Cursor vs Claude Code vs Codex: What Each Costs in 2026
Compare Cursor, Claude Code, and Codex as AI coding tools: plans, usage limits, overages, and how each one bills.
TL;DR: Cursor, Claude Code, and Codex price AI coding differently. Cursor combines a paid editor with included model usage and overages. Claude Code subscriptions use rolling allowances and can draw from prepaid usage credits when extra usage is enabled; direct API or cloud-provider use is billed by tokens. Codex is available through ChatGPT plans and workspace credits.
- Choose Cursor for an editor-first workflow with Tab completion, model choice, and coding agents in one workspace.
- Choose Claude Code for Claude-powered repository work across terminal, desktop, web, and IDE.
- Choose Codex for OpenAI-powered coding across local and cloud surfaces, especially delegated and parallel tasks.
- Use SuperPenguin alongside any of them to compare AI coding spend, usage limits, repository costs, and cost per merged pull request across all three.
How the products differ
Verified against vendor documentation on August 29, 2026. Seat prices are list prices.
Cursor: an AI-native code editor
Cursor is a VS Code-based editor with Tab completion, inline edits, and agent mode in the same workspace. The agent can independently edit code, search the codebase, and run terminal commands using the selected model. Its subscription combines the editor with monthly model-usage pools, then charges for additional usage. Auto is a router rather than a model: it selects a model according to Cost, Balance, or Intelligence mode.
The product is designed for day-to-day development without leaving the editor. Composer and Grok sit beside third-party models including Claude Opus 5, GPT-5.6 Sol, Gemini, and open-weight options such as Kimi K3 and GLM 5.2. Bugbot reviews pull requests, and Cloud Agents run work away from the laptop.
OpenAI model access is changing. OpenAI says it notified SpaceX that it intends to end the contract supplying OpenAI models directly to Cursor on November 12, 2026. Existing hosted models remain available during the proposed notice period, but OpenAI says it will not provide future models to Cursor. Cursor documents OpenAI BYOK only for standard, non-reasoning chat models. Tab still uses Cursor's built-in models. Neither company has said BYOK will replace hosted OpenAI models after the proposed cutoff.
Cursor plans
Cursor does not publish a token count for each individual pool. Pro, Pro Plus, and Ultra share the same product features and differ mainly in how much included usage you get. Cursor recommends Pro Plus for daily Agent users and Ultra for power users.
| Plan | Price | Included usage | What this tier covers | What you pay extra for |
|---|---|---|---|---|
| Hobby | Free | Limited Agent requests; Composer | Editor, no credit card | Frontier models, Bugbot, and extended Agent limits start on Pro. Cloud Agents are on Start and every plan above it |
| Start (India only) | ₹649/mo, tax inclusive | Cursor Models pool only (Grok 4.6, Grok 4.5, Composer 2.5). No Fast mode or effort controls | Cloud Agents, Cursor for iOS, plugins, MCP, hooks, and skills | Other Models pool, on-demand usage, Bugbot, Auto, Automations, and the Cursor SDK. Upgrade to Pro for those |
| Pro | $20/mo | Both monthly pools (Cursor Models and Other Models); unlimited Tab | Frontier models, Cloud Agents, Bugbot on usage-based billing, MCP, skills, and hooks | On-demand at the same API rates when a pool empties |
| Pro Plus | $60/mo | Larger included pools than Pro | Same product features as Pro | On-demand after included usage |
| Ultra | $200/mo | Largest individual pools | Same product features as Pro | On-demand after included usage |
| Teams Standard | $40/user/mo ($32 annual) | Two monthly pools per seat | Team billing, marketplace for rules/skills/plugins, shared Cloud Agents and automations, Bugbot, usage analytics, team-wide privacy mode, SAML/OIDC SSO. Mix Standard and Premium seats | Cursor Token Rate on third-party models; on-demand; team spend limits |
| Teams Premium | $120/user/mo ($96 annual) | 5x Standard Agent usage | Same team features as Standard | Same extras as Standard |
| Enterprise | Custom | Pooled usage | Everything on Teams, plus invoice/PO billing, SCIM, repository/model/MCP access controls, auto-run/browser/network controls, audit logs, service accounts, AI code tracking API, and priority support | Cursor Token Rate; per-user and group spend limits. Until 7 Sep 2026, Enterprise Auto is billed per million tokens regardless of model |
Included usage arrives as two monthly pools, one for Cursor's own models (Grok 4.6, Grok 4.5, Composer 2.5) and one for everyone else's. Emptying either on Pro, Pro Plus, Ultra, Teams, or Enterprise moves that plan onto pay-as-you-go at the same API rates. Start has no on-demand path and no Other Models pool. First-party Grok and Composer requests are exempt from the Cursor Token Rate; everything else on Teams and Enterprise pays $0.25 per million tokens on included usage, on-demand, and BYOK. Auto bills at the list price of whatever model it routes to. Spend limits are meant to stop on-demand usage, though Cursor says enforcement is not instant. On Teams, hitting the team cap stops AI features for every member still consuming on-demand. Spend alerts notify and do not stop usage. Individual list prices other than Start are before tax.
Claude Code: Anthropic's agent across terminal, desktop, web, and IDE
Claude Code is available in the terminal, a desktop app, the web, and IDE integrations, with GitHub and Agent SDK options for additional workflows. It is designed for Claude models: Opus, Sonnet, and Haiku. Organizations can access those models through Anthropic, Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry, or a compatible LLM gateway. Third-party gateways such as LiteLLM can also translate Claude Code traffic to non-Claude models, but Anthropic does not support that configuration and compatibility can vary.
It is built for repository-scale work such as debugging, refactors, tests, multi-file features, and pull-request preparation. Sessions can include large amounts of code, tool calls, and multi-step reasoning. Anthropic therefore tells customers to budget more for coding than chat, and says agent teams use approximately 7x more tokens than standard sessions when teammates run in plan mode.
Claude Code plans
Claude Code is included on every paid Claude plan. It is not on Free. The seat allowance is unpublished: Anthropic documents a rolling five-hour window and a weekly window, shared across Claude Code, chat, Cowork, and other Claude surfaces on the plan, but no token or message count for either tier.
| Plan | Price | Claude Code | Included usage | What this tier covers | What you pay extra for |
|---|---|---|---|---|---|
| Free | $0 | No | Chat on web, iOS, Android, and desktop | Chat, web search, artifacts, memory, connectors | Claude Code starts on Pro |
| Pro | $17/mo billed annually ($20 monthly) | Yes | At least 5x Free per five-hour window, plus weekly limits. 200k context | Cowork, Design, Science, Research, projects, more models, Claude for Microsoft 365 | Usage credits at API rates after the seat allowance |
| Max 5x | $100/mo | Yes | 5x Pro per five-hour window; higher output limits; priority access | Everything in Pro, plus early access to advanced features. Fable uses up to 50% of weekly limits on both Max tiers | Usage credits |
| Max 20x | $200/mo | Yes | 20x Pro per five-hour window | Same Max features as 5x, with more usage | Usage credits |
| Team Standard | $20/seat/mo annual ($25 monthly); 2 to 150 people | Yes | More usage than Pro; mix Standard and Premium seats | Central billing, SSO, connector admin, enterprise desktop deployment, enterprise search. No model training by default | Usage credits; admin spend limits if extra usage is enabled |
| Team Premium | $100/seat/mo annual ($125 monthly) | Yes | 5x Standard | Same Team features as Standard | Same extras as Standard |
| Enterprise (self-serve) | $20/seat/mo plus usage at API rates, billed annually | Yes | Seat fee is access only. All model usage is billed at API rates. 500k context on the default model | Team features plus user and org spend limits, RBAC, SCIM, audit logs, Compliance API, custom retention, network controls, HIPAA-ready option | There is no included token pool. Usage is the bill |
| Enterprise (sales-assisted) | Custom, billed annually | Yes | Custom usage terms | Self-serve Enterprise features plus invoicing, AWS Marketplace, consultation, and customer success at committed-spend thresholds | Custom |
A third path skips seats entirely: point Claude Code at an API key, or run it through Amazon Bedrock, Microsoft Foundry, or Google Cloud, and pay per token. Usage through a cloud provider is invisible to Anthropic's own dashboards. On a subscription, prompt cache lifetime is an hour inside the seat, and the default drops to five minutes once you start drawing on usage credits. You can keep the one-hour lifetime on usage credits by choosing that TTL.
Crossing the allowance can make the next hour more expensive than the last. More cache misses means more input tokens reprocessed, at the moment you begin paying per token. Anthropic states the billing consequence: usage inside the seat allowance is not metered in dollars.
Codex: OpenAI's agent across local and cloud work
Codex spans a CLI, IDE extension, desktop app, web and cloud surfaces, GitHub code review, and Slack. An API key supports the local agent, while ChatGPT plans add the hosted product surfaces. Codex uses OpenAI models by default. The CLI can also connect to custom providers or locally hosted open-source models, but these require manual setup.
Its strongest distinction is delegation across local and hosted environments. Teams use it for scoped bug fixes, test updates, refactors, pull-request review, and parallel cloud tasks that should not occupy a local machine.
Codex plans
Codex rides on ChatGPT plans rather than a Codex-specific seat, so there is no Codex-only pool on current plans. OpenAI's Codex docs list Plus local GPT-5.6 Sol messages per five-hour window as 10 to 100. The consumer Codex pricing page lists 15 to 90. Both are ranges, because similar-looking tasks consume very different amounts of the allowance.
| Plan | Price | Codex surfaces | Included usage | What this tier covers | What you pay extra for |
|---|---|---|---|---|---|
| Free | $0 | Explore Codex on quick tasks | Tight limits | Limited local Codex | Upgrade for cloud review, Slack, and credits |
| Go | $8/mo | Lightweight coding | Lower than Plus | A cheaper personal entry than Plus | Credits and higher limits on Plus or Pro |
| Plus | $20/mo | Web, CLI, IDE, iOS, GitHub code review, Slack | Five-hour window plus weekly limits on some plans. Learn docs: 10-100 local Sol messages. Consumer page: 15-90 | GPT-5.6 Sol, Terra, and Luna. Shared with ChatGPT Work and ChatGPT for Excel | Purchased credits after included limits |
| Pro 5x | $100/mo | Everything in Plus, plus GPT-5.3-Codex-Spark (research preview) | 5x Plus | Spark has a separate usage limit that can change with demand | Purchased credits |
| Pro 20x | $200/mo | Same as Pro 5x. Unlimited ChatGPT Voice; tasks still draw from the Codex budget | 20x Plus | Same Pro features at the higher multiplier | Purchased credits |
| API key | Token rates | CLI, SDK, and IDE only | None included | Local and scripted Codex. Model availability follows the key | No GitHub code review, Slack, or other cloud Codex features. Pay API rates |
| Business Standard | $20/user/mo annual, $25 monthly; 2+ seats | ChatGPT Chat, Work, and Codex | Included seat limits; five-hour window. Mix Standard and Premium | Shared workspace. Pool also includes ChatGPT for PowerPoint and Workspace Agents when those features are on the plan. Card billing only | Workspace credits are optional. Without credits, included limits stop eligible usage |
| Business Premium | $100/user/mo annual, $125 monthly | Same Business surfaces | 5x Standard allotment; no five-hour usage limit. Codex docs still mention additional weekly limits on some plans | Same workspace features as Standard. Seat cap 200 on workspaces created on or after 24 Aug 2026 | Workspace credits |
| Enterprise / Edu | Custom | Business surfaces plus SSO, analytics, retention, and admin APIs | Flexible pricing: no fixed rate limits; usage scales with credits. Without flexible pricing: per-seat limits like Plus | Invoice billing, SCIM, EKM, and compliance APIs on supported contracts | Owner-set overage cap. A new credit workspace has none, so usage can continue and bill |
| Legacy Codex-only seat | No included allowance | Codex only | Draws on workspace credits from first use | Only workspaces that had a Codex-only seat, or a pending invite, as of 24 Jun 2026 can still use and add it | New workspaces, and old ones that never added it, cannot |
Fast mode consumes 2.5x credits on GPT-5.6 and GPT-5.5, and 2x on GPT-5.4. GPT-5.4 and GPT-5.4 mini leave ChatGPT-signed-in Codex on 31 Aug 2026. API-key Codex is unaffected.
What each costs
The product sections above are the seat and allowance for each tool. This section puts those plans side by side, then the model rates and how to read each vendor's cost number.
| Plan | Cursor | Claude Code | Codex |
|---|---|---|---|
| Free / cheapest | Hobby free (limited Agent); first paid is Pro $20/mo. India Start ₹649/mo | Free has no Claude Code. Pro $20/mo or $17/mo billed annually | Go $8/mo |
| Mid individual | Pro Plus $60/mo | Max 5x $100/mo | Plus $20/mo |
| Top individual | Ultra $200/mo | Max 20x $200/mo | Pro $200/mo (20x); 5x starts at $100/mo |
| Team | Standard $40/user/mo ($32 annual); Premium $120/user/mo ($96 annual) | Standard $25/user/mo ($20 annual); Premium $125/user/mo ($100 annual) | Business Standard $20/user/mo annual, $25 monthly; Premium $100/user/mo annual, $125 monthly; 2+ users |
| Enterprise | Custom | Self-serve: $20/seat/mo plus usage at API rates, billed annually; sales-assisted: custom | Custom |
| Dimension | Cursor | Claude Code | Codex |
|---|---|---|---|
| How included usage resets | Two monthly pools: Cursor Models and Other Models | Rolling five-hour window and a weekly window | Five-hour window, plus weekly limits on some plans |
| What the pool is shared with | Split by model family, not by product | Claude Code, chat, Cowork, and other Claude surfaces on the plan | Plus and Pro: Codex, ChatGPT Work, and ChatGPT for Excel. Business also shares ChatGPT for PowerPoint and Workspace Agents when those features are on the plan |
| After the allowance | Pay-as-you-go at the same API rates | Usage credits | Purchased credits |
| Default overage cap | Individual and team hard caps. On Teams, hitting the team cap stops on-demand for every member still consuming it. Enterprise adds per-user (and group) caps. Cursor says enforcement is not instant | Seat allowance is the default ceiling unless usage credits and admin spend limits are enabled. Self-serve Enterprise has no included pool | Enterprise/Edu credit workspaces: no cap until an owner sets one. Business: included limits stop usage unless the workspace has purchased credits |
| Surcharge | Cursor Token Rate: $0.25 per million tokens on third-party models, Teams and Enterprise. Applies to included usage, on-demand, and BYOK, not only after the allowance | 1.1x for inference_geo: "us" on Claude 4.6+ | Fast mode: 2.5x credits on GPT-5.6 and GPT-5.5; 2x on GPT-5.4. GPT-5.4 and GPT-5.4 mini leave ChatGPT-signed-in Codex on 31 Aug 2026 |
Model rates
Input / output per million tokens, unless noted. Codex dollars below convert credits at $0.04, the list implied by matching OpenAI's credit and USD rate cards.
| Rate | Cursor | Claude Code | Codex |
|---|---|---|---|
| Frontier models | Grok 4.6 $2 / $6; GPT-5.6 Sol $4 / $20; Claude Opus 5 $5 / $25 | Opus 5 $5 / $25; Sonnet 5 $2 / $10 | Sol 100 / 500 credits, about $4 / $20 |
| Cheapest first-party model | Composer 2.5 $0.50 / $2.50 | Haiku 4.5 $1 / $5 | Luna 5 / 30 credits, about $0.20 / $1.20 |
Rate cards move in weeks. Cursor repriced Teams in June 2026. Bugbot switched from a $40 per seat subscription to usage-based billing, announced 11 May 2026 and effective at each customer's next renewal after 8 June 2026. Anthropic cancelled a scheduled Sonnet 5 increase, leaving $2 / $10 as standard. OpenAI's Sol promotional pricing runs at least through November 21, 2026 for eligible usage paid with purchased credits; included plan usage, five-hour and weekly limits, and legacy credit rates are unchanged. Re-check any figure you plan to put in a budget.
Published cost benchmarks
Anthropic publishes a deployment benchmark: about $13 per active developer day and $150 to $250 per developer month across enterprise deployments, with 90% of users below $30 per active day. It is a distribution, not a quote.
OpenAI publishes a Codex average of about $100 to $200 per developer per month, with costs varying by model, concurrent instances, automations, and fast mode. Cursor does not publish a monthly per-developer average. Cursor's Bugbot blog cites an average run cost of $1.00 to $1.50 depending on PR size and complexity; this comparison does not treat that average as a rate card. Codex code-review costs vary with usage.
How to read each vendor's cost data
A dollar sign does not always mean an invoice charge. Use the cost fields to compare tools only after identifying how each vendor calculated them.
| Product | Number you receive | How to interpret it |
|---|---|---|
| Cursor | Included usage and on-demand spend | Teams and Enterprise Admin API events and invoices report charges and separate included usage from pay-as-you-go spend. Self-serve Usage views may show tokens rather than dollars. |
| Claude Code | A session estimate in /usage; usage-credit spend in organization reports | /usage estimates token cost at list or administrator-configured rates and is not an invoice. Organization reports cover paid usage credits, not work covered by the seat allowance. Usage-based Enterprise analytics can return both post-discount and list-price amounts. |
| Codex | Credits and estimated dollars | The dollar value uses that workspace's credit conversion rate, so the same number of credits can represent a different cost under another contract. |
The $0.04 per credit rate used in this article is a list implied by matching OpenAI's credit and USD rate cards, not a per-credit price OpenAI publishes as a constant. A negotiated workspace can convert credits differently, so every Codex dollar figure carries that caveat.
Cursor vs Claude Code vs Codex compared
| Dimension | Cursor | Claude Code | Codex |
|---|---|---|---|
| Primary workflow | Standalone AI editor, plus CLI and cloud agents | Terminal, desktop, web, and IDE | CLI, IDE, desktop, web, and cloud tasks |
| Models | Composer, Grok, Claude Opus 5, Gemini, GPT-5.6 Sol, and open-weight options such as Kimi K3 and GLM 5.2; direct OpenAI supply proposed to end November 12, 2026 | Claude models by default; third-party gateways can route to other models, but Anthropic does not support that configuration | OpenAI by default; custom providers and local OSS models are available in the CLI with setup |
| Cost basis in the product UI | Invoice and Admin API charge; Usage UI may be tokens-only | List-rate estimate | Credits, with an optional dollar conversion |
| Controls that can stop paid usage | Individual and team spend limits; Enterprise also has per-user (and group) limits | Admin spend limits; --max-budget-usd (print mode only; does not apply to interactive sessions) | Plan-dependent ChatGPT workspace credit limits; separate from API billing |
Matching tools to teams
Cursor suits developers who want AI throughout the editing process, and finance teams that can reconcile Admin API or invoice charges. The VS Code foundation lowers migration cost. The Token Rate and Auto routing are the two billing behaviors to watch.
Claude Code suits repository-wide work from the terminal. Compare a typical month of local list-rate usage with the seat price, including any API discount, before deciding between Max and the API. Cloud-provider deployments need their own accounting from day one.
Codex suits scoped, delegated work that can run locally or in the cloud. Confirm the workspace overage rate before treating any Codex dollar figure as a bill. On an Enterprise or Edu credit workspace, set an overage cap; a new one has none. On Business, included limits already stop usage unless credits are in the pool.
Many teams run an editor-based assistant beside a terminal or cloud agent. When both can modify the same codebase, review requirements matter as much as the seat price.
How SuperPenguin tracks Cursor, Claude Code, and Codex
SuperPenguin brings Cursor, Claude Code, and Codex into one view. Engineers can track their spend and usage-limit health, while team leaders can see aggregate activity on the web. Connecting GitHub adds the shipped outcome by linking coding activity and cost to merged pull requests.
Track each engineer's spend and usage limits
SuperPenguin does this without routing coding-agent traffic through a proxy. Each engineer installs the free macOS app, which meters local Cursor, Claude Code, and Codex sessions. It also shows current usage limits, remaining allowance, reset times, and whether the engineer is on pace to exhaust a limit before it resets.
An engineer can see their own activity in the app. On a SuperPenguin Pro workspace, or an Enterprise workspace with the AI Coding entitlement, engineers connect Desktop to the company workspace and the web dashboard aggregates the activity they share. Owners, Viewers, and Policy administrators can see organization-wide reporting, while the installation roster shows which engineers are connected.
Provider billing connections are optional for Desktop reporting. A Cursor connection can add provider-billed team spend. Claude Code and Codex coding activity comes from Desktop, not the Anthropic or OpenAI billing APIs, so their session costs remain estimates rather than invoice charges.
Connect spend to shipped pull requests
Connecting GitHub adds repository, commit, and merged-pull-request data. SuperPenguin links that data to Desktop activity and calculates the coding cost associated with each merged pull request.
Together, usage health and pull-request outcomes show how coding spend translates into shipped work. This gives engineering leaders useful evidence when deciding whether to grant additional allowance, adjust budgets, or compare tools across teams. Activity that cannot be matched stays visible as unattributed instead of being spread across the pull requests that did match.
Choose the tool that fits how your team works, then evaluate it against the outcomes the team ships. Seat price, allowance health, and cost per merged pull request provide a stronger basis for budget decisions than any one vendor's usage total.
Sources
- Pricing, Cursor.
- Models and pricing, Cursor docs.
- Models and pricing, Cursor docs.
- Cursor Token Rate, Cursor help.
- Teams pricing update, Cursor.
- Changes to Bugbot pricing, Cursor.
- Spend limits, Cursor docs.
- Spend alerts, Cursor docs.
- Cursor Agent, Cursor docs.
- Bring your own API key, Cursor docs.
- Our decision on Cursor following its acquisition by SpaceX, OpenAI.
- Manage costs effectively, Claude Code docs.
- Analytics, Claude Code docs.
- Cost tracking, Claude Agent SDK docs.
- Pricing, Claude Platform docs.
- Claude Enterprise Analytics API, Claude Platform docs.
- Claude plans and pricing, Anthropic.
- Claude Code LLM gateways, Claude Code docs.
- CLI reference, Claude Code docs.
- Codex pricing, OpenAI.
- Codex pricing, OpenAI (consumer page; Plus Sol five-hour range differs from the Learn docs).
- ChatGPT rate card: Business, Enterprise and Edu, OpenAI help.
- Managing billing and seats in ChatGPT Business, OpenAI help.
- Codex usage and billing, OpenAI help.
- Managing legacy Codex seats in ChatGPT Business, OpenAI help.
- Speed and fast mode, OpenAI.
- ChatGPT usage limits and spend controls, OpenAI.
- Manage usage limits and overages in ChatGPT Enterprise and Edu, OpenAI help.
- Codex advanced configuration, OpenAI.
- Install SuperPenguin Desktop, SuperPenguin docs.
- Set up AI Coding and PR Costs, SuperPenguin docs.
- Connect GitHub and enable PR Cost Attribution, SuperPenguin docs.
- Use Claude Code with non-Anthropic models, LiteLLM docs.
Keep reading
SuperPenguin vs LLM Observability vs Gateways
Compare SuperPenguin with LLM observability (Langfuse, LangSmith) and gateways (Helicone, OpenRouter).
SuperPenguin vs Vantage vs CloudZero: AI Cost Management
Compare SuperPenguin with Vantage and CloudZero for AI cost tracking, attribution, pricing, integrations, chargeback, and cost per customer.