Every figure, version, benchmark and quoted limit this video puts on screen, chased to a primary source. Checked 2026-08-31.
OpenClaw 2.0 is the marketing name; the shipped version is 2026.8.1.
The release is published as v2026.8.1 and the documentation page for it is titled
“v2026.8.1 (AKA OpenClaw 2.0)”.
Source: https://docs.openclaw.ai/releases/2026.8.1
Source: https://github.com/openclaw/openclaw/releases/tag/v2026.8.1
Distribution. The release ships native macOS downloads and npm packages alongside its other install routes. Provider packages for several vendors are installable on demand rather than bundled. Source: https://github.com/openclaw/openclaw/releases/tag/v2026.8.1
The headline features. Search across session transcripts from the Control UI; moving live sessions between the gateway and runners; surfacing sessions that need attention; typed questions presented as live option cards; dashboards; credential handling during setup; recurring approvals; and richer media attached to sessions. Source: https://docs.openclaw.ai/releases/2026.8.1
Storage. 2.0 moves sessions and transcripts into SQLite, and the release notes ask for verified backups before upgrading. Source: https://docs.openclaw.ai/releases/2026.8.1
Background consolidation is on by default. OpenClaw enables model backed background
memory consolidation by default, promoting provenance qualified material into long term
memory. It is disabled with
plugins.entries.memory-core.config.dreaming.enabled: false.
Source: https://docs.openclaw.ai/concepts/dreaming
The Dream Diary. Phase summaries and diary entries are written to DREAMS.md for human
review, including rewrite counts and highlights. The diary is explicitly human only and is
not itself a promotion source.
Source: https://docs.openclaw.ai/concepts/dreaming
Provenance gating. Before the consolidation prompt is built, candidates whose indexed provenance is untrusted or system are removed. The documentation describes this as a structural taint gate rather than a score penalty, and states that promotion and injection are both gated on provenance rather than on content appearing safe. Source: https://docs.openclaw.ai/concepts/memory-architecture
Bounded recall. With Active Memory enabled, OpenClaw retrieves bounded same agent private conversation context by default on personal installs without configured DM isolation. Explicit controls disable recall, and groups and channels remain excluded. Source: https://docs.openclaw.ai/concepts/memory
Minimum. 1 GB RAM, 1 core, 500 MB free disk, and a 64 bit OS. Recommended. 2 GB or more of RAM, and a 16 GB or larger card or USB SSD. Supported boards. Raspberry Pi 4 or 5 with 2 GB or more, with 4 GB recommended. A 64 bit Raspberry Pi OS is required. Source: https://docs.openclaw.ai/install/raspberry-pi
What it measures. A benchmark for agent reliability in evolving, multi session information environments, scored by a Composite Reliability Score that combines task completion rate with a robustness term rewarding sustained success streaks and penalising clustered failures, macro averaged across the scenario set.
Composition. 12 multi turn scenarios, 337 evaluation rounds, 95 multiple choice reasoning checks, 242 execution checks, and 45 dynamic updates that change files or chats mid evaluation to test belief revision.
Leaderboard, top three by Composite Reliability Score.
| Rank | Model | Harness | Score |
|---|---|---|---|
| 1 | GPT 5.5 | OpenClaw | 68.28 |
| 2 | Claude Opus 4.7 | Claude Code | 66.31 |
| 3 | Gemma 4 31B | OpenClaw | 63.80 |
Source: https://github.com/aiming-lab/ClawArena Source: https://arxiv.org/html/2604.04202v2
Measured between the April baseline v2026.4.14 and the May stable point v2026.5.28.
| Measure | Before | After |
|---|---|---|
| Cold agent turn | 9,819 ms | 1,908 ms |
| Warm agent turn | 7,458 ms | 1,870 ms |
| Agent peak RSS | 686.2 MB | 581.0 MB |
Install footprint. The May line spiked at 1,020.6 MB fresh install size in v2026.5.22,
with nested openclaw/node_modules accounting for 911.8 MB of it. By v2026.5.28 the
footprint was 361.7 MiB total, with nested modules down to 259.7 MiB.
Dependency roots. 300 unique package name and version roots in v2026.5.28, down from
645 at the February high of v2026.2.26, a reduction of 53.5 per cent. The step from 371
roots in v2026.5.27 to 300 came from dropping an all platform native canvas package.
Source: https://docs.openclaw.ai/reference/release-performance-sweep
GPT 5.6 Sol window. OpenAI documents a 1,050,000 token provider window and 128,000 maximum output tokens. OpenClaw derives the safe active input budget as 1,050,000 total minus 128,000 maximum output, giving 922,000, and sets the automatic compaction threshold at 700,000 active tokens.
The pricing boundary. OpenAI applies higher long context pricing once a GPT 5.5 or GPT 5.6 request exceeds 272,000 input tokens: the whole qualifying request is billed at 2x input and cache rates and 1.5x output rates. OpenClaw defaults the active runtime budget for those models to 272,000 tokens.
Fast mode. GPT 5.6 Sol Fast mode is a further 2x over Standard, so combined long context Fast traffic is 4x short context Standard input side pricing and 3x short context Standard output pricing.
Expanding it is explicit. A model’s active input budget is overridden per model entry
with models.providers.openai.models[].contextTokens, so an expanded window is an opt in
configuration choice rather than a default.
Source: https://docs.openclaw.ai/providers/openai Source: https://docs.openclaw.ai/reference/token-use
Bundled providers named in the documentation include OpenAI, Anthropic, Google for Gemini, DeepSeek, Groq, NVIDIA, Ollama, LM Studio, vLLM, SGLang, OpenRouter, Qwen, Mistral and Cohere.
Allowlists and remembered selection. agents.defaults.modelPolicy.allow is the explicit
override allowlist. Normal onboarding and reauthentication preserve an existing explicit
primary model, and adding authentication does not change the primary model unless
--set-default is passed.
Source: https://docs.openclaw.ai/concepts/model-providers
OpenProse. The bundled OpenProse plugin and its /prose command are removed. Users are
pointed at openclaw doctor --fix to clean stale configuration and at the upstream Agent
Skill migration documentation. Existing .prose source files are retained.
OpenAI route naming. Model references using codex/* and openai-codex/* migrate to
openai/*. openclaw doctor --fix migrates provider configuration, stored sessions and
automation routes, flagging conflicts for review.
Plugin SDK. Several plugin SDK imports move to focused openclaw/plugin-sdk/ subpaths,
effective 2026-09-01, covering configuration mutation, channel inbound and outbound
operations, and runtime helpers.
Source: https://github.com/openclaw/openclaw/releases/tag/v2026.8.1
Meta as a listed provider. The model providers page does not name Meta as a bundled first party provider. Meta models are reachable, but through OpenRouter proxying rather than a dedicated provider entry. The narration lists Meta among the providers the docs name, and that is the one item in that list which is not a first party entry.
“Many more” providers. The bundled provider table is the verified set. The size of the full surface beyond it, including community provider plugins, was not enumerated.
The 200 tokens per second figure used when describing what a gateway does not need to do is a round illustrative number rather than a documented OpenClaw threshold, and is not put on screen as a measurement of anything.