Pricing one Cursor agent run against Factory

In Cursor, one agent request can draw on three different budgets. Cursor's own models come out of a weekly usage pool. Third-party models such as Claude, GPT, and Gemini are billed at API list price from prepaid credit. On Teams and Enterprise, eligible third-party requests also carry a per-token Cursor Token Rate, including requests made with the team's own API key.

The model picker and the Auto mode decide which budget a request uses.

Factory meters agent work differently. Factory-managed models consume Standard Credits scaled by a published per-model multiplier, and BYOK requests go to the customer's provider account. The example below prices one agent run each way, using published rates as of October 2026. For the trial method behind a fair comparison, see Factory and Claude Code: compare completed work.

What the weekly pool covers

Cursor is sold as part of The X Subscription. According to Cursor's models and pricing documentation, Plus costs $8 per month, Premium $30 with 4x Plus usage, Super $100 with 15x Plus usage, and Ultra $200 with 50x. Teams seats cost $30, $100, or $200 and match those tiers. Each paid plan includes one pool for Cursor models (Grok 4.7, Grok 4.6, Grok 4.5, and Composer 2.5). That pool is shared across Agent in the editor, the CLI, Cloud Agents, Grok Bot, and the Grok apps, and it refreshes weekly.

Third-party models are not part of the pool. On Premium and above they are billed at API list price from the first request. Plus covers only Cursor models at fixed settings, with Fast off and medium effort. A team running Claude in Cursor therefore pays per token from the start, on top of the seat price.

Autocomplete is billed separately from agent work. Cursor's legacy Pro, Pro+, and Ultra plans include unlimited Tab completions. The current models and pricing page describes the pool in terms of model requests and lists no separate Tab rate. Check the usage dashboard for how Tab is metered on a current plan before modeling it.

The same agent run on six Cursor billing paths

The token profile here is an illustrative assumption for a multi-file bug fix with test runs. It is not a measured Cursor or Factory session. Assume 2.0 million input tokens over about 40 turns, made up of 1.8 million cache reads, 150,000 cache writes, and 50,000 uncached input tokens, plus 50,000 output tokens. That is 2.05 million tokens in total.

Cursor lists Claude Opus 5.5 at $4 input, $5 cache write, $0.20 cache read, and $20 output per million tokens. That matches Anthropic's published Opus 5.5 rates. On an individual Premium plan, the run costs 0.05 × $4 + 0.15 × $5 + 1.8 × $0.20 + 0.05 × $20 = $0.20 + $0.75 + $0.36 + $1.00 = $2.31, paid from prepaid credit.

On Teams, the Cursor Token Rate adds $0.25 per million input, output, and cached tokens on eligible third-party requests. This example counts all 2.05 million tokens, which adds $0.5125 for a total of $2.8225. The same rate applies when the team uses its own Anthropic key. In that case, Anthropic bills $2.31 directly and Cursor charges $0.5125 from the team's prepaid credit. On individual plans, BYOK requests are billed only by the provider and do not draw on Cursor usage.

Cursor's own models skip the Token Rate. Beyond the weekly pool, Composer 2.5 is listed at $0.50 input, $0.20 cache read, and $2.50 output, with no cache-write rate. Billing the cache writes as input, the run costs 0.2 × $0.50 + 1.8 × $0.20 + 0.05 × $2.50 = $0.585. Grok 4.7 at $2, $0.50, and $6 gives 0.2 × $2 + 1.8 × $0.50 + 0.05 × $6 = $1.60. Inside the weekly pool, both draw on included usage instead.

Two more settings change the Opus figure. Regional data residency adds 10% on eligible models, which brings it to $2.541. Max Mode, now available only on legacy request-based plans, bills at the API rate plus 20%, which brings it to $2.772.

The six paths for this one run:

Auto modes decide which line of the bill a request lands on

Cursor Auto has Cost, Balance, and Intelligence modes. The Cursor Router documentation says Cost mode bills as a Cursor-model request from the weekly pool, whatever model serves it, with no Token Rate. On Teams and Enterprise, Balance and Intelligence classify each request and bill at the routed model's list price. A third-party route is charged from prepaid credit, plus the Token Rate where it applies. Individual plans have Cost mode only.

Two documented details affect measurement. Subagents can run a named third-party model while the picker still shows Auto, Grok, or Composer, and those requests bill as third-party usage. Cursor also hides the underlying routed model by default. For a cost trial, turn on the display of the underlying model so each request can be matched to its rate.

Factory's version: multipliers, BYOK, and routing

On Factory, the models page assigns each managed model a usage multiplier. Claude Opus 5.5 is 1.6, Claude Sonnet 5.5 and Grok 4.7 are 0.8, and Claude Haiku 5.5 is 0.04. For the same token profile, an Opus 5.5 session therefore consumes Standard Credits at twice the rate of a Sonnet 5.5 session. Individual plans are $20, $100, and $200 per month, with rolling 5-hour, 7-day, and 30-day limits. Teams is $60 per month plus $40 per seat, for up to 10 seats.

With BYOK, Anthropic would bill the example's $2.31 on the customer's key. Factory individual plans include a free BYOK allowance, after which BYOK usage is charged according to the plan. The CLI /cost command shows cache hit rates, which matters for a profile where 88% of tokens are cache reads.

Factory publishes measured results for its own router. At launch, Factory Router reached 99% of Claude Opus 4.7's Terminal-Bench 2 pass rate at 20% lower cost per session, and 96% on Legacy-Bench at 25% lower cost. Charged only against tasks that finished, its cost per successful run was 80.5% of Opus on Terminal-Bench 2. In production from September 5 to 18, 2026, Factory reports median savings of 72.5% per routed session against frontier-model pricing, with 84% of routed sessions saving at least half.

These figures describe Factory Router on Factory workloads and are not a measurement of Cursor Router.

Factory has also published a same-model harness comparison that includes Cursor. In its October 2026 Droid 1.0 research post, Factory ran Droid and cursor-cli@2026.10.19-6bbf2c4 on all Terminal-Bench 4 tasks with Claude Opus 5.5 at high reasoning, three trials per task. Its figure shows Droid at 65.2% and Cursor at 62.1%, with Droid's cost per passed task 1% lower and its time per passed task 26% lower. The methodology notes that Cursor undercounts tokens when it reconnects mid-task, so for the two trials where that occurred, Factory estimated the missing usage from the transcript.

Context size is a cost line on both products

The Token Rate is charged on cached tokens as well, so a cache-heavy agent run pays it on context that is re-read each turn. In this example, the 1.8 million cache reads account for $0.45 of the $0.5125 Token Rate. Cutting repeated context lowers that charge directly, and it lowers the model charge on both products.

Factory has measured one source of repeated context. Its deferred context engine loads MCP tool schemas only when needed. Among measured turns in MCP-triggered sessions, it cut estimated input tokens by 15.1% on average and by 50.8% in sessions with 100 or more hidden tools. Cursor's legacy Max Mode works in the other direction, since it extends the context window and can increase the tokens a request uses.

Further reading

• Factory Droid vs Cursor
• Token efficiency starts with smaller tool results
• Factory Router cuts inference costs by 63%
• Factory individual plans