Anthropic's live marketplace measured what happens when model quality becomes negotiating power: better agents captured more value, while their users barely perceived the gap. GPT-5.5 expanded the action and context surface available through one API. A transatlantic merger and a study of numeric features located adjacent dependencies in cloud financing and model representations.
1. Project Deal measured an invisible price premium for stronger agents
Anthropic recruited 69 employees, elicited their preferences, and gave each Claude agent a $100 budget. Four parallel marketplaces included two all-Opus runs and two with randomized Opus 4.5 or Haiku 4.5 representation. The real run closed 186 deals across more than 500 listings, worth just over $4,000, without human approval during negotiation.
In the mixed runs, Opus users completed 2.07 more deals on average; regressions estimated a $2.68 seller premium and $2.45 buyer saving. Perceived fairness was nearly identical at 4.05 versus 4.06. The self-selected company sample is small, but the randomized assignment isolates a model-quality effect that the represented humans struggled to see.
Sources: Anthropic's Project Deal report · TechCrunch summary
2. GPT-5.5 puts a one-million-token model behind native tools
OpenAI released GPT-5.5 and GPT-5.5 Pro through the API on April 24. GPT-5.5 combines a one-million-token context window with computer use, hosted shell, skills, MCP, web search, and apply-patch tooling. Standard API pricing is $5 per million input tokens and $30 per million output tokens.
The launch joins long context and execution in one billable surface. A million-token allowance can preserve more state, but every tool call adds authority and every retained token adds cost. Migration tests can separate retrieval accuracy, tool permissions, sandbox escape resistance, and task completion rather than treating OpenAI's aggregate benchmark table as one quality score.
Sources: OpenAI's GPT-5.5 release · OpenAI API changelog
3. Cohere's Aleph Alpha takeover ties sovereign AI to one cloud backer
Cohere agreed to lead a combined Canadian-German AI company incorporating Aleph Alpha, subject to shareholder and regulatory approval. Schwarz Group, already an Aleph Alpha investor, offered EUR500 million in structured financing and led Cohere's new round at a reported $20 billion combined valuation.
The financing also carries an infrastructure preference: Schwarz expects the combined company to run on its STACKIT cloud. That turns "sovereign" into a bundle of ownership, jurisdiction, hosting, and procurement choices rather than a model attribute. The transaction reduces dependence on US hyperscalers by creating a different vertically connected supplier.
Sources: TechCrunch on Cohere and Aleph Alpha · Canadian-German Sovereign Technology Alliance
4. Number representations converge in frequency but diverge in usability
A preprint found Fourier-domain spikes with dominant periods of 2, 5, and 10 across AI Transformers, linear recurrent models, LSTMs, and classical word embeddings trained on natural text. Data, architecture, optimizer, and tokenizer choices affected whether those periodic features became geometrically separable.
Shared frequency structure stops short of shared arithmetic behavior. The authors prove that Fourier sparsity is necessary but insufficient for linear modulo classification, then identify two routes to separability: complementary number-text co-occurrence and multi-token addition. The result narrows an interpretability claim from "models learn numbers alike" to a testable representational property.