OpenAI and Anthropic Burned $27 Million Fighting One Lawmaker and Came Out Even

01OpenAI and Anthropic Spent $27 Million Against One New York Lawmaker. It Ended in a Draw.

Alex Bores, a New York state Assemblyman, narrowly lost the Democratic primary for New York's 12th Congressional district. The race carried a $27 million price tag, money that came out of a proxy fight between Anthropic and OpenAI, according to The Verge. That is an enormous sum for a single House primary. Few primaries for a House seat draw money on that scale.

The spending did not work the way its backers intended. Bores became more popular after a pro-AI super PAC began targeting him. The campaign against him drew attention beyond his Assembly district. He still lost, and by a narrow margin. The $27 million produced no decisive outcome for either side.

Behind the money sits a policy split. Anthropic and OpenAI hold opposing positions on how AI should be regulated, and that disagreement turned a local Democratic primary into a test case. The two companies poured resources into a race far from their headquarters.

A second AI fight in Washington runs the other direction. Rep. Anna Paulina Luna (R-FL) spent the week denying that her staff used AI to draft legislation. Accounts on X had circulated screenshots of an amendment summary attached to a major defense funding bill. The screenshots prompted questions about who, or what, wrote the text.

Luna's answer drew a sharp line. She says the AI handled "spellcheck" in the summary, not the bill language itself. "NO Legislation is ever drafted with AI," she said. By her account, the disputed material is the amendment summary, not the law.

The two episodes pull against each other. In New York, AI companies spent millions trying to choose who writes federal law. The Washington fight runs the opposite way, with a lawmaker denying that AI already sits inside the law. One side is about influence over legislators; the other is about AI reaching into legislation directly.

Bores lost his primary by a thin margin. The defense funding bill is still moving through Congress. Luna's only public response so far is the spellcheck denial.

02Computer use lands in Google's cheap Flash tier, as two new benchmarks map where these agents break

Google moved computer use into Gemini 3.5 Flash, the model it sells on speed and price. The capability had lived in a standalone Gemini 2.5 computer use model. Now it ships as a built-in tool, alongside the function calling and Search grounding Flash already does. Developers can point it at browser, mobile and desktop environments to build agents that read a screen, reason about it, and click.

Google pitches the upgrade at long-horizon work: continuous software testing, knowledge tasks across professional apps, automation that runs for many steps. It also ships guardrails for that setting. The company says it used adversarial training to blunt prompt injection, and released two optional enterprise systems. One demands user confirmation before irreversible actions. The other halts a task when it flags an indirect injection.

Long-horizon is exactly where two papers released the same week say these agents still fail.

PlanBench-XL drops an agent into an ecosystem of 1,665 tools across 327 retail tasks, then limits what it can see. The agent has to retrieve usable tools, invoke them to surface hidden sub-goals, and keep planning as conditions shift. Existing benchmarks, the authors write, rarely test planning under that retrieval-limited visibility.

MemGUI-Agent comes at the problem from mobile GUIs. Its authors report that multimodal agents handle short tasks but stay unreliable on long ones, the kind that require holding intermediate facts across many steps and app switches. They blame ReAct-style prompting, which piles up per-step records until the prompt bloats and critical cross-app facts get diluted.

One group working on tool retrieval, one on phone screens, both land on the same bottleneck: the longer the task, the worse the agent holds its thread. Flash makes that kind of task cheaper to attempt, not more reliable to finish.

03A24's fans spent years championing the anti-Hollywood studio. Its new backer is Google DeepMind.

The fans who turned A24 into a brand did it by treating the studio as the opposite of everything they distrusted about big Hollywood. So when Google DeepMind put $75 million into the company, the people who had championed it took the deal personally, and they said so loudly, according to Wired.

The studio has not pretended the anger away. A24 "knows you're mad," as Wired framed it, casting the backlash as something the company is bracing for rather than waving off. The money arrives as AI firms push deeper into film and television, buying footholds in an industry that spent two years fighting them in writers' rooms and on picket lines.

That history is what makes this particular check sting. A24's pull came from a promise of human authorship, of bets placed on directors instead of formulas. A frontier AI lab writing the check reads, to its core audience, as a quiet vote for the other side. The fans who built the brand by word of mouth now feel they boosted a company that crossed a line they care about.

The reaction is not confined to one fanbase. Cory Doctorow, the science-fiction author and tech journalist, has a new book out, The Reverse Centaur's Guide to Life After AI, on how to live once the technology sits in everything, according to Ars Technica. The piece runs under the headline "How to burst the AI bubble: Strike at its roots."

Doctorow's framing treats opposition as something to organize rather than wait out, giving the A24 revolt a wider intellectual backdrop. The studio sits at the front of that mood because it asked its audience to see it as a partner, not a vendor. That audience is now deciding what the partnership was worth.

A24 has not signaled any change to the Google DeepMind arrangement, and the AI money flowing into Hollywood is growing, not slowing.