The post-GPT-5.6 dust is settling, and this week the story shifted from destructive autonomy to platform consolidation. OpenAI shipped Codex CLI v0.145.0 — the biggest feature release since the ChatGPT desktop merger — with stabilized multi-agent collaboration, a cross-tool migration importer, and audio input. Fifteen legacy model snapshots including every Codex-branded model hit their scheduled shutdown on July 23. ChatGPT Work gained a small business program. And Google’s Gemini 3.5 Pro missed its third launch deadline, shipping Gemini 3.6 Flash as a stopgap instead.

Here is our weekly breakdown of what matters across OpenAI’s models, the Codex ecosystem, and the broader competitive landscape.

1. Codex CLI v0.145.0 — Multi-Agent V2, Migration, and Audio

The July 21 release is the most feature-rich Codex CLI update since v0.144.0 merged Codex into the ChatGPT desktop app. It ships in three areas simultaneously: agent collaboration, tool migration, and input modalities.

Multi-agent V2 goes stable. The opt-in multi-agent experience graduates from preview to stable. A root session can now spawn configurable sub-agents — each with its own model, reasoning level, concurrency limit, and sandbox profile — using spawn_agent, send_input, resume_agent, wait_agent, and close_agent. Roles and navigation have been restored from the earlier preview, and the implementation is classified stable but remains disabled by default. For engineering teams, this means a single Codex session can now parallelize work across different model tiers — Sol for planning, Luna for mechanical edits — without external orchestration.

/import expanded to Cursor and Claude Code. Codex can now detect and migrate settings, MCP servers, plugins, sessions, commands, and project-scoped memories from both Cursor and Claude Code. Plugin commands convert to Codex skills during import. The migration is one-way — changes in Codex don’t write back — and Claude Code’s chat conversations are not included, only agent configuration and recent sessions. The importer runs locally in the TUI, rejects remote sessions, and caps session import at 50 sessions changed within 30 days. This is a meaningful switching-cost reduction for teams evaluating a move to Codex.

Experimental paginated thread history. Thread history now supports efficient resume, search, persisted names, sub-agent history, and memory integration. This addresses a long-standing pain point: long conversations that became expensive to reload and impossible to search. The feature is labeled experimental.

Audio inputs and streaming realtime V3. Local audio formats can now be fed as inputs, and tool outputs can include audio. Streaming realtime V3 conversations are introduced but remain under development in the tagged registry.

Amazon Bedrock login (experimental). Custom endpoint and authentication support lands with GPT-5.6 Sol as the default Bedrock model. This gives AWS-native teams a first-class path into Codex without routing through OpenAI’s API directly.

Reliability improvements. Editing an earlier prompt or retrying a safety-buffered turn now creates a proper contextual branch instead of mutating the original conversation. Windows sandboxing gets native exec-server support and correctly quoted hook commands. MCP startup deadlines cover client creation, OAuth discovery no longer blocks startup, and refreshes are serialized.

2. Fifteen Model Retirements — Including Every Codex Snapshot

On July 23, OpenAI shut down 15 model snapshots in a single batch. The most consequential: every Codex-branded API model is now retired.

Codex models retired: gpt-5-codex, gpt-5.1-codex, gpt-5.1-codex-max, gpt-5.1-codex-mini, and gpt-5.2-codex all map to gpt-5.5 (or gpt-5.4-mini for the mini variant). If your production code calls any of these endpoints, requests now fail outright.

Chat models retired: gpt-5-chat-latest and gpt-5.1-chat-latest both fold into gpt-5.5.

Computer use and search: computer-use-preview-2025-03-11 moves to gpt-5.4-mini. Search preview models (gpt-4o-search-preview, gpt-4o-mini-search-preview) also map to gpt-5.4-mini.

Deep research: o3-deep-research and o4-mini-deep-research both fold into gpt-5.5-pro at $30 input / $180 output per million tokens — a real cost increase for automated research pipelines.

Audio: gpt-4o-mini-tts-2025-03-20 moves to gpt-4o-mini-tts-2025-12-15, and gpt-audio-mini-2025-10-06 maps to gpt-audio-1.5.

Action items for engineering teams: Audit your codebase for any hardcoded model IDs in the retired list. Update API calls to the recommended replacements. Pay particular attention to deep research pipelines — the cost jump from o3-deep-research to gpt-5.5-pro is roughly 6× on output. The next retirement batch hits October 23, covering GPT-3.5, GPT-4, GPT-4 Turbo, GPT-4.1 Nano, GPT-4o, and several o-series models.

3. ChatGPT Work — Small Business Program and Enterprise Maturation

ChatGPT for Small Business (July 21). OpenAI launched a comprehensive SMB program built around ChatGPT Work. It includes hands-on virtual training webinars for accounting, marketing, and ecommerce workflows; in-person AI academies across the US through the OpenAI Academy team; interactive guides uploadable directly into ChatGPT Work; and partner integrations with Dropbox, Shopify, Intuit, Slack, Atlassian, and Wix. Five free live sessions are scheduled from July 30 through August 27. ChatGPT Work is powered by GPT-5.6 and available across all subscription plans — not locked behind an enterprise tier. The platform has reached approximately 10 million combined ChatGPT Work and Codex users.

Enterprise governance updates. ChatGPT Enterprise and Edu workspaces gained credit usage analytics and updated spend controls in the Global Admin Console. Workspace agents moved to token-based pricing on July 6 — credit consumption is now based on input tokens, cached input tokens, and output tokens rather than fixed costs. ChatGPT for PowerPoint reached general availability for Business workspaces, with token-based pricing taking effect after August 6 for Enterprise customers.

Service account API key management. The Go SDK v3.44.0 (July 18) added programmatic management of service account API keys via the /organization/projects/{project_id}/service_accounts/{service_account_id}/api_keys endpoint. Admin keys — distinct from standard API keys and scoped to organization management — are now available for workspace owners and admins to connect tools to ChatGPT and Codex administration APIs, with role-based permissions (All, Read only, Restricted).

Desktop chat/work split. The ChatGPT desktop app added a global switcher between Chat and Work modes on July 16. Work conversations sync across web, mobile, and desktop. Custom instructions limit increased from 1,500 to 5,000 characters for Plus, Enterprise, Business, and Education users.

4. Assistants API — 33 Days and Counting

The August 26 shutdown is 33 days away. No new migration tools have shipped. The recommended internal deadline of mid-July has passed.

Critical reminders: There is no automated migration tool and none will ship. There is no Thread export — extract conversation history now. Silent failures during migration are common: 200 OK responses with lost context, empty grounding, or broken streaming. The Responses API is at full feature parity including deep research, MCP tool connections, and computer use. Azure OpenAI follows the same timeline. Zapier has deprecated all ChatGPT steps that use the Assistants API — those workflows stop working on August 26.

For teams still on Assistants API: Treat this as a P0 migration. Inventory every beta.assistants, beta.threads, and beta.threads.runs call. Map threads to Conversations, runs to Responses. Build regression tests for streaming, background work, retries, and cancellation. Run a shadow cutover with side effects suppressed before flipping production traffic.

5. Competitive Landscape — Google Stalls, xAI Opens Source

Google — Gemini 3.5 Pro Misses Third Deadline

Gemini 3.5 Pro did not ship. The model that Sundar Pichai publicly committed to delivering in June at Google I/O has now missed three deadlines: the original June window, a July 17 target, and a mid-July revised estimate. Google’s official model catalog still lists no Gemini 3.5 Pro — the Pro tier remains Gemini 3.1 Pro (Preview) and Gemini 2.5 Pro (Stable).

Bloomberg reports that internal coding benchmarks fell short of targets and key researchers have departed. Google’s only public statement: “Gemini 3.5 Pro is currently testing with partners and we plan to make it broadly available as soon as it’s ready.” No date, no benchmarks, no pricing.

What shipped instead: Gemini 3.6 Flash and Gemini 3.5 Flash-Lite went GA on July 21. Gemini 3.6 Flash is priced at $1.50/$7.50 per million tokens with a March 2026 knowledge cutoff and is available across Google AI Studio, Android Studio, Google Antigravity, GitHub Copilot, and the Gemini app. A specialized Gemini 3.5 Flash Cyber model for vulnerability detection is in restricted pilot.

xAI — Grok Build Open-Sourced After Data Controversy

xAI open-sourced Grok Build on July 15 under Apache 2.0 — a 1.3 million line Rust codebase covering the agent loop, tool implementations, terminal UI, and extension system. The release was precipitated by a data exfiltration controversy: users discovered the grok CLI was uploading entire directories including SSH keys and password manager databases to xAI’s Google Cloud buckets by default. xAI disabled the upload behavior, deleted retained data, and published the full source to restore trust.

What it means for developers: Grok Build is the only major coding agent harness that is fully open source with a public wire protocol (Agent Client Protocol). It supports MCP servers, skills, plugins, and hooks — reusing existing Claude Code MCP configs verbatim. However, external contributions are not accepted; it’s a source-transparency release, not a community project. Grok 4.5, the default model, is priced at $2/$6 per million tokens — aggressively cheaper than Opus 4.8 or GPT-5.6 Sol.

Anthropic — Claude Code Accelerates

Claude Code shipped v2.1.217 on July 21, keeping pace with Codex’s release cadence. Recent highlights include the /fork background-session clone, session-wide caps on web searches and subagents (200 each), MCP auto-backgrounding for long-running calls, and a permission-analyzer overhaul with fail-closed behavior. Claude Sonnet 5 remains the default model with a 1M-token context window at $2/$10 per million tokens through August 31. Fable 5 access has shifted to usage-based billing for Pro users — a signal of the broader industry move toward consumption-based pricing for top-tier models.

Open-Weight Ecosystem

DeepSeek V4 continues to offer cache-hit input pricing as low as $0.07 per million tokens. Kimi K3 from Moonshot AI remains a 2+ trillion-parameter open model. The gap between frontier proprietary models and open-weight alternatives has narrowed to single benchmark digits on several coding evaluations. For self-hosting scenarios, these models merit evaluation.

Final Thoughts

This week marks a transition from crisis management to platform building. Codex CLI v0.145.0’s multi-agent stabilization and cross-tool migration importer signal that OpenAI is focused on reducing friction for teams consolidating on its platform. The model retirements are a necessary cleanup — but teams running any of the 15 deprecated snapshots needed to have migrated yesterday.

Three actions for engineering leaders this week: First, audit your API calls against the retired model list and update to recommended replacements — especially if you’re running any Codex-branded model or o3-deep-research. Second, if you haven’t started the Assistants API migration, start now — 33 days is not enough for a production cutover with proper testing. Third, evaluate Codex CLI v0.145.0’s /import to assess switching costs between Codex, Claude Code, and Cursor — the multi-agent V2 feature and Bedrock support may change the calculus for your team’s architecture.

The competitive landscape is producing genuine alternatives. Grok Build’s open-source harness, Claude Sonnet 5’s value pricing, and the open-weight frontier models from DeepSeek and Moonshot all deserve evaluation. Google’s continued struggles with Gemini 3.5 Pro, meanwhile, are a reminder that not every lab can sustain this pace.

Follow the conversation on X.