Three AI coding agents launched in 72 hours last week compressed what developers pay for frontier-level capability more sharply than any single release this year. On May 18, Cursor released Composer 2.5, its latest in-house coding model. The next day, Anthropic held Code with Claude London, its first developer event outside the United States, and announcedtwo new infrastructure features for Claude Managed Agents.
Also on May 19, Alibaba’s Qwen 3.7 Max API went live. Each of these would have been its own story. Together, they landed on a market where pricing had already been sliding for months and pushed it further down.
What Cursor shipped
Composer 2.5 is Cursor’s third-generation proprietary coding model, built on the same open-source Kimi K2.5 base as its March predecessor, but trained on 25 times as many synthetic coding tasks. Cursor named the base model upfront this time, having faced community criticism in March when the Kimi foundation was not clearly disclosed at launch.
The standard tier is priced at US$0.50 per million input tokens and US$2.50 per million output tokens. A faster default variant runs at US$3.00 input and $15.00 output, and Cursor doubled usage limits for the first week after launch. On Cursor’s own CursorBench v3.1, Composer 2.5 scores around 63% accuracy at roughly US$0.50 per task, while Claude Opus 4.7 at its default setting scores comparably at approximately US$7 per task.
That is the same capability bracket at a fraction of the cost, by Cursor’s own measurement. Vendor benchmarks on vendor infrastructure warrant the usual scepticism, but the pricing gap is real regardless of how the numbers hold up in independent testing.
What Anthropic shipped in London
Code with Claude London was Anthropic’s first dedicated developer event in Europe. The two announcements addressed what has kept many enterprise and regulated-sector teams from deploying Claude agents at scale: data leaving their perimeter.
Self-hosted sandboxes, now in public beta, let teams run Claude Managed Agents and execute tools within their own infrastructure. The agent orchestration loop stays on Anthropic’s side, but the surface where code runs, files are written, and network calls go out stays inside the customer’s environment. Launch partners include Cloudflare, Daytona, Modal and Vercel, with a bring-your-own-sandbox option available as well.
MCP tunnels, currently in research preview, let Claude agents connect to private internal systems and databases without requiring a public endpoint. Traffic is encrypted end-to-end and routed through a lightweight gateway deployed inside the private network.
Both features come with caveats that are worth stating plainly. Self-hosted sandboxes are public beta, not general availability, and MCP tunnels carry explicit as-is language in documentation and require access approval before use. Teams that need stability guarantees today are not the target audience yet, but the architectural gap that has been blocking enterprise adoption is narrower than it was a week ago.
What Alibaba shipped
Qwen 3.7 Max API went live on Alibaba Cloud Model Studio on May 19, a day before its formal announcement at the Alibaba Cloud Summit in Hangzhou. The model is closed-weight, a departure from Alibaba’s usual pattern of releasing open weights alongside hosted APIs. No weights have appeared on Hugging Face as of this writing.
Pricing sits at US$2.50 per million input tokens and US$7.50 per million output tokens, with a 90% discount on cached input tokens, bringing that down to $0.25 per million. On the Artificial Analysis Intelligence Index, it scores 56.6, ranking as the highest-placed Chinese model on that leaderboard at launch. SWE-Bench Verified sits at 72.5.
There is a practical catch worth flagging. Extended thinking is enabled by default, which makes Qwen 3.7 Max verbose in long agent sessions. Developers running it in coding loops report effective costs running three to four times the headline rate unless max_tokens is capped manually. The $2.50 rate card is not what most teams will actually pay in production without that adjustment.
One detail relevant to Claude Code users specifically: Qwen 3.7 Max natively supports the Anthropic Messages protocol, meaning it can be dropped into an existing Claude Code harness without rewriting integration logic.
What this means for tooling decisions right now
The practical effect of all three launches together is that frontier-adjacent coding capability now has multiple competitive price points rather than one. Six months ago, running a capable AI coding agent in production meant budgeting for Opus-class API rates or accepting a meaningful capability drop. That trade-off is narrower than it was.
Cheaper tokens do not automatically mean better outcomes, though. The real cost of any AI coding agent is the token rate plus the review burden the agent’s output creates, plus the overhead of managing it inside a security and compliance stack. A cheap model is only cheap if the resulting diff is scoped and easy to inspect, a point Cursor’s own documentation makes explicitly. That caveat applies across all three of these launches.
Alibaba’s move is also worth watching as a strategic signal beyond the pricing. Entering the API market with a closed flagship model aimed at enterprise developers, rather than seeding the open-source community as it has historically done, marks an intent shift. Whether Qwen 3.7 Max holds its benchmark position once independent testing catches up is a separate question from whether that intent has changed. It has.
See also: Software development in 2026: Curing the AI party hangover
Want to dive deeper into the tools and frameworks shaping modern development? Check out the AI & Big Data Expo, taking place in Amsterdam, California, and London. Explore cutting-edge sessions on machine learning, data pipelines, and next-gen AI applications. The event is part of TechEx and co-located with other leading technology events. Click here for more information.
DeveloperTech News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.