Skip to content
Source · Daily Brief

AI Daily Brief — 7 August 2025

The day the AI calendar bent around one event. OpenAI shipped GPT-5 across every Microsoft surface. Coding partners endorsed. Altman pitched “PhD-level expert” — backlash started within hours. Decart raised $100M at $3.1B.

Top stories

  • OpenAI launches GPT-5 in 10am PT livestream. One-hour “LIVE5TREAM” rolled GPT-5 across ChatGPT, API, and GitHub Models Playground simultaneously. Free users get access (tight caps), Plus higher limits, Pro unlimited + “GPT-5 Pro” thinking variant. via OpenAI
  • GPT-5 ships as unified system with real-time router. Two underlying models — “GPT-5-main” (fast) and “GPT-5-thinking” (deep reasoning) — orchestrated by a real-time router. OpenAI said it plans to merge them into a single model in the near future. via OpenAI System Card
  • GPT-5 API pricing undercuts GPT-4o. GPT-5 ($1.25 input / $10 output per M tokens), GPT-5 mini ($0.25 / $2.00), GPT-5 nano ($0.05 / $0.40). Four reasoning levels (minimal/low/medium/high). 400K context (128K output). Cutoffs: Sept 30 2024 (full) / May 30 2024 (mini, nano). via Simon Willison
  • GPT-5 benchmark numbers headline. 74.9% SWE-bench Verified, 88% Aider Polyglot, 94.6% AIME 2025 (no tools), 96.7% τ2-bench telecom (tool-calling benchmark released only two months earlier). HealthBench Hard: 46.2% vs o3’s 31.6%. via OpenAI
  • Coding partners ship same-day endorsements. Cursor called GPT-5 “the smartest model we’ve used”; Vercel: “the best frontend AI model”; Windsurf: half the tool-calling error rate of other frontier models and SOTA on their evals. OpenAI explicitly tuned GPT-5 for Cursor, Windsurf, GitHub Copilot, Codex CLI. via OpenAI Dev Blog
  • GPT-5 lands across Microsoft stack on day one. Microsoft 365 Copilot, consumer Copilot, GitHub Copilot, Azure AI Foundry on launch day — default reasoning option across Microsoft’s enterprise AI portfolio.

Who shipped

OpenAI shipped GPT-5. Microsoft shipped GPT-5 everywhere. Anthropic shipped revocation.

Open-source pulse

Altman pitched GPT-5 as “PhD-level expert” — backlash started quickly. Users posted screenshots of basic geography and math errors; Altman later admitted a broken auto-router “made GPT-5 seem way dumber” and eventually restored GPT-4o for Plus. via TechRadar

Money, infra & hardware

Decart raised $100M Series B at $3.1B valuation — Tel Aviv-based AI lab for real-time generative video. Sequoia, Benchmark, Zeev Ventures led with Aleph VC joining. Funds its “livestream diffusion” models Oasis and newly released MirageLSD. via Fortune Anthropic answered with Claude Opus 4.1 two days earlier (74.5% SWE-bench Verified) — pre-empted the GPT-5 event.

Quiet corners

Anthropic revoked OpenAI’s Claude Code access after catching OpenAI engineers using it ahead of GPT-5 launch. via BleepingComputer GPT-5 positioned as “best model ever for health” — new high on HealthBench (built with 262 physicians across 5,000 realistic clinical conversations).

By the numbers

  • 400K / 128K — GPT-5 context window / output max
  • $1.25/$10, $0.25/$2, $0.05/$0.40 — GPT-5 / mini / nano per M tokens
  • 74.9% / 94.6% / 96.7% — GPT-5 SWE-bench / AIME 2025 / τ2-bench telecom
  • $100M / $3.1B — Decart Series B / valuation
  • 120B / 20B / Apache 2.0 — gpt-oss sizes / license (released Aug 5)

Compiled by AI Feed’s editor from verified web sources for 7 August 2025.