GPT-5.6 Exists — But You Can't Have It Yet. Meanwhile, Etched Wants to Rewire AI Inference Hardware

OpenAI's GPT-5.6 is real, but US government restrictions are keeping it out of most hands for now. In parallel, AI inference chip startup Etched is emerging from stealth with $800M raised, a $1B+ order backlog, and first-try silicon success on TSMC 4nm.

OpenAI's GPT-5.6: Real, Restricted, Rolling Out Slowly

OpenAI has built GPT-5.6, a new model family comprising three variants — Sol, Terra, and Luna — but access is currently limited to "select partners" following intervention from the US government.

Sol is the most capable of the trio, outperforming Mythos on several benchmarks. Notably, its performance on cybersecurity exploit tasks sits just below Mythos-level — a gap that appears deliberate. Sam Altman has indicated broader rollout is coming, though an initial US-only launch seems likely.

Separately, OpenAI published an economics paper examining Codex adoption inside and outside the company. A key finding: non-technical teams are adopting AI coding tools nearly as fast as engineering departments — a signal of how broadly these tools are spreading.

Etched Comes Out of Stealth — and It's a Big Deal for Inference

While model releases dominate headlines, the real competitive battleground in 2026 is shifting to AI inference: lower latency, lower cost, lower power draw, and more tokens per second. The market demand is unlike anything seen before.

Etched, an AI inference hardware startup, is now emerging from stealth — and the numbers are striking:

$800M raised to date
$1B+ in backlog orders
Production underway with TSMC on 4nm
First chip (A0) worked on the very first attempt — an exceptionally rare outcome in silicon development
First product shipped in under three years, versus a 7+ year norm for hardware peers
Team of 400+ drawn from NVIDIA, Google TPU, Broadcom, SK Hynix, TSMC, and other elite chip programs

Etched's approach is extreme vertical integration — chips, racks, software, manufacturing, and production are co-designed end to end.

"Who wins in AI won't just be decided by who has the best models. But by who can serve them."

The company was founded by Gavin, a 21-year-old Harvard dropout, who attracted backing from Peter Thiel, Jane Street, Two Sigma, Jump, HRT, Ribbit, Stripes, and VentureTech Alliance — the latter holding a deep partnership with TSMC.

Other Notable Developments

Cursor for iOS now lets users launch always-on cloud agents from their phones and remotely control agents running on desktop. Composer 2.5 is available at 75% off through July 5.

X (formerly Twitter) launched a hosted MCP server, enabling tools like Grok and Cursor to connect directly to the X API and developer docs without needing to build the server layer themselves.

Replit launched a desktop app for both Mac and Windows.

A US national design studio released Rampart — a 14.7MB in-browser ML model that redacts PII before data is ever sent to a server.

Tools Worth Watching

Zaro — build live apps, agents, and workflows from Slack, email, docs, and calendars
Unpeel — a native Mac terminal for agents with persistent sessions and git worktrees
Inference.net — test GLM 5.2 on mirrored production traffic before switching your live model
smolmachines — spin up hardware-isolated Linux microVMs
Odessia Travel — an opinionated AI trip planner that searches and books flights, stays, and activities