GPT-6 Astra’s rough first week dominated the headlines, but the bigger story for engineering teams is how much infrastructure OpenAI shipped around it. The Agents API entered public beta, exposing the Codex harness as a managed service. Codex CLI 0.154.0 turned the terminal tool into a parallel working system with git worktrees. A quality postmortem confirmed three bugs degrading Astra output and triggered a full usage reset. And OpenAI paused ChatGPT Pro sign-ups under capacity pressure.

Here’s the week’s breakdown.

1. Astra Quality Postmortem: Three Bugs, One Reset

By September 11, side-by-side comparison videos and developer complaints about Astra “feeling worse” than launch week reached a tipping point. Tibo Sottiaux, Codex/ChatGPT lead, published a postmortem on September 12 confirming three concrete issues:

  1. Legacy skills misfiring — Skills written for previous models triggered too aggressively or prevented Astra from checking its own work.
  2. Context-management experiment bug — An opt-in experiment caused early stops and replies to stale messages. Affected approximately 4,000–5,000 users. Disabled.
  3. Misconfigured serving engines — Badly tuned engines degraded quality for a long tail of traffic. Removed.

OpenAI issued a full usage reset for Codex and Astra users at midnight September 12→13. Unresolved: complaints about xHigh reasoning burning less usage than Medium, and reports of usage caps up to 4× tighter than launch week. ChatGPT Pro new subscriptions were also paused due to capacity constraints.

Takeaway for engineering leads: if your team rolled out Astra in the first 48 hours and formed opinions about quality, re-evaluate now. The postmortem fixes materially changed follow-through consistency and latest-message tracking.

2. Agents API: The Codex Harness as a Managed Service

OpenAI shipped the Agents API in public beta on September 10. This is not a renamed Assistants API — it’s an entirely new service built on the Codex harness infrastructure that powers Codex CLI and ChatGPT internally.

What OpenAI Manages

Sessions, orchestration, context compaction, crash recovery, subagent coordination, lazy tool loading, and MCP server connections. You supply the agent definition (model, instructions, tools), optional sandbox environment, and business logic.

Sandbox Options

  1. OpenAI-hosted sandbox — same infrastructure behind Codex and ChatGPT. Configurable with files, packages, skills, and plugins.
  2. Self-hosted — run codex exec-server in your environment. All connections are outbound WebSocket.
  3. Nine partner sandboxes — Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, Vercel.

The Two Enterprise Blockers

  • US-only data residency during beta. A hard blocker for EU/UK workloads with data-residency requirements.
  • Zero Data Retention not supported — in any configuration, including self-hosted sandboxes. The retention limitation sits on the API layer, not the sandbox. You can move compute into your own VPC and still not get ZDR.

Pricing is straightforward: no separate fee for the Agents API. You pay standard model token rates, hosted tool call rates, and container time for OpenAI-hosted sandboxes.

Assessment: The Agents API is the fastest path from prototype to production for teams already committed to OpenAI’s model stack. The harness/compute separation — where tool execution runs in unprivileged sandboxes while orchestration stays privileged — is architecturally sound. But US-only residency and no ZDR mean enterprises with strict compliance postures should treat this as a US-region beta, not a global production platform.

3. Codex CLI 0.154.0: Parallel Worktrees and Inline Q&A

The headline CLI release of the week shipped September 9 with two features that change how teams work with Codex sessions.

Git Worktrees

--worktree or /worktree creates isolated git checkouts for new or forked sessions. Parallel tasks no longer collide in a single working directory. A session that modifies package.json won’t conflict with another session running tests in the same repo. Worktree sessions can be browsed and resumed independently.

Inline Q&A

Codex can now answer questions — suggested choices or custom text — while it continues working. Previously, a prompt for clarification blocked the entire session. Now the agent keeps executing while you respond on your own timeline.

Other Notable Additions

  • Windows shared background server with daemon lifecycle commands and managed updates.
  • MCP OAuth handling — connections coordinate token refreshes and surface login challenges when refresh fails.
  • Startup trust hardening — workspace-controlled helpers no longer run before trust is established.
  • macOS terminal-injection protection — sandbox blocks terminal input injection attacks.
  • Plugin refresh — existing sessions pick up newly installed plugin tools without restart.

The deprecated codex mcp-server entry point (running Codex as an MCP server over stdio) is no longer available in 0.154.0. Teams using that integration need to migrate.

4. GPT-Live-1 in the API: Voice at $0.05/Minute

OpenAI’s full-duplex speech-to-speech model entered the API on September 10. GPT-Live-1 listens and speaks simultaneously with natural turn-taking and interruptions. The architecture separates the voice layer from backend reasoning — GPT-Live-1 handles speech and delegates deeper work to any backend text model (Astra, GPT-5.6 Luna, or third-party).

Pricing is structurally different from Realtime models: $0.05 per minute for the voice layer, billed per second, plus standard API rates for the backend model. For context, gpt-realtime-2.1 charges $32 per million audio input tokens and $64 per million audio output tokens. At typical speech density, GPT-Live-1 is roughly an order of magnitude cheaper for the voice layer alone.

Concurrency ranges from 25 sessions (Tier 1) to 500 (Tier 5). No free tier.

5. ChatGPT Images 2.5 and Data Agent

Two product releases worth noting for engineering teams building internal tools:

ChatGPT Images 2.5 (September 8) shipped with a Sketch tool for drawing visual references inline, comment-based edits for targeted image modification, and 50% lower latency versus Images 2.0. Two API models — Flare (default, fast) and Sunburst (premium, precise) — are priced at 2× the GPT-Image-2 rate. SynthID watermarking was added across ChatGPT, Codex, and the API.

Data Agent for ChatGPT Work (September 10) connects to Snowflake, BigQuery, Redshift, Databricks, ClickHouse, MongoDB, and Datadog, plus reads existing Power BI, Tableau, and Sigma dashboards. It answers business questions in natural language, investigates why metrics changed, and builds interactive dashboards with approval workflows.

6. Competitive Context: Anthropic’s Revenue Lead Widens

The competitive data this week reinforced a narrative that’s been building since spring:

  • Anthropic Q2 2026 revenue: $11.6B vs OpenAI’s $6.7B
  • Anthropic annualized run rate: $65B vs OpenAI’s $40B+
  • New enterprise procurement: Anthropic 65%, OpenAI 32%
  • Anthropic private valuation: $965B vs OpenAI’s $852B

OpenAI still leads on total user scale (900M+ weekly active users, 50M+ paying subscribers) and holds top benchmark positions. But the procurement momentum has shifted. For teams evaluating platforms, the practical implication is that OpenAI is no longer the default enterprise choice — it’s one of two, and the decision increasingly depends on specific workload characteristics rather than brand.

The Week Ahead

OpenAI DevDay 2026 is scheduled for September 29 in San Francisco. Expect Agents API updates, possibly GA timing for ZDR or international residency, and likely model roadmap details. For teams building on the Agents API today, the US-only and no-ZDR constraints are the key items to watch.

Follow along at https://x.com/kkaminsk for daily Codex updates. The Codex Weekly series publishes every Monday.