AI News

Cursor Composer 2.5 review: pricing, changes, and one tested repair

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.

Commercial relationship for this article: None. This is based on Curtis Pyke’s direct confirmation on October 7, 2026.

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: Token-rate observation from September 5, 2026, rechecked October 7. Purchase-plan correction dated October 20 appears below. Token rates, included usage, subscription price, and total job cost are separate measures.

The September 5 Models & Pricing capture, rechecked October 7, placed Composer 2.5 in the Cursor Models pool and listed 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

For new subscribers, Cursor’s current plan documentation lists Plus at $8/month, Premium at $30/month, Super at $100/month and Ultra at $200/month, before tax. The plan documentation says Fast mode is off on Plus. A reader choosing Composer 2.5 Fast should check the higher-tier access rules before subscribing. The new paid tiers use a weekly Cursor-model pool shared with Grok; third-party models cost extra on Premium and above. Legacy Pro, Pro+ and Ultra keep separate monthly-pool terms. A subscription price does not establish the cost of a completed repair.

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:

  1. One file changed: src/temperature.js.
  2. Cursor reported npm test with 3 pass, 0 fail.
  3. An independent local rerun produced the same 3 pass, 0 fail result.
  4. 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.

Synthetic Composer 2.5 test trace showing one constrained JavaScript repair, one changed file, and an independently confirmed three-pass test result.
One synthetic repair; not a benchmark or real-project test.

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.

Buying details checked October 22, 2026: Cursor Composer buying profile: Fast is the product default where enabled; subscription price is separate from token rates. New and legacy plan allowances differ. The original article keeps its existing dates and evidence scope; the linked profiles describe their own demonstrations.