This is the week everything landed. After a two-week government vetting delay, GPT-5.6 hit general availability on July 9 — and OpenAI didn’t just release a model, they reshaped the entire product surface around it. A new three-tier model family (Sol, Terra, Luna), the ChatGPT Work super-app, full-duplex voice with GPT-Live, a merged desktop experience, and a Codex CLI that continues its relentless weekly release cadence. The Assistants API shutdown clock is also ticking: 47 days remain until the August 26 hard cutoff.

Here is our weekly breakdown of what matters across OpenAI’s models, the Codex ecosystem, and the broader competitive landscape — and what it means for your technical strategy.

Read the full breakdown on X.

1. GPT-5.6 Family: Sol, Terra, Luna — Three Tiers, One Launch

GPT-5.6 launched on July 9 with three distinct tiers instead of a single flagship — a deliberate pricing strategy that signals a price war with Anthropic and Google.

Sol is the new frontier — 91.9% on Terminal-Bench 2.1 (vendor-reported), beating Claude Opus 4.8’s 78.9% and matching Claude Mythos 5’s 84.3%. Sol also tops the Artificial Analysis Coding Agent Index at 80, edging Claude Fable 5 by 2.8 points with less than half the output tokens. On Agent’s Last Exam (ALE), Sol scores 53.6 vs Fable 5’s 40.5 — a 13.1-point gap. The one outlier: SWE-bench Pro at 64.6%, still behind Mythos 5 (80.3%) and Opus 4.8 (69.2%).

Terra is the balanced tier — GPT-5.5-level performance at half the cost ($2.50/$15 per 1M tokens vs GPT-5.5’s $5/$30). This is the default for Free and Go users in Work and Codex.

Luna is the fast-and-cheap option at $1/$6 per 1M tokens — beating Opus 4.8 on the coding agent index at a fraction of the price.

Reasoning modes: “max” is available to all GPT-5.6 users in Codex; “ultra” deploys multi-agent acceleration for the hardest problems, available to Plus and higher in Codex, and to Pro/Enterprise in ChatGPT Work.

Token efficiency: Sam Altman claims 54% improvement for agentic coding tasks. Prompt caching now has explicit cache breakpoints and a 30-minute minimum cache life. Cache writes billed at 1.25x uncached input; cache reads get a 90% discount.

Government context: GPT-5.6 was held in limited preview since June 26 (~20 partner organizations) due to a voluntary U.S. government national security review. The White House clarified it did not “approve” the release — companies work with the administration voluntarily. Sol is classified as High cybersecurity/bio risk (below Critical threshold).

2. ChatGPT Work — The Super-App Launch

The biggest ChatGPT update in history dropped alongside GPT-5.6. ChatGPT Work merges Codex agent technology into a cross-application task execution environment.

What it does: Gathers context from connected apps (Slack, Teams, Google Drive, SharePoint, email, calendars, CRM, project management tools) via plugins. Creates finished deliverables — documents, spreadsheets, presentations, web apps. Operates continuously for hours on complex projects, breaking them into multi-step plans. Uses @app mentions to direct queries to specific information sources.

Desktop app consolidation: The Codex desktop app is now the ChatGPT desktop app — update in-place, keep projects and settings. Three modes: Chat (conversation), Work (multi-step tasks), Codex (developer/technical). Available globally for macOS and Windows, all plans including Free. The old ChatGPT desktop app is renamed “ChatGPT Classic” — no further updates.

New Codex features in the merged app: Inline diff editing, PR review in side panel, multi-repo project support, faster computer use powered by GPT-5.6.

Sites (public beta): Turn work or ideas into interactive websites or web apps shareable via URL — live dashboards, project trackers, launch calendars, prototypes, internal portals. ChatGPT can update sites as underlying information changes. Available to all paid users.

Availability: Web and mobile rolled to Pro, Enterprise, Edu first — Plus and Business within days. Desktop available to all plans globally.

Other Work features: Built-in browser on desktop, Chrome extension updated for sidebar access, Scheduled Tasks (replaces retired Pulse daily briefing), Plugin Directory replaces old App Directory. Group chats retired as of July 9.

3. Codex CLI v0.143 — Relentless Cadence

The CLI shipped 26 releases in Q2 2026 and is now at v0.143.0 (July 8), written in Rust, Apache-2.0, with 94K+ GitHub stars and 5 million weekly active developers (up from 3M in April).

Model lineup in CLI: GPT-5.6 Sol (new flagship, GA), GPT-5.6 Terra (balanced, default for Free/Go), GPT-5.6 Luna (fast, affordable), GPT-5.5 (still available, strong at 88.0% TB 2.1), GPT-5.4 (general fallback), GPT-5.4-mini (fast subagent work), gpt-5.3-codex-spark (research preview, near-instant, Pro only).

Key capabilities: Two-axis safety system (approval_policy × sandbox_mode) — still the differentiator vs Claude Code’s single permission layer. Goal mode (GA since May 21) for long-horizon objectives running hours to days. MCP support with deferred schema loading since v0.142.2. Plugin marketplace with remote plugins enabled by default in v0.143. /import for selectively importing setup from Claude Code. Cloud handoff for long tasks. codex exec for non-interactive CI/CD automation. Multi-agent worktree system for parallel sub-agents.

Q2 strategic themes:

  1. Realtime & voice — WebRTC v2 default (v0.119), background agent progress streaming
  2. Plugin ecosystem — codex marketplace add (v0.121), /plugins menu (v0.142), remote plugins default (v0.143)
  3. Remote execution — Authenticated, end-to-end encrypted Noise relay channels (v0.141), per-thread MCP servers
  4. Enterprise readiness — AWS Bedrock support (v0.123), monthly usage analytics (v0.137), Windows system proxy (v0.142.1), Bedrock GPT-5.6 models (v0.143)

Plan pricing: Free ($0, GPT-5.6 Terra, limited), Go ($8/mo, more quota), Plus ($20/mo, model picker, ~15-80 messages per 5hr window), Pro ($100/mo, 5x-20x Plus limits, Sol Ultra, codex-spark), Business ($20/user/mo annual), Enterprise (custom). Billing shifted from per-message to token-based credits on April 2. Annual plan prices reduced: Go $4/mo, Plus $16/mo, Pro $160/mo.

4. GPT-Live — Full-Duplex Voice

OpenAI launched GPT-Live on July 8, replacing Advanced Voice Mode with true full-duplex voice models.

Architecture: Simultaneously listens and speaks — no turn-taking. Evaluates interaction cues multiple times per second, dynamically deciding to speak, pause, listen, or process tools. GPT-Live handles natural conversation flow while heavy reasoning and search are delegated to GPT-5.5 in the background — the model keeps chatting while the frontier model works silently, then seamlessly weaves the answer back into the vocal stream.

Models: GPT-Live-1 (paid — Go, Plus, Pro) and GPT-Live-1 mini (free). Rolling out gradually on iOS, Android, and ChatGPT.com. API access coming soon.

Realtime API models also released: gpt-realtime-2.1 ($4/$24 text, $32/$64 audio per 1M tokens) and gpt-realtime-2.1-mini ($0.60/$2.40 text, $10/$20 audio). ≥25% lower p95 latency. Knowledge cutoff: September 30, 2024.

5. Competitive Landscape — The Heat Increases

Terminal-Bench 2.1 public leaderboard (verified June 28): Codex CLI with GPT-5.5 at 83.4% (#1), Claude Code with Fable 5 at 83.1% (#2). GPT-5.6 Sol’s 88.8% would top the list but isn’t yet on the public leaderboard.

Anthropic: $30B annualized revenue run-rate (April 2026) — rapidly closing the gap with OpenAI. Fable 5 was suspended June 12 under U.S. export-control order, restored July 1. Mythos 5 still partially restricted. IPO target October 2026 at ~$965B valuation. Claude Code v2.1.205 shipped background subagents by default, 5-level recursive subagent spawning, Chrome GA.

Google DeepMind: Gemini CLI at 105K stars, Gemini 3.1 Pro at 70.7% TB 2.1. Research leader but adoption lag in enterprise.

opencode: 165K+ stars, MIT, provider-agnostic — lost Claude Pro/Max subscription login after Anthropic dispute.

Open-weight models: Ornith-1.0 (MIT, 9B-397B params, SWE-bench Verified 82.4%), DeepSeek V4, GLM-5.2, Qwen 3.7, MiniMax M3, Kimi K2.6 — all viable for self-hosting.

Cursor found benchmark contamination — models retrieving public eval solutions from internet/git history. A reminder to take vendor benchmarks with a grain of salt.

The $3.8T IPO wave: SpaceX/xAI, OpenAI (September target, ~$1T valuation), and Anthropic (October target, ~$965B) collectively targeting ~$3.8T in combined market cap. Staggered timing to avoid cannibalizing investor demand.

6. Assistants API — 47 Days and Counting

Hard deadline: August 26, 2026. All calls to /v1/assistants, /v1/threads, and /v1/threads/runs will return errors. No grace period, no degraded mode.

Migration target: Responses API — now at full feature parity including deep research, MCP tool connections, and computer use.

Critical warnings: No automated migration tool — OpenAI will not ship one. No Thread export — extract conversation history now via the Assistants API before shutdown. Silent failures during migration — the Responses API returns 200 OK even when context is lost, grounding is empty, or streaming breaks. Test thoroughly. Zapier affected — all Assistants API steps in Zaps fail on August 26. Azure OpenAI — same timeline, not exempt.

Fine-tuning policy change (July 2): Organizations with no fine-tuned model inference in the past 60 days can no longer create fine-tuning jobs. Broader deadline: January 6, 2027 — all existing customers lose the ability to create new fine-tuning jobs. Inference on existing fine-tuned models remains available until base model deprecation.

Recommended internal deadline: mid-July — now.

7. OpenAI Corporate — IPO, Credit, and Government Stake

IPO: September 2026 target, confidential filings submitted, valuation approaching $1 trillion.

Government stake: OpenAI floated giving the U.S. government a 5% equity stake (~$42.6B), modeled on the Alaska Permanent Fund — a national “AI dividend” vehicle. Pooled with similar stakes from other leading AI labs, tied to voluntary engagement with frontier model standards.

Credit: $520M credit line from Bank of America (reported July 8). BofA had previously passed on lending to OpenAI — the reversal signals maturing credit profile.

Microsoft: >$100B in combined investment, Azure infrastructure, and hosting costs through June 2026.

OpenAI Deployment Company: Standalone entity majority-owned by OpenAI, backed by >$4B from TPG-led group, focused on deployment/commercial operations.

Funding dominance: OpenAI + Anthropic together accounted for $217B of global startup funding in H1 2026 — ~43% of all global VC capital ($510B total).

Final Thoughts

This is the most consequential week in OpenAI’s history. GPT-5.6 delivers a three-tier pricing strategy that forces a price war, ChatGPT Work merges Codex into a cross-application super-app, and the Codex CLI continues its relentless weekly cadence with 5 million developers. Meanwhile, the Assistants API shutdown in 47 days means teams still on the old API need to move now. The September IPO at ~$1T valuation will only intensify OpenAI’s enterprise push. Engineering leaders should: migrate off the Assistants API immediately, evaluate GPT-5.6 Terra as the cost-performance sweet spot, and begin treating ChatGPT Work as a legitimate enterprise automation platform — not a consumer chatbot with extra features.