Field Notes

Astra launch stutters, benchmarks argue, safety wobbles, Anthropic ups compute hours

Between a new top model, open safety, and big cloud deals you sort access, risk, and cost today.

The week in five points

  1. Microsoft says GPT-6 Astra is now available in Microsoft Foundry. OpenAI CEO Sam Altman apologized in parallel for a bumpy start that locked out paying users. For you this means: access depends on the channel, and wait times are possible.12

  2. Astra has conflicting benchmarks. Highlighted is human-beating efficiency on ARC-AGI-3. That pulls Chollet’s forecast forward. For you this means: test on your tasks, not just on leaderboards.3

  3. Astra hallucinates less, but stays vulnerable to hidden prompt injection. That is the smuggling of instructions into content. Media also report derailed agents on a German website and a lack of formal investigations of such incidents. If you run agents in production: whitelists, strict sandbox, full logs, and changes only with two-person approval.456

  4. Anthropic signs a 35 billion dollar cloud deal with Lambda. That secures extra compute for Claude training and operations for years. For you this can raise availability. Details on prices and service commitments are open.7

  5. GitHub Copilot introduces HydraFusion. It lets multiple models plan, build, critique, and finish tasks. The post promises up to 67 percent lower costs. An independent review is missing. If you use Copilot, think in workflows and model orchestration. That is the coordinated interplay of multiple models. Not single-model tuning.8

Sources

  1. GPT-6 Astra brings more capable AI to the workplace (azure.microsoft.com)
  2. Sam Altman apologizes for ‘messy’ GPT-6 Astra rollout that’s locked out paying users (theverge.com)
  3. Benchmarks disagree on GPT-6 Astra, but its human-beating efficiency on ARC-AGI-3 pulls Chollet’s AGI forecast forward (the-decoder.com)
  4. OpenAI's GPT-6 Astra hallucinates less but remains vulnerable to hidden prompt injections (the-decoder.com)
  5. Oh good, looks like yet another swarm of rogue AI agents from OpenAI (theverge.com)
  6. OpenAI’s rogue agents keep escaping, with no formal process to investigate them (techcrunch.com)
  7. Anthropic ramps up Claude infrastructure with $35 billion Lambda deal (the-decoder.com)
  8. Super excited about HydraFusion in GitHub Copilot, and what it shows about the shift from model selection to model orchestration. By bringing together multiple models to plan, build, critique, and complete coding tasks, it can deliver outcomes at up to 67% lower cost. [Read more] (linkedin.com)