Verdict: Cursor Composer 2.5 Fast fixed one constrained JavaScript defect in 14 seconds, changed one file, and passed the same three tests in Cursor and in Kingy’s independent local rerun.
That result is evidence for one synthetic repair. It is not a benchmark, speed guarantee, pricing test, reliability finding, or proof that Composer 2.5 is ready for a large production repository.
Source status: Cursor’s official Composer 2.5 changelog and Models & Pricing page were rechecked on September 5, 2026. Recheck both on the day of publication.
What Cursor officially says changed
Cursor’s dated changelog says Composer 2.5 became available in Cursor on May 18, 2026. The same entry describes it as an improvement over Composer 2 in intelligence, sustained work, complex-instruction following, and collaboration quality.
Those are Cursor’s release claims. Kingy’s synthetic task does not independently validate any of them, and this article should not turn them into measured performance conclusions.
For broader company context, see Kingy’s report on Cursor’s move toward a frontier lab.
Source: Cursor Composer 2.5 changelog, published May 18, 2026 and accessed September 5, 2026.
Current documented model rates
Pricing snapshot: Official Cursor rates and individual-plan prices, checked September 5, 2026. Token rates, included usage, subscription price, and total job cost are separate measures.
Cursor’s current Models & Pricing page places Composer 2.5 in its Cursor Models pool and lists the following rates per million tokens:
| Variant | Input | Cache read | Output | What this table does not mean |
|---|---|---|---|---|
| Composer 2.5 | $0.50 | $0.20 | $2.50 | It is not a subscription price or total-job estimate |
| Composer 2.5 Fast | $3.00 | $0.50 | $15.00 | It is not a latency or quality guarantee |
Cursor separately lists individual plan prices of $20 per month for Pro, $60 for Pro Plus, and $200 for Ultra. Subscription price, included usage, model token rates, and a user’s final bill are different measures. This review does not estimate what a representative project will cost.
Source: Cursor Models & Pricing, accessed September 5, 2026.
The synthetic repair Kingy actually ran
The test repository was newly created, local, disposable, and non-sensitive. It contained a small temperature utility with two deliberately seeded defects: an incorrect Celsius-to-Fahrenheit calculation and parsing that accepted a partial numeric string.
Composer 2.5 Fast was instructed to make the smallest repair only in src/temperature.js, add no packages, use no network access or external services, create no commit, avoid every other workspace, and run the existing tests.
The task record showed:
- One file changed:
src/temperature.js. - Cursor reported
npm testwith 3 pass, 0 fail. - An independent local rerun produced the same 3 pass, 0 fail result.
- The recorded task duration was 14 seconds.
The final code corrected the conversion and rejected partial strings such as 12C. The repository remained a tiny synthetic fixture, so the result says nothing about large codebases, long-horizon work, collaboration, security, or production readiness.
Evidence table
| Question | Evidence available | Confidence | Boundary |
|---|---|---|---|
| Was Composer 2.5 announced and available? | Dated Cursor changelog | High | Availability in Cursor only; no surface-by-surface audit |
| What improvement did Cursor claim? | Dated Cursor changelog | High that Cursor made the claim | Not independently verified by Kingy |
| What are the documented current rates? | Current Cursor Models & Pricing page | High as of access date | Not a cost forecast |
| Did the synthetic repair pass? | Cursor task result plus independent local rerun | High for this fixture | Not a benchmark or reliability result |
| Did only one file change? | Post-run local file inspection | High for this fixture | Not evidence about normal agent scope |
| Is Fast generally faster or better? | No suitable Kingy evidence | None | Do not claim it |
What this review does not establish
This review does not make benchmark comparisons, competitive rankings, generalized reliability claims, or predictions about how Composer 2.5 will behave across real projects. It also does not claim Composer 2.5 is limited to a particular set of product surfaces unless a current official source supports that exact list.
The evidence boundary is simple: Cursor’s changelog establishes what Cursor announced; the synthetic task establishes what happened once in a controlled local fixture. Neither establishes a product-wide win.
Practical decision guide
| If you need to decide… | Verify before adopting |
|---|---|
| Whether Composer 2.5 fits your repository | Run a representative, non-sensitive task with written success and intervention criteria |
| Whether Fast is worth its rate | Compare Standard and Fast on the same tasks, model settings, and review process |
| Whether an agent can work safely | Scope files and commands, require review, and treat destructive/network actions as separate approvals |
| What it will cost | Record actual token usage and consult current pricing; do not extrapolate from this task |
| Whether it is reliable | Repeat tasks across multiple repositories and failure classes; one passing fixture is insufficient |
For the programmatic agent surface, read Kingy’s Cursor SDK guide. For a wider category comparison, see the AI coding agents comparison.
FAQ
Did Kingy benchmark Composer 2.5?
No. Kingy ran one constrained synthetic repair. It was not a comparative benchmark.
Did the task really take 14 seconds?
The task record showed 14 seconds for this specific fixture. That is not a latency guarantee or a basis for comparing models.
What changed in Composer 2.5?
Cursor says the release improves intelligence, sustained work, instruction following, and collaboration behavior. Those remain vendor claims unless independently evaluated. Source: Cursor Composer 2.5 changelog, accessed September 5, 2026.
What are the current token rates?
On September 5, 2026, Cursor documented $0.50 input, $0.20 cache read, and $2.50 output per million tokens for Composer 2.5; Fast was $3 input, $0.50 cache read, and $15 output. Recheck Cursor Models & Pricing on publication day.
Does the passing test prove production readiness?
No. The fixture was intentionally small, local, and synthetic. It did not test security, real-project behavior, collaboration, or reliability.
The Kingy Brief
Get future Kingy Brief editions.
Source-checked AI changes, original tests and one practical thing to try.
Free · Choose your subjects · Double opt-in · Unsubscribe anytime
Regular sending is paused; no restart date is set.