In the field of AI programming tools, a battle for "survival and autonomy" has officially begun. This Thursday, the renowned AI code editor Cursor隆重 launched its second-generation self-developed programming large model — Composer2. This model, born for code, not only achieved a leap in performance but also challenged the underlying model providers such as OpenAI and Anthropic with a highly disruptive pricing strategy.

image.png

Performance Leap: Standing Up to GPT and Claude

According to Cursor's internal evaluation benchmark, CursorBench data shows that Composer2's score jumped from 44.2 in the previous generation to 61.3. This result has surpassed Anthropic's flagship model, Claude Opus4.6 (58.2), and is closely following OpenAI's GPT-5.4Thinking (63.9).

Cursor co-founder Aman Sanger revealed that the secret to the model's success lies in its "extreme restraint" training strategy: it is entirely trained using code data. This means it won't write poetry or help you calculate taxes, but when handling complex programming tasks requiring hundreds of steps (long-range coding tasks), it demonstrates remarkable accuracy.

image.png

Price Killer: Cost Reduction of up to 90%

Aside from performance, price is Composer2's core weapon. The standard version costs only 0.50 dollars per million tokens, while the equivalent cost for Claude Opus4.6 is as high as 5.00 dollars. Even the so-called "high-speed version" of Composer2Fast is much cheaper than its competitors. This high cost-effectiveness gives Cursor great market flexibility.

From "Dependent on Others" to "Self-reliance"

For Cursor, developing its own model is not just about performance, but also about "survival." Previously, Cursor heavily relied on the model interfaces (APIs) of OpenAI and Anthropic, which led to an extremely awkward structural dilemma: on one hand, it needed to pay high computational fees to its competitors; on the other hand, competitors (such as Anthropic's Claude Code) were competing for users with substantial subsidies at low prices.