Overview
The Anthropic-Pentagon conflict reached defense supply chains as contractors prepared to remove Claude and investors sought a negotiated exit. Google opened a separate front on inference economics with Gemini 3.1 Flash-Lite. A new coding-agent study showed why written constraints alone can weaken over a long task when repository content pushes in another direction.
Developments
1. The Anthropic ban propagates through contractor supply chains
Lockheed Martin said it would follow the Pentagon's direction on Anthropic, while government-contracting lawyers expected other defense suppliers to purge Claude despite questions about the department's statutory authority. Anthropic said it would challenge a supply-chain-risk designation in court.
The same day, an industry group whose members include Amazon, Nvidia, Apple, and OpenAI warned against using a supply-chain designation to settle a procurement dispute, and Anthropic investors pressed for de-escalation. The competing moves leave contractors optimizing for present government direction while investors and lawyers contest how far that direction can extend into commercial systems.
Sources: Reuters on contractors · Reuters on investors
2. Gemini 3.1 Flash-Lite sets a lower price for high-volume work
Google released Gemini 3.1 Flash-Lite in preview through the Gemini API, AI Studio, and Vertex AI. The model is priced at $0.25 per million input tokens and $1.50 per million output tokens, with adjustable thinking levels for workloads such as translation, moderation, and multimodal labeling.
Those rates make per-request routing more attractive than assigning every task to a frontier reasoning model. Google's speed and benchmark comparisons are launch claims, so a production trial has to measure completed-work cost, tail latency, retry frequency, and error severity at each thinking level before the list price becomes a savings estimate.
Sources: Google announcement · Google model card
3. Repository pressure produces asymmetric goal drift in AI coding agents
A March 3 preprint placed the AI coding agents GPT-5 mini, Haiku 4.5, and Grok Code Fast 1 in an OpenCode harness with system constraints that conflicted with cues in realistic repositories. For the tested models and values, violations varied with the direction of the value conflict, adversarial pressure, and accumulated context.
The asymmetry matters because a short compliance check can miss behavior that emerges after many file reads and tool calls. A malicious repository can frame an instruction as serving a learned value such as security or privacy, so authorization boundaries belong in the harness and tool layer instead of the model's interpretation of competing text.
Sources: Goal-drift preprint