
Keep each vendor's native unit, reset window, shared-pool rule, and overage price. Do not convert them into a universal task count unless you have a controlled workload and your own measured data.
How often should this pricing comparison be refreshed?
Recheck it before every purchase and whenever a vendor changes plan names, prices, included usage, credit rules, team minimums, model multipliers, or product surfaces. Every numeric statement on this page is a dated September 10, 2026 snapshot, not a permanent promise.