The Brief

New briefs daily around 7 AM Eastern

Gemini 4 Carbon tops AA Intelligence Index at 59, edging Opus 5.5

Yesterday's v5 index had Opus 5.5 alone on top at 58. Carbon takes the lead by one point at $4/$20 per million tokens, but its hallucination rate rises to 11% from Argon's 4%, and access is limited to selected users and vetted cyber defenders.

• 01Read

  GitHub's Copilot Fusion routes Claude and GPT at half Claude Code's cost

  Artificial Analysis measured $5.59 per task for Fusion's max setting against $11.42 for Claude Code on Opus 5.5, with a score 2 points lower. Teams paying frontier rates on every coding task now have a router that halves that bill for a small quality drop.

  GitHub Universe: Copilot Fusion, Automations and dreaming→Copilot Fusion: model routing and critique patterns explained→Artificial Analysis→
• 02Read

  AutomationBench finds the top agent completes about half its tasks

  It grades what actually changed across 47 simulated apps instead of what the agent reports, and most failures are agents claiming success on tasks they did not complete. Any workflow that trusts an agent's own completion message carries that same blind spot.

  Fireworks→
• 03Read

  OpenAI adds up to $500 monthly API credits to ChatGPT Pro and Business

  API usage was billed separately from a ChatGPT subscription. Pro users now get up to $500 a month by tier and Business workspaces share a pooled $500, so prototyping an agent within that amount no longer needs its own API budget.

  OpenAI Developers→
• 04Read

  Mistral opens its Responses API in public preview for agent builders

  Each response returns the answer, the reasoning and every tool call as separate items, so agent code reads each step directly instead of parsing one Chat Completions message. It works with every Mistral-hosted chat model, including Large 4.

  MistralAI→
• 05Read

  AssemblyAI-TTS-1 ranks #12 in AA voice arena at $15 per 1M characters

  It is now callable through OpenRouter with 4,000+ voices and US or EU serving. Artificial Analysis measured 162 characters per second and 81.7% pronunciation robustness, so voice-agent teams have independent numbers rather than only the vendor's claims.

  OpenRouter→Artificial Analysis→
• 06Read

  Anthropic details preclinical Claude results with six drug partners

  One partner cut manufacturing document review time by 89%, and Claude produced 35,000 antibody designs on its own. Every result is preclinical, so the evidence is about time saved in design and paperwork, not about drugs proven in patients.

  Anthropic→