Summary

July 9 was dominated by OpenAI's three-tier GPT-5.6 launch (Sol, Terra, Luna) and its most striking infrastructure angle: the top-tier Sol model running on Cerebras WSE-3 wafer-scale chips at up to 750 tokens per second, roughly 10x GPU throughput for a frontier model, under a $20B+ multi-year contract. Around it, the day's signals clustered into a broader shift toward faster, cheaper, and more openly sourced compute: Moonshot's open-weight Kimi K2.7 Code reached both GitHub Copilot and Microsoft Foundry, SpaceXAI shipped Grok 4.5, and OpenAI debuted its full-duplex GPT-Live voice models. Agentic coding tooling advanced in parallel (Codex controls in ChatGPT for iOS, multi-chat Claude sessions in VS Code 1.128), while two governance moves — Anthropic's biometric ID checks for consumer Claude and Cloudflare's purpose-based AI-crawler controls — signaled tightening rules around identity and automated traffic.

Key themes

  • Frontier inference goes wafer-scale: GPT-5.6 Sol on Cerebras WSE-3 hits ~750 tok/s (~10x GPU) via MemoryX weight streaming, tied to a $20B+ OpenAI contract, marking a serious non-GPU path for frontier serving.
  • Tiered, cheaper frontier models: GPT-5.6 arrives as Sol/Terra/Luna ($5/$30 down to $1/$6 per M tokens) with 90% cache-read discounts and parallel-subagent Ultra reasoning; SpaceXAI's Grok 4.5 lands at $2/$6.
  • Open-weight coding models enter proprietary toolchains: Moonshot's MIT-licensed Kimi K2.7 Code (1T params / 32B active) becomes the first open-weight model in GitHub Copilot and hits public preview in Microsoft Foundry the same week.
  • Agentic developer tooling matures: ChatGPT for iOS gains full Codex task management and diff filters, while VS Code 1.128 gives a single Claude agent session multiple parallel chats and surfaces subagent transcripts as peer chats.
  • Identity and traffic governance tightens: Anthropic can now require government ID + facial-geometry selfies (via Persona) from consumer Claude users, and Cloudflare rolls out purpose-based Search/Agent/Training crawler controls to all tiers, blocking Agent/Training by default on ad pages from Sept 15.
  • Real-time multimodal push: OpenAI's GPT-Live full-duplex voice models (GPT-Live-1 and mini) begin global ChatGPT rollout, listening and speaking simultaneously and delegating to a frontier model for deeper reasoning.

Notable items

  • GPT-5.6 Sol runs on Cerebras WSE-3 wafer-scale hardware at up to 750 tokens/second — ~10x GPU throughput for a frontier model — under a disclosed $20B+ multi-year OpenAI–Cerebras contract, initially for a limited set of customers.
  • OpenAI releases GPT-5.6 in three tiers: Sol ($5/$30 per M tokens, 91.9% on Terminal-Bench 2.1), Terra ($2.50/$15), and Luna ($1/$6), with prompt caching (90% cache-read discount) and an Ultra mode running parallel subagents.
  • Moonshot AI's Kimi K2.7 Code becomes the first open-weight model in the GitHub Copilot picker (MIT-licensed, 1T total / 32B active params, Azure-hosted), reaching Pro/Pro+/Max then Business and Enterprise on July 7.
  • Microsoft adds Kimi K2.7 Code to Microsoft Foundry in public preview (from July 1), giving Azure agent builders a first-party path to the same Chinese open-weight coding model.
  • SpaceXAI (formerly xAI) launches Grok 4.5 on a ~1.5T-parameter V9 foundation with Cursor coding data, available in Grok Build, Cursor, and the SpaceXAI console at $2/$6 per M tokens (not yet in the EU).
  • OpenAI introduces GPT-Live, full-duplex voice models (GPT-Live-1 and mini) rolling out globally in ChatGPT from July 8, using GPT-5.5 in the background and delegating to a frontier model for search and reasoning.
  • ChatGPT for iOS 1.2026.181 adds create/search/fork/manage for Codex coding tasks in-conversation, richer staged/unstaged/branch diff filters, an SSH host shortcut, and per-task model/reasoning and usage-limit visibility.
  • VS Code 1.128 lets one Claude agent session hold multiple related chats for branching and parallel work, shows subagent transcripts as read-only peer chats, and makes Copilot Vision image/PDF support generally available.
  • Anthropic's updated privacy policy (effective July 8) can require Free/Pro/Max consumer Claude users to submit a government photo ID and live facial-geometry selfie via Persona; API, Team, and Enterprise plans are exempt.
  • During Agents Week 2026, Cloudflare launches purpose-based control of Search, Agent, and Training crawlers for all customers including Free; from Sept 15, newly onboarding domains block Agent and Training crawlers by default on ad-serving pages.

Source coverage

Source rows used: 10