Is ChatGPT Down?

ATLAS SUNSET DATE CONFIRMED (announced 2026-07-09; stops working **August 9, 2026** per press): OpenAI is killing the Atlas browser less than a year after its October 2025 launch -- TechCrunch's framing: 'the browser is a feature, not the destination.' Atlas's capabilities redistribute into (a) a ChatGPT **Chrome extension** (page context, summarization, page Q&A) and (b) the unified **ChatGPT desktop app** (built-in browsing + a remote cloud browser for agent tasks), which together span 'a continuous workspace' across Chrome, desktop, and agents. Part of the broader 'no side quests' consolidation under Fidji Simo (same push that killed the Sora app). If you use Atlas, migrate before Aug 9. GPT-Live note (unchanged as of 7/18): the consumer voice models shipped 7/8 but there is still **no GPT-Live API** -- the developers.openai.com changelog shows no GPT-Live entry (the Jul 6 GPT-Realtime-2.1 is the older, separate realtime line -- don't conflate them); API access remains a notify-me form

GPT-5.6 WEEK-ONE GOTCHAS (2026-07-10, press/analyst-reported -- single-source, verify before relying): the 24h global rollout completed on schedule with no major outage, but early friction is real: (1) **ChatGPT Work and Codex share ONE usage pool** -- long Work sessions silently drain Codex weekly limits; (2) **Codex clients older than 0.144.0 (desktop 26.707.30751+) hide the GPT-5.6 models entirely** -- the top 'where's my model' complaint; update the client first; (3) Terra/Luna are not selectable in the standard chat UI (Work/Codex/API only); (4) cloud/desktop Work conversations didn't sync at launch; (5) UAE inference-residency workloads unsupported at launch. OpenAI's quick fix: enterprise admins got per-user monthly limits + workspace overage caps. Pricing unchanged from launch

GPT-5.6 GENERAL AVAILABILITY (2026-07-09): The **GPT-5.6 family went GA across ChatGPT, Codex, and the API** (global rollout over ~24h), ending the 6/26 limited preview. Three durable tiers: **Sol** (flagship, $5/$30 per 1M in/out), **Terra** (balanced, $2.50/$15 -- previously unpublished), **Luna** (fast/cheap, $1/$6). In ChatGPT chat, Sol is available to Plus and up at medium+ effort; Sol Pro is Pro/Enterprise-only. New capability settings beyond xhigh: **`max`** (more reasoning time, all GPT-5.6 users in Work/Codex) and **`ultra`** (coordinates 4 parallel agents by default; Pro/Enterprise in ChatGPT Work, Plus+ in Codex; multi-agent beta in the Responses API). Vendor headline benchmarks: **Agents' Last Exam 52.7% Sol** (new SOTA, +13.1 over Claude Fable 5 adaptive per OpenAI's charting), **AA Coding Agent Index 80 (max)**, **BrowseComp 90.4% / 92.2% ultra (SOTA)**, **OSWorld 2.0 62.6% (SOTA)**, Terminal-Bench 2.1 88.8%, SWE-Bench Pro 64.6%, GPQA 94.6% -- note AA Intelligence Index 58.9 still trails Fable 5's 59.9, and SWE-Bench Pro trails Mythos 5/Fable 5; OpenAI's pitch is performance-per-dollar (Sol ~half Fable's estimated cost). Also new: **Programmatic Tool Calling** in the Responses API (ZDR-compatible), predictable prompt caching (explicit breakpoints, 30-min minimum cache life, cache writes now billed 1.25x uncached input), and a much stricter cyber-safeguard stack (~10x more blocking than prior models; defensive-cyber depth gated behind Daybreak Trusted Access -- **individuals must enable hardware-passkey Advanced Account Security by 2026-09-01 to retain top cyber-capable model access**). All first-party numbers pending third-party verification

PRODUCT (2026-07-09): **ChatGPT Work launched** -- a long-running agent built on Codex technology and powered by GPT-5.6 that works across your connected apps (unified plugins directory: Drive, SharePoint, Slack, Teams, Gmail, Outlook, Salesforce, GitHub, Canva...) for hours and returns finished sheets/slides/docs/web apps, plus Scheduled Tasks and Sites (public beta) -- full review at /tools/chatgpt-work. Platform consequences for ChatGPT itself: a **new unified desktop app** (Mac/Windows) puts Chat + Work + Codex on EVERY plan including Free; the **Codex app merges into it**; the old desktop app is renamed **ChatGPT Classic**; and OpenAI is **sunsetting the standalone Atlas browser** in favor of the desktop app's built-in browser + an updated Chrome sidebar extension. Rollout: Pro/Enterprise/Edu on web/mobile 7/9; Plus/Business over the following days

VOICE OVERHAUL (2026-07-08): **GPT-Live launched** -- full-duplex voice models (listen while speaking, natural backchannels, live translation) replacing the default ChatGPT Voice experience globally. GPT-Live-1 is the new default for Go/Plus/Pro, GPT-Live-1 mini for Free; hard questions delegate to GPT-5.5 in the background (Instant/Medium/High effort). Massive deltas vs Advanced Voice Mode (GPQA 84.2% vs 45.3% at high effort; BrowseComp 75.2% vs 0.7%). Launch limits: no API yet, no video/screen-share (legacy voice modes remain for those), predefined voices only. Full review at /tools/gpt-live

FINE-TUNING STAGE 2 IN EFFECT (2026-07-02): per the staged sunset announced 5/7, organizations with **no fine-tuned-model inference in the past 60 days can no longer create new fine-tuning jobs** as of July 2. Stage 3 (no new FT jobs for anyone) lands 2027-01-06