Claude Code cost management for teams: Cursor and Codex in 2026

SpaceXSI introduced The X Subscription on October 12, 2026; new Cursor team seats start at $30 per user per month. To manage Claude Code, Cursor and Codex costs, inventory seats and API billing paths, set native and gateway limits, and reconcile spend against accepted work. We make Requesty, a gateway for the configured API traffic in this guide.

How Claude Code, Cursor and Codex charge a team

The editor, coding agent, model and bill are separate choices. Claude Code accepts a Claude Team login or an API credential; local Codex also has separate ChatGPT and API authentication paths.

Prices are US monthly rates as of October 2026, before tax. Annual billing changes the Claude Team and ChatGPT Business seat prices. Codex and ChatGPT Work share usage.

Cursor's October plan change

The X Subscription became available to new customers, with switching for existing subscribers rolling out afterward. Existing teams can keep their legacy plans: Cursor Teams Standard costs $40 and Premium costs $120 per user per month. Switching is permanent.

Legacy paid Teams seats include an Other Models pool. On The X Subscription, a member's weekly pool is shared between Cursor and Grok, but does not transfer to another member. Enterprise has custom pricing and pooled usage. See X Subscription usage limits.

OpenAI has proposed November 12, 2026 as the shutoff date for its direct model supply to Cursor. OpenAI says the official date remains to be confirmed, and Cursor could end access sooner. It describes API keys and gateway credentials as separate options for supported local requests; test the intended model and feature.

Where a gateway fits

A gateway records and controls requests sent through its endpoint. A Claude Code gateway credential replaces the subscription credential for those calls, which are billed separately. Keep a subscription path for engineers who need Claude Code's Chrome integration or Remote Control. Anthropic-hosted cloud and Slack sessions use Anthropic's API outside the local gateway configuration.

For Codex, API-key authentication is separate from ChatGPT subscription access. The Requesty API-key path does not spend the ChatGPT allowance; Codex cloud requires ChatGPT sign-in. OpenAI also documents a different gateway configuration that forwards a ChatGPT credential upstream.

Build one monthly cost ledger

Add seats and usage billed outside those seats. Record the vendor, period, team, seat or credential owner, charge type and whether a charge is included in another line.

Assume 20 engineers each have Claude Team Standard, Cursor Teams Premium on The X Subscription, and ChatGPT Business Standard. All seats are retained and billed monthly. The following usage amounts are assumptions for the same period. Requesty usage is separate from the three credit lines; no Claude Platform API credits are applied.

Seat inputs come from Claude Team, Cursor Teams and ChatGPT Business. Value consumed ChatGPT credits using their purchase cost and agreement; credit units are not automatically dollars. This is a management view of consumption. Finance should apply its own accounting policy.

Assume the Requesty line is $400 of model usage paid with standard credits. Allocate the 5% fee paid when those credits were purchased: $400 x 1.05 = $420. Requesty charges the fee at top-up, so payment could fall in a different month. Track that payment and the unused balance separately; do not add the purchase and consumption together.

For an eligible Cursor BYOK token fee, record the paid amount separately from gateway inference. On The X Subscription, it draws from prepaid credit. On legacy Teams, it can draw from the included Other Models allowance first, so add only the amount paid beyond the seat allowance. See Cursor's token-rate guide and BYOK billing guide.

Cursor's Admin API distinguishes spend excluding included usage from overallSpendCents, which includes it. Do not add the latter to seat fees as another invoice charge. For Requesty provider BYOK, reconcile the provider invoice with Requesty analytics, count provider inference once, and add any separately charged Requesty fee only if it is not already included. Apply your BYOK agreement's terms.

Set native controls before rollout

Claude Code: seats and usage credits

Every Claude Team seat includes Claude Code. Under Organization settings, then Usage, owners enable or disable usage credits, select members and set organization and per-user monthly caps. Eligible usage beyond the included allowance draws from credits at API rates. Depending on the seat, Fable can use credits instead of included limits; the /model picker identifies when credits are required. Anthropic says a cap can be exceeded slightly because accounting happens after a request.

Usage-based Enterprise charges from the first token on top of its seat fee; use its organization and individual caps. Claude Console API and cloud-provider access need separate controls. Claude Code's /usage session cost is a locally calculated estimate. Team owners can export an Analytics spend report for usage-credit spend; included seat usage has no dollar amount there. Bedrock, Google Cloud and Foundry usage is absent from Claude Code analytics, so use OpenTelemetry metrics and the cloud bill.

Claude Team's separate Platform API credits

Eligible teams can claim pooled monthly Claude Platform API credits: $20 per Standard seat and $100 per Premium seat, capped at $500. The 20 Standard seats above could claim 20 x $20 = $400. Credits cover self-run claude -p with an API key from the linked Console organization, but exclude interactive Claude Code and usage-credit overages. The ledger's Requesty usage uses Requesty credits instead.

Cursor: limit additional usage

For The X Subscription, admins manage the prepaid balance, Auto Top-up and team monthly additional-usage limit under Additional usage credits at the billing dashboard. X Subscription spending caps stop paid additional usage when the team limit is reached or the balance is empty.

Legacy Teams on-demand usage is on by default. Turn it off or set a team-wide limit. The legacy spending guide documents an admin-only setting to prevent members from changing usage settings. Cursor says enforcement is not instantaneous.

Cursor's token-rate guide lists a $0.25 per million input, output and cached tokens fee on eligible third-party requests by Teams and Enterprise, including eligible BYOK requests. Cursor models, Cost mode and the named third-party models in that guide are exempt.

When you put a Requesty key in Cursor, it leaves the editor. Cursor says it sends the key to its backend over an encrypted connection with each request and does not persist it after the request completes. Cursor also says its Zero Data Retention policy does not apply to BYOK. Review Cursor's, Requesty's and the selected provider's data handling before rollout.

Codex: inspect allowances and workspace credits

Run /status in the Codex CLI to see usage and resets, or open the Codex usage page. Business owners buy shared credits under Workspace settings, then Billing. Set a monthly recharge limit if automatic reload is enabled: a blank limit permits unlimited automatic reload purchases. Owners and admins can set monthly credit limits by seat type and add per-user overrides; limits are unspecified by default. Credit limits govern additional usage, while the recharge limit governs purchases. See OpenAI's Business credit controls. When shared credits run out after included usage, the affected feature is blocked.

For credit-based Enterprise and Edu, set the workspace overage limit under Admin Console, then Billing. OpenAI says it defaults to No limit. Set it to 0 to stop new credit-consuming requests after the committed pool is exhausted; in-progress work can settle afterward. Set per-person defaults and overrides under Usage limits. Some Enterprise agreements use USD token billing instead: set a monthly workspace budget under Billing and user limits under Usage limits. An old credit overage limit does not become a USD budget. API Platform spend is separate, as OpenAI's token-billing guide explains.

Add Requesty for configured API traffic

Give each engineer a distinct Requesty key for API-billed Claude Code and Codex work, cap the keys and review their usage in one place. Selected Cursor BYOK chat can use the same gateway. ZoomInfo's case study describes 1,300+ engineers routing AI access through Requesty, including Claude Code, GitHub Copilot, coding agents and internal apps. Provider credentials stay in the gateway while engineers use Requesty keys.

Decide who needs dashboard access. Enterprise admins can invite members from the console or use SSO for automatic enrollment. For the pay-as-you-go member-joining workflow, contact Requesty support. Pay-as-you-go members with console access can see organization-wide analytics and logs; Enterprise scopes those views by role.

1. 1
   Set user limits and issue keys

   For engineers who have joined the organization, keep their role as Member. In Admin Panel, then Users, set an individual monthly limit for each engineer and a default for new members. Members cannot change admin-set user limits. Have each member create a personal key on the API Keys page and set its monthly limit. In Analytics, Group By, Member identifies the key owner.

   For engineers without console access, an admin creates a separately named key, caps it and delivers it through an approved secrets channel. Reconcile by key; the recipient is not necessarily the key's owner or an organization member.

   Members can create and modify personal keys, so keep the cap on the user as well as on each key.

2. 2
   Separate automation and shared budgets

   For each automated workload, create an account under Admin Panel, Service Accounts, then Create Service Account. On API Keys, select Create New API Key, choose Service Account, select the account and create its key. Set a monthly key limit and deliver the credential through the automation's secrets manager. Service-account usage is separate from individual user attribution.

   Requesty groups require Enterprise. For a shared engineering budget, ask support to enable Group Budget mode. In the default global mode, a group limit is granted to each member, not shared across the group. See group budget modes. Requesty also has an organization-wide cap; ask support about setup, plan availability, reset timing and in-flight behavior.

3. 3
   Set alerts and approve models

   In Admin Panel, then Alerts, configure a Slack, Teams or JSON webhook and add key or user thresholds. Start with 80% and 95% as warning points. Alerts notify an owner; limits control spending.

   Configure Approved Models and an access list where a workload needs a narrower selection. A list attached directly to a key overrides its group list. Test an allowed and a disallowed model with a member-created key, including after any permitted key-list change. Confirm precedence with support before relying on it as an organization-wide boundary.

   Tell engineers who can approve a higher limit and what to do while they wait.

Requesty charges a 5% fee on top of prepaid credit top-ups. Pay-as-you-go has no seat fees or minimums; Enterprise pricing is custom. For paid-model calls, fund the prepaid balance first or use an agreed post-paid arrangement. The free tier covers free models. See Requesty pricing.

Connect each client

For Claude Code, point ANTHROPIC_BASE_URL at https://router.requesty.ai and set ANTHROPIC_AUTH_TOKEN to the engineer's key; follow the Claude Code guide.

For Codex, configure a custom model provider with https://router.requesty.ai/v1 and the Requesty key; follow the Codex integration guide.

For supported Cursor chat, set the OpenAI base URL override to https://router.requesty.ai/v1 and use the engineer's Requesty key where Cursor permits BYOK; see the custom endpoint guide and Cursor integration. Cursor Tab uses built-in models, Auto may take a different path, and Requesty's documented setup covers selected chat traffic. Reconcile other paths in Cursor's billing records; see also OpenAI's Cursor guidance.

Send a test request from each configured client and confirm it appears under that engineer's key in Requesty analytics; an admin can check for engineers without console access.

Reduce spend without hiding rework

A long session can resend earlier history and tool output. Prompt caching makes eligible repeated prefixes cheaper, but a model switch or change to tool definitions can cause a cache miss. Anthropic's prompt-caching guide explains the mechanics.

Give an agent a bounded goal, relevant files and a stopping condition. Inspect large tool results and repeated calls before adding parallel workers. Anthropic reports that agent teams in plan mode use approximately seven times the tokens of a standard session. In Claude Code, use /context to inspect the active context, /clear for an unrelated task, or /compact with instructions about what to preserve when the task continues. See Anthropic's context guidance.

For a model change, record model, billed cost, retries, test outcome, acceptance and reviewer time on the same task classes. Divide relevant tool and inference cost by accepted changes for a defined period. Review time belongs next to that number: a cheaper first attempt that needs more correction can cost the team more overall.

A pilot with two model settings

Assume two groups each work for one month on the same 12 representative tasks: six bug fixes and six test additions. Use the same starting revisions, tests and acceptance criteria. Each group has four Claude Team Standard seats dedicated to the pilot, allocated at 4 x $25 = $100. Assume all additional credit consumption below belongs to the pilot and is excluded from the seat charge.

Setting B costs $0.50 less per accepted change but needs four more hours of review and leaves two more tasks unaccepted. Pick A if those extra review hours and unfinished tasks cost more than the $40 difference in total tool spend. If seats also support other work, allocate a documented share to the pilot and keep the remaining seat cost elsewhere in the ledger.

Review the bill and handle exceptions

Assign one engineering owner and one finance owner to the ledger. Each week, inspect the most expensive users and models alongside accepted work. At month end, reconcile the invoice, credit balance and routed usage. For a budget stop, inspect the effective cap and request approval before raising it. For an upstream 429, inspect provider quota and retry behavior before changing a spend limit; see the provider rate-limit guide. For a Requesty in-flight 429, wait for active requests to finish and inspect the organization's concurrency limit.

Requesty organization admins can use the API key security view to inspect the last 30 days of client activity. It identifies clients from User-Agent and X-Title and flags a new tool, a traffic spike of about 3x normal, a key used by four or more tools, and mostly unattributed traffic. A new tool combined with a spike is flagged Critical. A signal alone does not prove a leak. Use Expire now to disable the key while investigating; the action is reversible. Inspect affected requests and issue a replacement if needed.

Revoke credentials and remove a departing person's access. Vendor seat-reduction rules and contract dates determine when those removals affect the bill. See Claude Enterprise billing and Cursor team pricing.

What to do this week

Start with a small group: create a Requesty account, choose clients from the coding agent guides, and assign an owner to the first ledger. For the tool choice itself, read Cursor vs Claude Code; use the October plan prices and prepaid top-up fee described here for budgeting.