Cursor says Composer 2 is “frontier-level at coding” and is priced at $0.50/1M input tokens and $2.50/1M output tokens, with a faster variant costing 3x more
38.0 — 40.0 — 56.9 — These quality improvements come from our first continued pretraining run …
Cursor
Context & Ripple Effects
Cursor’s Composer 2 launch was framed around an agent trained solely on coding data for autonomous, lengthy tasks, making the model’s token pricing relevant not just to chat use but to extended software workflows. The company later positioned Composer 2.5 as more reliable on long-running tasks and complex instructions, extending the same product arc from model capability toward sustained agent execution.
The announcement also ties quality gains to Cursor’s first continued-pretraining run. That makes it a concrete example of a coding-tool vendor seeking to differentiate through model work while exposing a speed-versus-cost choice at the API layer.
First-order effects
Cursor users can choose between a standard Composer 2 tier at $0.50 per million input tokens and $2.50 per million output tokens, or a faster option at roughly triple those rates.
Cursor pairs its coding-quality claim with explicit usage pricing, giving teams a direct basis to trade latency against model cost for coding-agent workloads.
Second-order effects
Competing coding-model providers face pressure to compete on the combined equation of coding performance, throughput, and token cost rather than benchmark claims alone; later coverage of Claude Sonnet 5’s token pricing shows how visible this comparison has become.
For customers running lengthy autonomous tasks, output-token pricing and the 3x fast-tier premium can shape which tasks are delegated to agents and when faster completion justifies higher spend.
Third-order effects
If coding agents continue moving toward sustained, autonomous work, pricing will increasingly be assessed as cost per completed software task rather than as isolated input and output token rates.
Continued pretraining by application-layer vendors could blur the boundary between developer tools and model providers, although the durability of that shift depends on whether such specialization produces repeatable task-level gains.
The trend: Coding-agent competition is shifting toward specialized models priced around the practical economics of long-running, latency-sensitive software work.
Cursor released their next in house model — Open weight base models + incredible ML teams in a specific niche can create immense value. Congrats to all my friends on the cursor research team! — They said they did continued pretraining to enable effective scaled RL, fun — c…
Composer 2 is out! Cursor is an example of a new type of company, not a pure app maker and not a model provider. Our aim is to build the most useful coding agents by combining the best API models and our domain-specific models.
fun fact about the Composer 2 RL run: we ran training distributed across 3 (sometimes 4) different clusters around the world using some secret sauce we built together.
We were able to significantly improve the model quality and cost to serve. These quality improvements come from our first continued pretraining run, providing a far stronger base to scale our reinforcement learning. [image]
Cursor have released Composer 2, a frontier-level coding model with higher token efficiency, and a faster default variant Looks like the rumor going around about it beating Opus 4.6 was in regard to Terminal-Bench 2.0, the only benchmark score they released aside from their own […
Composer 2 has been my daily driver for the past few weeks. It took a lot of self control to not post about how much I love using it for work. Give it a try, you won't regret it. Your wallet will thank you too.
Composer 2 just dropped in Cursor - i got early access for a few days and here are my first impressions: - it's very very fast so it feels amazing for sync coding or potentially as a daily driver - i've been using it for “flow-state coding” alongside Cursor Tab which has been
Cursor just launched Composer 2 - their own model. It's 10× cheaper than Opus 4.6 and supposed to rival it. I've been using it for a few days, I don't have any skewed graphs to show you but from a pure vibes POV I can tell you it's pretty good™ My litmus test right now is if it…
Open weight base models + incredible ML teams in a specific niche can create immense value. Congrats to all my friends on the cursor research team! (they said they did continued pretraining to enable effective scaled RL, super fun)
composer 2 is here beats opus 4.6 in terminal bench 2.0 input 0.5/M, 2.5/M output there's a faster variant too which has 3x price than normal variant. [image]