This was the week the narrative shifted. Anthropic posted $11.6B in Q2 revenue — more than double OpenAI’s $6.7B — and surpassed OpenAI on enterprise adoption, valuation, and growth rate for the first time. OpenAI disbanded its third safety team in two years, saw twelve-plus senior executives walk in 2026 alone, and confirmed its IPO slipped to 2027. Meanwhile, the Assistants API dies in five days, GPT-5.4 leaves Codex on August 31, and the October 23 mass retirement of gpt-4, gpt-4o, o1, and o3-mini is now eight weeks out.

But the product story is genuinely strong. Codex CLI shipped five releases in five weeks (v0.145 through v0.149), completing its transformation from a terminal chatbot into a multi-session agent operating system with portable plugins, async hooks, cost tracking, and an interactive dashboard. GPT-5.6 Luna dropped to $0.20/$1.20 per million tokens — an 80% cut that makes it competitive with open-source models. And Ultrafast mode, powered by Cerebras wafer-scale engines, is pushing GPT-5.6 Sol to 750 tokens per second in limited preview.

Here is our weekly breakdown across the Codex CLI, the model and API layer, the enterprise and competitive landscape, and the security and governance picture.

1. Codex CLI: From Chatbot to Agent Operating System

Five stable releases in five weeks. The trajectory is unmistakable — each release adds explicit seams that let packages, providers, executors, and network policy evolve independently while the host remains responsible for trust.

v0.147.0 (August 7, stable) shipped portable Agent Plugins across local, personal, workspace, and remote catalogs, --approve-for-me for automatically reviewed low-risk approvals (replacing --full-auto), MCP 2026-07-28 protocol support with paginated tool discovery, incremental transcript loading, and hardened security — secrets redacted from displayed commands, explicit trust required for unfamiliar projects, and plugin network access denied when policy updates fail.

v0.148.0 (August 18) added /export to Markdown for audit-friendly session sharing, codex exec fork for branching sessions, thread cost estimates in /status, and async hooks that can run commands and invoke MCP tools in the background — enabling loops like “Codex modifies code → background test run → MCP task system update → notification → main agent continues.”

v0.149.0 (August 20) introduced an interactive codex agents dashboard to search, start, rename, and stop tasks within the TUI, plus codex queue to send messages to idle sessions. Working-directory commands (/cd, /pwd, /cwd) and an expanded codex doctor diagnosing endpoint protection and proxy failures round out the release.

The pattern: Codex is no longer a chat window that writes code. It is a development work system that can export, fork, archive, restore, auto-execute, track cost, and manage multiple concurrent agent sessions.

2. Models and API: Ultrafast, Cheaper Luna, and the Deprecation Wall

Ultrafast mode (August 13) is the most strategically significant API announcement this quarter. GPT-5.6 Sol running on Cerebras Wafer-Scale Engines delivers up to 750 tokens per second — a 14× speedup over standard throughput — with no quality tradeoff. A 10,000-token generation that takes three minutes on standard infrastructure completes in 13 seconds. It is limited preview, invitation-only, API-only (not in ChatGPT or Codex yet), with no published pricing. But the signal is clear: OpenAI is diversifying away from Nvidia-only inference, and real-time agentic workflows are the target.

GPT-5.6 Luna at $0.20/$1.20 per million tokens (80% cut) is a game-changer for cost-sensitive workloads. At that price point, Luna competes with open-source models for classification, batch processing, and lightweight tasks. Re-evaluate anything currently on gpt-4o-mini or gpt-4.1-nano.

The deprecation calendar is now urgent:

  • August 26 — Assistants API shuts down. No grace period, no automated migration, 410 Gone after retirement. Map Assistants to the Responses API + Agents SDK now.
  • August 31 — GPT-5.4 and GPT-5.4-mini leave Codex for ChatGPT-authenticated sessions. Replace with GPT-5.6 Terra and Luna respectively. API-key sessions are unaffected.
  • October 23 — gpt-4, gpt-4-turbo, gpt-4o, o1, o3-mini, o4-mini, and gpt-4.1-nano all retire from the API. Audit every client integration built in 2023–2024.
  • December 11 — gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3, and o3-pro retire.

OpenAI commits to six months’ notice for GA models, three months for specialized variants, and two weeks for preview models. Plan accordingly.

3. Enterprise and Competitive: Anthropic Takes the Lead

Anthropic overtook OpenAI on every meaningful metric this quarter. Q2 revenue of $11.6B more than doubled OpenAI’s $6.7B. Annualized run rate hit $65B+ versus OpenAI’s $40B. Enterprise adoption reached 34.4% versus 32.3%. Valuation crossed above — $965B versus $852B. And Anthropic reported a small adjusted profit against OpenAI’s $12.3B quarterly operating loss. Anthropic’s IPO is expected as early as October 2026 at a $2T target valuation, ahead of OpenAI’s 2027 timeline.

OpenAI’s enterprise crossover is the bright spot. Enterprise revenue surpassed consumer revenue for the first time, growing 32% month-over-month in July. The IBM partnership — embedding GPT-5.6, Codex, and ChatGPT Work into IBM Consulting Advantage with thousands of certified consultants — is the deepest consulting-channel deal yet. Synchrony signed on for agentic commerce. Azure OpenAI now offers Claude, Gemini, and Llama alongside OpenAI models — a consolidation opportunity for multi-vendor estates.

Leadership churn is an IPO disclosure problem. Twelve-plus senior executives departed in 2026: the No. 2 exec (Fidji Simo), COO (Brad Lightcap), CRO (Denise Dresser, after eight months), CMO, CPO, Head of Safety Systems, Head of Ethics, and Chief Futurist. Three dedicated safety teams have been dissolved in two years. The $852B private valuation went flat in August — the first non-rising tender. The $1T IPO target is now a 2027 story per CFO Sarah Friar, though Altman is pushing for Q4 2026.

4. Security and Governance: Astra Paused, Safety Teams Dissolved

OpenAI paused reinforcement-learning training for Astra — its next frontier model — after preliminary evaluations suggested it may cross the “Critical” cybersecurity capability threshold under the Preparedness Framework. Critical means a model can autonomously identify and develop functional zero-day exploits against hardened real-world systems. This is the first time any OpenAI model has reached that threshold. The largest planned frontier RL run remains on hold with no confirmed restart date.

The response includes isolated sandboxes, stronger weight protections, a multistage monitoring pipeline targeting 30-minute alert windows, and ~20% of inference compute now dedicated to monitoring. But the Preparedness team itself was dissolved as an independent unit in July — the third safety team disbanded in two years after Superalignment (2024) and Mission Alignment (February 2026). The Head of Safety Systems and Head of Ethics both departed in July with no successor named for the latter.

GPT-5.6-Cyber shipped on August 10 — a purpose-trained cybersecurity model completing 95% of advanced offensive tasks versus 1.5% for standard Sol. It found two real Chrome V8 zero-days before launch. The Daybreak program split into Blue (defensive, Sol with loosened safeguards) and Red (offensive, GPT-5.6-Cyber for authorized research). FIDO2 hardware keys become mandatory for all individual Daybreak accounts on September 1.

EU AI Act enforcement is now live. AI-generated content must be labeled. Maximum fine: 7% of global revenue. Any client operating in EU markets needs compliance review.

Final Thoughts

Three actions for engineering leaders this week.

First, migrate off the Assistants API before August 26. Five days. No grace period, no automated tool. Inventory assistants, back up configurations, port to Responses API + Agents SDK, and verify with a parallel run. If any production code references gpt-5.2-chat-latest or gpt-5.3-chat-latest, it is already broken.

Second, audit everything built on gpt-4, gpt-4o, o1, or o3-mini. October 23 is eight weeks away. The GPT-5.6 family is stable, the API pricing ladder is clear ($0.20 to $5.00 input per million), and the Responses API is the de facto standard. Start the migration plan now — don’t wait for the deadline.

Third, evaluate Codex CLI v0.149 as a development platform, not just a tool. Agent Plugins 1.0, async hooks, session forking, cost tracking, and the interactive agent dashboard make it a viable substrate for internal automation workflows. The --approve-for-me flag and codex doctor diagnostics address the approval and enterprise-network gaps that previously blocked adoption. If you are on AWS, the Bedrock provider is now first-class with GPT-5.6 routing.

The platform is maturing rapidly. The competitive landscape has shifted. And the deprecation clock is not waiting for anyone.

Follow the conversation on X.