AI coding agents and the new price race
AI coding agents are interactive development assistants that read code, run tools, and update files autonomously or semi-autonomously, while charging for each token of context they process across often long-running sessions. Within a 72-hour window, Cursor’s Composer 2.5, Anthropic’s new Claude Managed Agent features, and Alibaba’s Qwen 3.7 Max API all arrived on top of an already falling price floor, tightening competition on both capability and cost for developers. Cursor pushed its own proprietary coding model forward with a clear token-based price sheet, Anthropic focused on infrastructure control rather than headline rates, and Alibaba added another frontier-scale option to cloud APIs. In parallel, Reasonix entered from a different angle as a DeepSeek-native terminal agent that tries to cut long-session expenses via prefix caching instead of larger models, giving developers a fresh choice in cost-first design.
Cursor’s Composer 2.5 shifts cost benchmarks
Cursor’s Composer 2.5 is its third-generation AI coding agent model, built on the Kimi K2.5 base and trained on 25 times as many synthetic coding tasks as its March predecessor. Cursor now states the foundation model explicitly, a response to criticism when the Kimi origin was not clear earlier. The company prices its standard tier at USD 0.50 (approx. RM2.35) per million input tokens and USD 2.50 (approx. RM11.75) per million output tokens, with a faster default at USD 3.00 (approx. RM14.10) input and USD 15.00 (approx. RM70.50) output. Cursor says Composer 2.5 scores around 63% accuracy on CursorBench v3.1 at about USD 0.50 (approx. RM2.35) per task, while Claude Opus 4.7 scores similarly at about USD 7 (approx. RM32.90) per task, highlighting how developer pricing models now vary widely even at similar capability levels.
Anthropic and Alibaba focus on infrastructure and reach
Anthropic’s update did not change per-token rates but changed how enterprises consume AI coding agents. Self-hosted sandboxes, in public beta, let teams run Claude Managed Agents while executing tools entirely inside their own infrastructure, so agent orchestration stays on Anthropic’s side but files and network calls remain on the customer’s systems. MCP tunnels, in research preview, give Claude agents encrypted routes into private internal systems without a public endpoint. Both features are early-stage and carry clear caveats, but they narrow a long-standing gap for regulated teams that need stricter control. Meanwhile, Alibaba’s Qwen 3.7 Max API went live on its Model Studio platform, expanding the field of high-end models available through cloud APIs. Together, these moves show that API cost reduction now depends not only on token prices but also on where and how code runs in production.
Reasonix and cache-first design for long sessions
Reasonix enters the market as a DeepSeek-native terminal AI coding agent aimed at cheaper long shell sessions instead of bigger models. Its design uses DeepSeek prefix caching so that repeated repository context and instructions do not need full reprocessing on every turn, which can reduce API cost for long, continuous workflows in the terminal. The project highlights a cache-first loop, plan mode, and first-class MCP support, all wrapped in a shell-first workflow for macOS, Linux, and Windows with a Node.js 22 requirement. According to WinBuzzer, the project’s own single-day study reports about USD 12 (approx. RM56.40) in usage instead of about USD 61 (approx. RM286.10) under comparable conditions, although this remains self-published evidence rather than independent benchmarking. With an MIT license, Reasonix appeals to developers who want open-source control while experimenting with API cost reduction strategies for DeepSeek terminal agents.
Cost-to-performance trade-offs for developers
For developers choosing between AI coding agents, the emerging pattern is not a single cheapest option but a graph of cost-to-performance trade-offs. Cursor targets frontier-level capability at comparatively low token prices for hosted editor workflows. Anthropic focuses on enterprise-grade control with self-hosted sandboxes and MCP tunnels, which may justify higher spend for teams with strict compliance needs. Alibaba’s Qwen 3.7 Max broadens the supply of powerful cloud models, adding pressure on pricing without yet defining a clear developer pricing model narrative. Reasonix, by contrast, treats the DeepSeek terminal agent as a cache-first tool designed to curb long-session costs when developers reuse the same context for hours in the shell. In such a crowded market, MCP support, plan mode features, and pricing for extended sessions all matter; teams need to evaluate not only headline rates but also how each agent behaves over an entire working day.