[Oct 6 Recap] Four big announcements from OpenAI: 6GW partnership with AMD, Apps in ChatGPT, AgentKit, and Codex GA
TL;DR (grasp it in 1 minute)
- Strategic partnership with AMD: Rolling out 6 gigawatts of AMD GPU capacity in phases. The first 1GW goes live in H2 2026 on MI450. This adds a second pillar alongside the NVIDIA-centric setup.
- “Apps” arrive in ChatGPT: Launched Apps in ChatGPT that you can invoke and operate inside the conversation. Apps SDK (preview) is also available.
- AgentKit announced: A complete toolkit for agent development covering design (Agent Builder) / embeddable UI (ChatKit) / evaluation (enhanced Evals) / RFT fine-tuning.
- Codex is now generally available (GA): Adds Slack integration / SDK / admin features, built on GPT-5-Codex and designed for real engineering operations.
Who should read this (audiences & impact)
This article organizes the Oct 6 announcements from a “go/no-go for adoption” perspective for practitioners in Engineering / IT & DX / Product & Biz Planning / E-commerce & Customer Support / PR & Education.
- Engineering: AgentKit and Codex GA can compress the effort and risk across agent implementation → evaluation → operations.
- Product / Biz: With Apps in ChatGPT, you can consolidate user touchpoints into “conversation × UI” and accelerate external integrations using the SDK.
- IT & DX: The AMD partnership advances supply diversification of compute, opening room for risk hedging and cost negotiations.
- E-commerce & CS: Combine in-chat apps and agents to shorten the service flow from search → propose → execute.
- PR & Education: Understand safety designs (guardrails, policies) and translate them into internal rules for provenance displays and consent.
1. AMD × OpenAI: Establishing a “second pillar” with a 6GW compute base
OpenAI announced a multi-year, multi-generation strategic partnership with AMD, introducing 6GW of GPU capacity in stages. Phase 1 delivers 1GW on MI450 in H2 2026. It’s significant that AMD joins as a core strategic partner, complementing the NVIDIA-centric setup. External coverage also notes the scale and market reaction.
Operational takeaways
- Compute diversification can help smooth supply shortages and strengthen pricing leverage.
- Anticipate cross-generation optimization (mixed MI450/NVIDIA) and pre-design job routing policies for training/inference to ease the transition.
2. Apps in ChatGPT: “Call an app” inside the conversation
Apps now live within ChatGPT, letting you operate services in context—such as Booking.com, Canva, Coursera, Figma, Expedia, Spotify, Zillow—right in the chat. For developers, the Apps SDK (preview) is available and extends MCP (Model Context Protocol) as an open standard. Rollout starts with logged-in users outside the EU, expanding over time.
Use cases (examples)
- “Spotify: make a playlist for Friday’s office social” → edit via in-chat UI.
- “Show Zillow listings matching this budget on a map” → browse and filter within the chat.
- Turn Confluence or internal search into a company app, embedded in ChatGPT via the Apps SDK.
3. AgentKit: One package for design, embedding, evaluation, and RFT
AgentKit unifies the previously fragmented tasks of agent development into a single toolkit. Key components:
- Agent Builder: No-code visual canvas for workflow design and versioning.
- ChatKit: Simplifies embedding conversational UI (threading, streaming, presentation).
- Enhanced Evals: Datasets, trace scoring, automatic prompt optimization, cross-model evaluation.
- RFT (reinforcement-style fine-tuning): o4-mini GA, GPT-5 in private beta. Supports tool-use optimization and custom eval axes.
Adoption pointers
- Enable guardrails via Python/JS libraries to standardize PII masking and jailbreak detection.
- Use Connector Registry to govern integrations like Dropbox/Google Drive/SharePoint with admin controls and audit logs baked into operations.
4. Codex GA: “Side-by-side coding” built for real-world ops
With Codex now GA, you get:
- Slack integration: Ask @Codex in a channel → it works in the cloud → returns a PR link.
- Codex SDK: Embed GPT-5-Codex into your own workflows; includes a GitHub Action for CI/CD.
- Admin features: Environment management, usage controls, analytics dashboard for standardized governance.
Impact on the ground
- Automating routine changes, code cleanup, PR reviews shortens cycles.
- With local / cloud / IDE unified under one account, the boundary between human work and agent work is clearer.
5. Quick-start implementation plans (fastest route by role)
- App owners: Use the Apps SDK to build one internal mini-app, turning enterprise search or approval workflows into in-chat UI. Treat EU rollout as phased.
- Support operations: With AgentKit’s ChatKit + Evals, design the flow from FAQ → handoff to human. Use Guardrails to block unsafe replies and PII leaks.
- Engineering orgs: Wire Codex GA into Slack → PR; start with small repos and expand. Use the audit dashboard to make quality and governance visible.
- Cloud procurement: Reflect the AMD deal by updating allocation policies and spend caps across job type (train/infer/vision/audio) × GPU generation.
Bottom line: Oct 6 marks a declaration of “platform-ization”
- Compute advances diversification via AMD;
- App distribution opens a new surface—conversation × UI—with Apps in ChatGPT;
- Agent operations are unified through AgentKit for design / embedding / evaluation;
- Engineering teams normalize human–AI collaboration with Codex GA.
Together, these moves form the core of “making ChatGPT OS-like”, a milestone where development, operations, and distribution become more tightly connected.
References (primary sources & major media)
- OpenAI official news index (includes Oct 6 items): OpenAI News
- AMD × OpenAI:
- AMD and OpenAI announce strategic partnership to deploy 6 gigawatts of AMD GPUs (official)
- AMD Press Release (official)
- The Verge coverage / Bloomberg (market reaction)
- Apps in ChatGPT:
- AgentKit:
- Codex GA:
- Codex is now generally available (official)