Cursor’s 3× Tier Tests the IDE Margin Moat

On March 19, 2026, Cursor priced Composer 2 at $0.50 per million input tokens and $2.50 per million output tokens. Its faster variant cost three times more. Cursor published a token meter, while the cost of a completed, accepted coding task remained unreported.

Composer 2 separated the model from the product

On March 20, Moonshot said Kimi K2.5 provided the foundation for Composer 2 and that Cursor accessed it through Fireworks AI. Moonshot credited Cursor with continued pretraining and high-compute reinforcement learning before Cursor packaged the result in its coding product.

The cited announcement does not disclose Cursor’s training bill, its serving contract with Fireworks, ownership of the resulting weights, or how much each layer contributed to Composer 2’s reported performance.

• March 19, 2026: Cursor launched Composer 2 and published standard and fast-tier pricing.
• March 20, 2026: Moonshot identified Kimi K2.5 as the foundation and Fireworks AI as the inference provider.
• April 20, 2026: Moonshot released the open-weight Kimi K2.6 under a modified MIT license.
• May 19, 2026: Cursor released Composer 2.5, still built on Kimi K2.5, and claimed better sustained work on long-running tasks.

Cursor’s next named Composer release remained on Kimi K2.5 after K2.6 appeared. The cited sources leave Cursor’s rationale unresolved, but the sequence shows that a newer upstream release did not automatically become the next production base.

The 3× tier leaves task cost unanswered

At Cursor’s listed rates, one million output tokens cost $2.50 on the standard model and an implied $7.50 on the faster variant. The corresponding input cost rises from $0.50 to $1.50.

Consider a hypothetical routing rule: send an interactive patch to the fast tier while a developer waits, then use the standard tier for an unattended migration. At one million output tokens, that choice adds $5 to the interactive job and saves $5 on the unattended one. Cursor has not said it uses this policy.

The published prices also omit task-level latency, token consumption, retries, test results, and human acceptance rates. A faster run could justify its premium by finishing sooner or avoiding retries, but Cursor’s data do not establish either outcome. The 3× price therefore measures the tier premium, not the cost of accepted work.

On April 6, 2026, The Wall Street Journal reported that investor documents from OpenAI and Anthropic showed inference costs exceeding half of revenue. Those figures establish the scale of inference spending at two model labs. The report provides no corresponding figure for Cursor’s serving contract, discounts, revenue, or gross margin.

Security puts failure inside the route’s cost

Security researchers reported sandbox escapes or boundary bypasses across Cursor, Codex, Gemini CLI, and Antigravity. They used files later consumed by trusted tools, crossing a boundary between generated artifacts and higher-trust components. Vendors patched most of the identified vulnerabilities, while some remained at publication.

The research does not establish that multi-model routing caused those flaws. It does establish that agent handoffs can cross trust boundaries. When one model writes a file and a test or deployment tool later opens it with broader permissions, the IDE must preserve the file’s provenance and permission limits across review, fallback, and execution.

Reviews, sandbox checks, audit logs, and human escalation consume time and compute. An IDE vendor that compares routes must count those controls alongside tokens and latency rather than treat security failures as costs outside the task.

Accepted work is the missing denominator

Per-token prices expose one input. Comparing routes requires the total model and serving charges, elapsed time, retries, reviews, and failures for changes that pass the required tests and earn acceptance. Cursor’s public posts do not provide that denominator or compare its two Composer 2 tiers against it.

Three times the price is easy to see. The harder number is cost per accepted task after latency, retries, tests, and security review—a number Cursor has not published. An IDE earns durable advantage only when it can carry context across suppliers, choose routes from measured outcomes, and own the result.