OpenClaw 2.0 ships: one open client for your subs and local models
Reuse your existing Codex/ChatGPT/Claude logins, verify local models before they run, and boot the UI in 575ms — shared sessions aren't a security wall.
Copy markdown
No API tokens: your subscription is the key
OpenClaw signs in with your existing ChatGPT, Claude Max, or Codex CLI credentials and drives those models from one client — so agent sessions run against plans you already pay for instead of metered API spend.
Local models have to prove themselves first
Guided setup discovers your Ollama and LM Studio models and won't save one until it passes a live answer check; llama.cpp gets a 64K default context and Gemma 4 is the RAM-constrained default. No more silently-misconfigured local endpoints.
The control UI opens roughly 3x faster
A rebuilt web UI cut JavaScript requests from 140 to 45 and cold start from ~1.6s to 575ms in a 50ms-latency test, and new sessions now get a real setup screen for agent, model, reasoning level, and target computer.
Shipped by a swarm: 933 contributors, 16k+ PRs
This release folds in 16,000+ merged PRs from 933 contributors — roughly half the project's entire history — a reminder of how fast a community client can outrun a single vendor's roadmap.
Team sessions go live — but aren't a security wall
Shared cloud gateways let a teammate join or take over an active session in real time, with one trust boundary per gateway; the docs are blunt that these controls are 'not a security boundary,' so sandbox any gateway you expose.
Elsewhere: Sonnet 5's Sept 1 price hike is off
Tomorrow's scheduled jump to $3/$15 per million tokens won't happen — Anthropic's docs now list the $2/$10 introductory rate as permanent. Note the newer tokenizer still counts ~1.0-1.35x more tokens on code.