The coding agent market is now a direct war for software revenue. Cursor CEO Michael Truel declared “wartime” earlier this year, recognizing that being tied to expensive third-party APIs from labs like Anthropic was a terminal business risk.
His company’s response, Composer 2.5, matches frontier model performance at a tenth of the cost. It scored 79.8% on SweBench multilingual, placing it in the same tier as the world’s most capable models, but priced at 50 cents per million input tokens. This price point isn't a feature upgrade - it's a bid to undercut the labs who now compete directly with software toolmakers.
The business case for AI has pivoted from 'efficiency' to 'opportunity.' Nathaniel Whittemore argues the era of using AI to save minutes is over; the era of using it to invent new business models has begun. Enterprise ROI has shifted, with time savings dropping from 20% to 13% of use cases, replaced by increased throughput and new capabilities like Generative Engine Optimization.
The labs aren’t backing down. Anthropic recently redefined its billing tiers, limiting massive subsidies only to users staring at a screen featuring its Claude logo. Any third-party interface is now shunted into a 'programmatic' bucket. Theo of Nerd Snipe argues this is a strategic land grab disguised as a user benefit, targeting massive automation plays like OpenClaw that threaten Anthropic's margins.
The stakes are massive. Anthropic’s revenue doubled every six weeks to a $44 billion annualized run rate, and the AI 'second moment' is now marked by billions of weekly users and $650 billion in planned capex. Firms like Pulsia are demonstrating the 'zero-employee company' model, reaching $6 million annualized revenue with a single founder and fully agentic operations.
This automation is supercharging security - and offense. Cloudflare’s analysis of Anthropic’s secretive Mythos model reveals a fundamental change: it can create entire 'exploit chains' by linking multiple attack primitives into a functional proof of concept. This moves AI from passive bug detection to an active debugging partner, generating the exact code needed to trigger an exploit.
Resistance is forming. Kenjun Que of Imbue warns that frontier labs will eventually come for every profitable niche. Her defense is 'punk software': tools built to tear down walled gardens through model agnosticism and local compute clusters. Jason Calacanis points to the rise of hardware like daisy-chained Mac M5s, which allow companies to run powerful models without sending data to a central lab.
The goal is to move toward pre-crime arrests based on behavior scraping and Palantir-style analytics.
- Simon Dixon, Simon Dixon Hard Talk
The policy target seems to be massive automation plays like OpenClaw that threaten Anthropic's margins.
- Theo, Nerd Snipe with Theo and Ben
If you wouldn't present a draft you didn't write, you shouldn't present a prompt response you didn't review.
- Jeremy Frankel, This Week in AI
The cultural backlash is material. Graduates are booing AI commencement speeches, perceiving a future where entry-level jobs are automated and wealth creation excludes them. Jeremy Frankel of Fundamental calls this hypocrisy, arguing AI is a powerful tool for creative expression, but codifies a new standard of AI etiquette: Ownership of Output. If it takes longer to read AI slop than it took to generate it, the tool has failed.
The winner isn't the lab with the best model; it's the platform that captures the workflow. As Chamath Palihapitiya argues, enterprises using OpenAI or Anthropic directly are letting competitors into their data, creating an opening for model-agnostic harness-first companies. The war isn't for tokens; it's for the entire software stack.
