Cursor releases Composer 2.5 SDK for Python and TypeScript

• cursor_ai's launch post says Cursor SDK now exposes Composer 2.5 in both Python and TypeScript for building custom agents.
• The launch came with a temporary pricing hook: cursor_ai's post and Eric Zakariasson's discount note both say Composer usage in the SDK is 90% off for the long weekend.
• Artificial Analysis' coding agent benchmarks, as summarized by ArtificialAnlys' benchmark post and ArtificialAnlys' model details thread, still place Composer 2.5 near the top on cost per task.
• ArtificialAnlys' model details thread says Composer 2.5 standard is priced at $0.50 per million input tokens and $2.50 per million output tokens, while Fast is $3.00 and $15.00.
• Early user chatter from TheZachMueller's repost and kevinkern's multitask post points to Composer 2.5 being used for long-running docs, planning, and multi-agent task splits, not just single-shot edits.

You can read Cursor's Composer 2.5 launch post, browse the SDK landing page, and jump straight to the Python SDK docs. The other useful reference is Artificial Analysis' coding agent index, which is where the cost-per-task claims in ArtificialAnlys' benchmark post and ArtificialAnlys' detailed thread come from.

Cursor kept the SDK announcement tight: custom agents, Composer 2.5, Python, and TypeScript. The official post links directly to the Python docs, while Eric Zakariasson's SDK pointer separately points builders to the SDK package page.

• cursor_ai says the SDK now supports building agents with Composer 2.5.
• The official docs linked from that post land on Python SDK docs.
• Eric Zakariasson's SDK pointer links to the broader SDK page, which gives Cursor a programmatic surface outside the IDE and CLI framing from the original Composer release.

The real ship vehicle here is pricing. Cursor paired the SDK release with a weekend discount large enough to push experimentation into impulse-buy territory.

Eric Zakariasson's discount note calls out 90% off Composer 2.5 in the SDK, and cursor_ai's launch post repeats the same offer in the official announcement. That discount matters more because Composer 2.5 was already competing on low cost in Cursor's own stack.

The benchmark backdrop is why the SDK story matters. Cursor is not exposing a prestige model first, it is exposing the cheap one that has been overperforming its price tier.

According to ArtificialAnlys' model details thread, Composer 2.5 scored 62 on the Coding Agent Index, up from 48 for Composer 2. The same thread says standard Composer 2.5 costs about $0.07 per task, Fast costs about $0.44, and both sit well below the higher-effort Opus 4.7 and GPT-5.5 configurations above them.

ArtificialAnlys' benchmark post adds that Composer 2.5 used 1.6 million tokens across its Coding Agent Index runs, versus as much as 5.7 million for other leading models. That is a cleaner explanation for the pricing story than token list price alone.

Artificial Analysis says Cursor is serving one model in two operating points, and the tradeoff is pretty blunt.

• ArtificialAnlys' Fast versus standard breakdown measured Fast at 6.7 minutes per task versus 9.3 minutes for standard, about 39% faster.
• The same post says Fast costs about 6x more per task, aligned with a 6x token pricing jump.
• WesRoth's benchmark summary separately notes that Fast is one of the quickest agents tested, but the premium buys responsiveness, not a different base model.

The most interesting early usage signals are about harness design, not raw evals. People are describing Composer 2.5 as something they hand heavier planning and coordination work to.

kevinkern's multitask post describes using Cursor's /multitask mode to send different models to different tasks and track plan history through plandb, with verify loops added when a task fails. Separately, TheZachMueller's repost says Composer 2.5 is already landing for heavy work on docs, configs, skills, plans, and prompts, which is a more specific early usage pattern than the benchmark charts show.