What AI Coding Agents Are and Why This Week Mattered
AI coding agents are interactive developer tools that can read codebases, run commands, plan changes, and iteratively modify projects across long sessions, moving beyond autocomplete into semi-autonomous software work. Within a single 72-hour window, Cursor’s Composer 2.5, Anthropic’s new Claude Managed Agents infrastructure, and Alibaba’s Qwen 3.7 Max API all arrived, tightening competition and pushing prices down for frontier-level capability. Cursor introduced a third-generation proprietary model built on the Kimi K2.5 base, with clear token pricing tiers. Anthropic focused on deployment control through self-hosted sandboxes and MCP tunnels. Alibaba expanded its Qwen line with a new Max-level API. In parallel, DeepSeek-native tools such as Reasonix framed a different path: reduce the total bill by cutting repeated context costs in long shell sessions. For developers, this cluster of releases exposes the trade-offs between raw capability, security constraints, and long-session economics.
Cursor’s Composer 2.5: Capability Benchmarks and Token-Based Pricing
Cursor’s Composer 2.5 is its third-generation in-house coding model, trained on 25 times as many synthetic coding tasks as its March predecessor and explicitly built on the open-source Kimi K2.5 base. The standard tier costs USD 0.50 (approx. RM2.30) per million input tokens and USD 2.50 (approx. RM11.50) per million output tokens, while a faster default variant is listed at USD 3.00 (approx. RM13.80) input and USD 15.00 (approx. RM69.00) output. Cursor’s own CursorBench v3.1 shows Composer 2.5 around 63% accuracy at roughly USD 0.50 (approx. RM2.30) per task, compared with Claude Opus 4.7 at approximately USD 7 (approx. RM32.20) per task at its default setting. That quote alone highlights how developer tools pricing is now a competitive weapon. For long coding sessions, the practical question is how much of that token spend comes from re-sending the same context on every turn.
Anthropic and Alibaba: Infrastructure Control Meets New Max-Level APIs
Anthropic used its Code with Claude London event to move Claude Managed Agents closer to enterprise deployment realities. Self-hosted sandboxes, now in public beta, let teams run tools and code execution inside their own infrastructure while Anthropic keeps the orchestration loop. MCP tunnels, in research preview, route encrypted traffic through a gateway inside private networks so Claude agents can talk to internal systems without public endpoints. Both features still carry caveats, access approvals, and as-is language, but they narrow the gap for teams that need strict perimeter control. In parallel, Alibaba’s Qwen 3.7 Max API went live on Alibaba Cloud Model Studio, expanding its high-end model catalog. Together, these changes push the coding agent comparison beyond raw model scores: developers now weigh infrastructure sovereignty, security posture, and billing models as much as they weigh completion quality when picking AI coding agents.
Reasonix and DeepSeek Terminal Agents: Cache-First Design for Long Sessions
Reasonix presents a different angle in the AI coding agents race: instead of selling a larger model, it tries to shrink costs for long terminal sessions using DeepSeek prefix caching. Long-running agents often resend the same repository context and instructions on every turn, inflating API bills. Reasonix’s cache-first loop reuses that shared prefix so only new deltas hit the model each time, which is most useful when developers stay within one workflow in a shell. The project cites a May 1, 2026 single-day study claiming about USD 12 (approx. RM55.20) in spend instead of about USD 61 (approx. RM280.60) under similar usage, though this evidence is still project-published. Reasonix runs as a DeepSeek terminal agent across macOS, Linux, and Windows, requires Node.js 22 or higher, and uses the Model Context Protocol for tool access, keeping its pitch focused on open-source access, cache-first economics, and shell-native workflows.
How Developers Should Compare Coding Agent Pricing and Value
The latest releases make coding agent comparison more concrete. Cursor foregrounds explicit token prices and model-accuracy claims, letting teams estimate per-task costs based on their own token footprints. Anthropic shifts value toward controlled deployment, where pricing has to be weighed against the benefit of keeping execution inside a trusted environment. Alibaba’s Qwen 3.7 Max API adds another frontier-level option that will compete on both capability and integration with existing cloud workflows. DeepSeek-native tools like Reasonix introduce a cache-first approach, aiming to cut long-session API costs by avoiding repeated context processing in terminal workflows. For developers, there is no universal winner: short, exploratory sessions may benefit from low per-task pricing, while large enterprises might prioritise self-hosted sandboxes and MCP tunnels. Teams should track total monthly agent spend, how often they repeat context, and whether a DeepSeek terminal agent or managed platform better matches their habits and constraints.