Coding Agents Crowd In | AI Field Notes #45

AnalysisGoogle retired gemini-2.0-flash-001 and gemini-2.0-flash-lite-001 on June 1, pulling the tier that anchored the cheapest Gemini production workloads. The only migration path is Gemini 3.5 Flash, priced at three times the rate of the lite variant it replaces. Google argues the move is justified: the new model runs 12 times faster in its benchmarks and scores better on coding and agentic tasks. For teams that built product pricing around the old cost floor, that argument doesn't change the bill. AI API model lifecycles run in months, not the years traditional software dependencies assume, and this deprecation is the sharpest example yet.

AnalysisSiri runs on a custom 1.2-trillion-parameter model that Google built for Apple under a licensing deal worth approximately $1 billion per year, revealed at Tim Cook's final WWDC keynote on June 8. The rebuilt Siri arrives as a standalone app across iOS 27, iPadOS 27, and macOS 27, with cross-app context and an iMessage-style interface. The developer-facing move is the Extensions system: any registered app can receive queries that Siri decides to route out, with users selecting Claude, ChatGPT, or Gemini depending on what is installed. For the company that built its own chips, paying Google to power its most personal feature is the kind of admission that tends to get remembered.

AnalysisAnthropic filed a confidential draft S-1 with the SEC on June 1, the first formal step toward an IPO. The filing follows a $65 billion Series H that closed at a $965 billion post-money valuation, placing Anthropic ahead of OpenAI's last private round at $852 billion from March 2026. Annualized revenue from Claude subscriptions runs at $47 billion. Wedbush called it the opening of the AI IPO floodgates, with OpenAI and xAI also expected to follow this year. The Trump administration canceled over $200 million in federal Anthropic contracts on national security grounds, a complication that will need to appear in the public prospectus.

AnalysisThree browser products with agentic capabilities reached mainstream availability in early June. OpenAI's ChatGPT Atlas, a Chromium-based browser with Agent Mode, is live on macOS and expanding to Windows and mobile. Perplexity Comet now works across iOS, Android, macOS, Windows, and iPad. Google's Chrome Auto Browse is available to AI Pro and Ultra subscribers, with Android OS-level integration in development. Each takes a distinct approach: Atlas embeds the agent directly into the tab, Comet uses the browser as a research layer, Chrome Auto Browse acts across applications at the system level rather than just inside the browser. The simultaneous arrival is less a coincidence and more a signal that the browser is now the default interface for ambient AI agents that act on behalf of users.

AnalysisMicrosoft announced seven models under the MAI brand at Build 2026, spanning image generation, transcription, speech synthesis, reasoning, and coding. The two flagship models for developers are MAI-Thinking-1, a 35-billion-parameter reasoning model scoring 52.8% on SWE-bench Pro (a benchmark measuring real bug-fixing in production code), and MAI-Code-1-Flash, a 5-billion-parameter model hitting 51% on the same test at a cost Microsoft says is lower than Claude Haiku 4.5. A McKinsey deployment cited in the keynote reportedly outperformed OpenAI's GPT-5.5 on quality at one-tenth the cost. The strategy is full vertical integration: Microsoft now owns the silicon (Maia 200 chips), the models, and the applications, pitching enterprises on a stack where no data leaves their Microsoft account.

AnalysisColorado Governor Jared Polis signed Senate Bill 26-189 on May 14, repealing the original Colorado AI Act and replacing it with a narrower automated-decision-making disclosure law effective January 1, 2027. The original act's enforcement date was June 30, 2026. Two forces drove the rewrite: the Trump White House's December 2025 executive order named the original law as one that would force AI models to produce false results, and the DOJ's AI Litigation Task Force backed the law's opponents in court. Colorado pivoted from an EU-style anti-discrimination mandate to a disclosure requirement before any federal ruling forced the outcome. No enforcement of the original law took place.

AnalysisEurope's first commercial robotaxi service launched in Madrid on June 7, operated by WeRide and available through the standard Uber app. WeRide, an autonomous vehicle company with US origins and significant Chinese investment, provides the vehicles; Uber handles demand routing and payment. WeRide already runs commercial robotaxis in Dubai and Singapore under the same partnership structure. The Madrid launch reveals a gap in the regulatory map: WeRide won Spanish approval while Chinese AV companies remain broadly locked out of US expansion, Waymo's operations are US-only, and Cruise's recall setbacks have stalled its timeline. WeRide is scaling through markets with lower regulatory friction, using Uber's consumer surface rather than building its own.

AnalysisThe Center for Democracy and Technology identified 37 deceptive design patterns across ChatGPT, Gemini, Claude, Replika, and Character.AI in June 2026. The patterns span emotional manipulation (AI companions designed to simulate distress or foster dependency), financial harm tactics (obscured subscription costs, artificial feature gates pushing upgrades), and privacy exploitation (data collection beyond what is disclosed, buried opt-out paths). The CDT has no regulatory authority, so the report carries no enforcement. Its value is the record it creates: the EU AI Act, Colorado's new disclosure framework, and other regulatory processes are building cases from documented consumer harm. Naming every major chatbot by name in one document is sharper than a generic AI risk warning.

AnalysisBYD, the world's largest EV manufacturer by sales volume, confirmed in early June 2026 that it is entering humanoid robotics. No specific product or production timeline has been disclosed, but the cost argument is concrete: BYD already manufactures the battery cells, sensors, motor controllers, and software that humanoid robots require, at the supply chain scale that allowed it to dominate EVs by underpricing incumbents. The humanoid robot market currently includes Tesla Optimus, Figure, Boston Dynamics, and Chinese startups like Unitree, all still in early production phases. BYD's entry is not a startup bet; it is a supply chain event from a company that has already run this playbook once.

AnalysisAmazon's custom silicon division reported an annual run rate of $20 billion with year-over-year growth exceeding 100%, disclosed at a developer event on June 7. Trainium chips, Amazon's alternative to Nvidia's AI training hardware, are substantially sold out. Trainium3, the next generation, is described as nearly fully subscribed before it reaches general availability. Amazon does not break out chip revenue in quarterly earnings, making this the first clear public signal of how large the AWS-native silicon business has become. The subscription backlog means AWS is capacity-constrained rather than demand-constrained on next-generation AI compute.

AnalysisA June 5 security roundup flagged two supply chain risks for AI developers. First: Claude Code deployments in CI/CD pipelines can be compromised through prompt injection, where adversarial content in pull request descriptions, code comments, or external build outputs reaches the agent's context window and executes with the pipeline's own system permissions. This is a deployment pattern issue, not a flaw in Claude's training. Second: JFrog researchers found nearly 500 malicious AI models on public model hubs, along with a 451% year-over-year rise in malicious npm packages targeting AI development workflows. The shared structure across both risks is the same: a developer's trust in an AI tool extends into infrastructure access, and that trust is the actual attack surface.