Cursor's in-house model just leapfrogged its own pricing. Composer 1.5 shipped in February at $3.50 per million input tokens. Composer 2, released on March 18, 2026, costs $0.50 - an 86% drop. It also outscores Claude Opus 4.6 on coding benchmarks while costing a tenth as much per token.

If you've been tracking Cursor as a line item, that's a meaningful shift in the cost of every agent request your team makes.

Composer 2 Pricing vs Opus 4.6 and GPT-5.4

Model	Per 1M Input Tokens	Per 1M Output Tokens
Composer 2 Standard	$0.50	$2.50
Composer 2 Fast	$1.50	$7.50
Composer 1.5	$3.50	$17.50
Claude Opus 4.6	$5.00	$25.00
GPT-5.4	$2.50	$15.00

Composer 2 pricing compared to Composer 1.5 and third-party frontier models. See the full list on the Vantage model pricing index.

Composer 2 comes in two variants. Standard at $0.50/$2.50 is the cheapest option. Fast at $1.50/$7.50 offers the same intelligence with lower latency, and Cursor is making it the default - so that's what your team will use unless they explicitly switch.

Even the pricier Fast variant undercuts every third-party model in the table. It's 70% cheaper than Opus 4.6 on both input and output. Against GPT-5.4, the gap is 40% on input and 50% on output.

To put that in team-level terms: say your org generates 10 million output tokens per month across Cursor - a reasonable number for a 20-30 person engineering team using agentic workflows daily. On Opus 4.6, that's $250/month in output costs alone. On Composer 2 Fast, it's $75. On Standard, $25. Over a year, the difference between Opus 4.6 and Composer 2 Standard is roughly $2,700 - just on output tokens, just from one team. For a larger org with multiple engineering groups, multiply accordingly and you're looking at five figures.

Cache-read pricing follows the same trend: $0.20 per million tokens for Standard, $0.35 for Fast. If your workflows involve repetitive context - and most agentic coding sessions do - the caching discount compounds on top.

Composer 2 Benchmark Scores

The pricing would be less interesting if Composer 2 were a weaker model. It's not.

Model	CursorBench	Terminal-Bench 2.0	SWE-bench Multilingual
Composer 2	61.3	61.7	73.7
Composer 1.5	44.2	47.9	65.9
Composer 1	38.0	40.0	56.9

Composer model benchmark progression (source). Third-party benchmark references: CursorBench, Terminal-Bench 2.0, SWE-bench.

The jump from Composer 1.5 is the largest single-generation improvement Cursor has shipped - 17 points on CursorBench, nearly 14 on Terminal-Bench 2.0, and 8 on SWE-bench. On Terminal-Bench 2.0, which measures how well AI agents complete real-world tasks in terminal environments, Composer 2 at 61.7 now sits ahead of Opus 4.6 at 58.0. GPT-5.4 still leads at 75.1, but it also costs 5x more per input token. On SWE-bench Multilingual - real GitHub issues across multiple languages - Composer 2 hits 73.7, up from 65.9 just a month ago.

These are Cursor-reported numbers, and each model was evaluated with a different harness, so the exact rankings deserve some skepticism. But the direction is clear enough: this is a model competing in the same tier as Opus 4.6, priced like a budget option.

How Self-Summarization Cuts Composer 2 Token Costs

How do you ship a competitive model at a tenth of Opus pricing? Part of the answer is distribution - Composer 2 only runs inside Cursor, not as a standalone API, which changes the economics. But the more interesting piece is a training technique Cursor calls self-summarization.

Agentic coding sessions generate long action histories. Every file read, edit, and terminal command adds to the context the model carries forward. Eventually that context exceeds the model's window, and something has to give. Traditionally, platforms either create a text summary of prior work or use a sliding window that drops older context. Both approaches lose information - the model forgets details, makes redundant decisions, and burns tokens redoing work it already completed.

Cursor's approach builds the summarization into the model's training. When a session hits a token-length threshold, the model pauses and compresses its own context to roughly 1,000 tokens - down from the 5,000+ that traditional compaction produces. Because this compression happens inside the reinforcement learning loop, the model learns which details to keep and which to discard. Cursor reports a 50% reduction in compaction errors compared to external summarization methods.

For your bill, this plays out in two ways. Smaller context per request means fewer input tokens billed on every subsequent action in a long session. And because the model retains the right context, it avoids re-reading files or redoing work - the kind of token waste that quietly inflates agentic session costs. The net effect: Composer 2 handles tasks spanning hundreds of sequential actions without the context bloat that makes long sessions expensive.

What Composer 2 Means for Your Cursor Bill

Composer 2 is available in Cursor now. If you're managing spend across a team, a couple of dynamics are worth watching.

Your model mix is shifting, possibly without anyone noticing. Cursor's Auto mode picks models based on intelligence, speed, and cost. With Composer 2 in the rotation, Auto will likely route more requests to the in-house model - competitive quality at dramatically lower pricing makes it the obvious choice. That should lower your average cost per token. But Cursor is also making the Fast variant the default rather than Standard, so the actual savings depend on which variant your developers end up on and whether anyone has overridden Auto in favor of pricier third-party models. Without visibility into your model distribution, you're guessing.

Composer 2 only exists inside Cursor. You can't call it through a standalone API. It's trained specifically for Cursor's agent tooling - semantic search, file edits, terminal operations - and that tight integration is part of what makes it competitive on the benchmarks. For teams already committed to Cursor, that's a feature. For teams weighing Cursor against direct Anthropic or OpenAI API access, it's a lock-in factor. The cheapest model per token only works in one place.

Wrapping Up

Cursor has shipped three Composer generations in five months, each one cheaper and more capable than the last. That pace makes the cost landscape for AI-assisted coding unusually dynamic - what your team spent per token last month may not reflect what it's spending now, even if nobody changed any settings.

The Cursor integration in Vantage breaks down spend by model, token type, and user. As Composer 2 enters the mix alongside Opus, GPT-5.4, and whatever ships next month, that visibility is how you'll know whether the savings are actually landing. For more on getting set up, see the Cursor integration docs.