← Back to OpenClaw News A paper-cut editorial illustration of a night harbor where a cargo ship departs, prompt scrolls are folded into compact capsules in a caching vault, and a violet comet arrives overhead
November 20, 2026Release WatchSecurityEnterpriseSkills

OpenClaw 2026.11.6 Ships as Gemini 4 Neon Arrives and Main Trims Prompt Costs With GPT-6.2 Caching Fixes

Yesterday's tags became real packages: 2026.11.6 is now latest on npm with tighter command approvals, and the 2026.11.10 beta is on the beta channel. Gemini 4 Neon reached OpenClaw through its hosted model catalog, main went after prompt costs, and Google's Antigravity terms name OpenClaw in an OAuth warning.

Yesterday's two release tags turned into real packages on Thursday morning. OpenClaw 2026.11.6 is now the latest build on npm and the newest GitHub release, and 2026.11.10-beta.1 sits on the beta channel. The project then announced Gemini 4 Neon support Thursday evening, hours after a fix got its hosted model catalog publishing again. Overnight, main focused on what each agent turn costs: a GPT-6.2 caching fix, a smaller system prompt, and turn costs that no longer read $0 for priced models. Below we also cover a Google terms clause every Gemini user of OpenClaw should read, a ClawHub plugin that hands your agent a cloud desktop, and a weekend hackathon. All times are Pacific.

2026.11.6 Ships, and Approval Prompts May Change

According to the npm registry, [email protected] was published at 8:52 AM Thursday. The OpenClaw v2026.11.6 release page on GitHub followed at 9:26 AM and is marked Latest. It counts 386 pull requests, 48 direct commits and 157 contributors. The beta went out first: npm received 2026.11.10-beta.1 at 7:09 AM, and the 2026.11.10-beta.1 pre-release page on GitHub appeared at 7:45 AM. When we checked at 6:27 AM Friday, npm view openclaw dist-tags showed latest at 2026.11.6, beta at 2026.11.10-beta.1 and extended-stable still at 2026.10.33.

The formatted 2026.11.6 release notes on docs.openclaw.ai, which returned a 404 yesterday, are now live. They call it "a patch release with bug fixes and quality-of-life improvements throughout OpenClaw." We covered its credential items yesterday, so here is the part most likely to surprise people after the upgrade: the Security and Privacy section tightens command approval in several places. Gateway shell commands that need approval are now checked more consistently on macOS and Linux, and the notes warn that "a command that previously ran automatically may now prompt for approval or be denied under your existing policy." A separate fix closes a path where the Gateway let some shell commands run without the approval your policy required. Some commands that change their own environment now require a prompt too. Reapplying an unchanged exec policy, including the cautious preset, no longer quietly switches off automatic command review.

Two more items are worth knowing. On Windows, approvals for inline POSIX shell commands stay one-time even after you choose allow always. In the Browser and Computer Use section, browser snapshots no longer expose filled passwords, or fields marked as one-time codes, card numbers or card security codes, to the model. 2026.11.6 is also well above both affected ranges in the GHSA-pw3x-g8v8-fvj2 delegation advisory we reported yesterday. That advisory's first patched stable was 2026.11.4, so anyone moving to 11.6 from an older build picks up that fix too.

SEN-X Take

Expect a few new approval prompts after you install 2026.11.6, and treat them as the release working. Each one marks a command that used to run without the check your policy asked for. Before you widen an allowlist to make them go away, read which command triggered the prompt. Commands that change their own environment are one of the cases the notes call out, so check scripts and wrappers first.

Gemini 4 Neon Arrives Through the Hosted Catalog

At 8:52 PM Thursday, the project posted OpenClaw's official announcement of Gemini 4 Neon support on X: "Gemini 4 Neon is now in OpenClaw!" The post repeats Google's claim that at Neon's default setting it costs 32% less per task than Carbon. In Google's Introducing Gemini 4 Neon post, dated November 18, the model is priced at $3 per million input tokens, $0.15 per million cached input tokens and $15 per million output tokens, with a 1 million-token context window. Google positions Neon between Carbon, for the hardest long-horizon projects, and Hydrogen, for fast everyday execution. It also describes four thinking levels, with high as the default.

The model reached users without a new OpenClaw release, and that path briefly broke. Pull request #201737, which lets the hosted model catalog publish again, merged at 1:01 AM Thursday. It says catalog.openclaw.ai had not published since 4:43 PM on November 17, "so new models such as google/gemini-4-neon never reach users." The cause was a release check that npm quietly upgraded from 2026.10.33 to 2026.11.5 while installing plugins. The PR notes that the fix changes nothing at runtime for OpenClaw users.

A second fix makes the spending numbers trustworthy. Pull request #201924, which stops OpenClaw recording a placeholder $0 over a real model price, merged at 9:11 AM Thursday. Its author reports that in a test against a real hosted catalog with 2,266 priced keys, where the transport reported an all-zero cost, main recorded a wrong $0 for 2,020 of them, and that 2026.11.5 and 2026.10.33 recorded $0 for the Gemini models tested. In the PR's end-to-end run, one Gemini 4 Neon turn priced at $0.050169. Tiered OpenAI models on the Codex runtime, such as GPT-6.2 Sol, had recorded no cost at all and now get an estimate. For setup, the OpenClaw Google (Gemini) provider documentation still lists google/gemini-4-argon as the default model and recommends a Google AI Studio API key, with Vertex AI for managed Google Cloud installations.

SEN-X Take

Neon is a reasonable model to trial for coding agents, but do not let the Usage page judge it yet. The pricing PR found 2026.11.5 and 2026.10.33 recording $0 for Gemini turns, and its fix merged after the 2026.11.6 tag, so it is only on main. Until it ships, compare Neon against Argon using Google's own billing console, and switch your default only after a week of real invoices.

Main Goes After Prompt Costs

The largest change for OpenAI users is pull request #201094, which caches system prompts on GPT-6.2 Astra and Sol, merged at 2:30 AM Thursday. It found that gpt-6.2-astra never caches the Responses API instructions field, where OpenClaw puts the system prompt, so the whole prompt was billed as plain input on every request. A new model capability, compat.cachesInstructions, moves the prompt into a developer message for models flagged false. The PR says only gpt-6.2-astra carries that flag today. In its six-turn test on a scratch Gateway, plain input per request fell from about 9,500 tokens to between 119 and 211. Two follow-ups landed overnight. Pull request #202477, which fails the live test if GPT-6.2 cache reads go missing, requires at least 1,024 cached tokens on a warm turn. Pull request #202359, which lets agent exec accept the new caching capability, fixes config overrides that the new key had broken.

Then, at 1:02 AM Friday, pull request #202326, which tightens the system prompt and renders a compact skills catalog, merged. By its own measurement, OpenClaw's generated prompt with the full toolset was 19,481 characters, against roughly 1,600 for Pi and 3,500 to 4,000 for Hermes. The PR folds rules that had been restated in two or three sections into one place each. When skills_read is available, it renders each skill as one line instead of a six-line XML block with a file path. In its behavioral A/B test, both models scored 12 of 12 before and after, while input tokens fell 52% on GPT-6 Luna and 49% on Claude Sonnet 5.5. In real agent turns, the median first-request prompt dropped 15%, and the matching skill was loaded in 9 of 9 runs instead of 7 of 9.

SEN-X Take

If you run OpenClaw on GPT-6.2 Astra with an API key, the caching fix could be the biggest cost change you see this month, because every request was paying full price for the system prompt. It merged on main after both Thursday release tags were cut, so neither 2026.11.6 nor the 2026.11.10 beta includes it. Write down your current cached-token ratio now so you can check the improvement when the fix reaches a release.

Security Tip: Keep Gemini on an API Key, Not Antigravity OAuth

  • Read Google's clause. The Google Antigravity Terms of Service say: "Using third party software, tools, or services to access the Service (e.g. using OpenClaw with Antigravity OAuth) is a breach of this Agreement. Such actions may be grounds for suspension or termination of your Antigravity and/or Gemini CLI accounts." That names OpenClaw directly, and the accounts at risk are your Antigravity and Gemini CLI accounts, not your Gateway.
  • Use the supported path. OpenClaw's Google provider docs say "OpenClaw does not offer new Gemini CLI OAuth or Antigravity OAuth setup" and point to an AI Studio API key or Vertex AI instead. If an old config, a third-party plugin or a forum recipe has your Gateway signing in with Antigravity or Gemini CLI credentials, swap it for an API key profile with openclaw onboard --auth-choice gemini-api-key, then revoke the old sign-in.
  • Keep secrets out of browser snapshots. 2026.11.6 stops snapshots from showing filled passwords and marked card or one-time-code fields to the model, but its own notes say to "continue to use a dedicated browser profile for sensitive work." Upgrade, and keep the agent's browser profile separate from the one holding your saved logins.
  • Watch for undelivered approvals. Pull request #202072, which logs approval notices that could not be delivered, merged at 1:56 PM Thursday. Before it, if native approval delivery reached no one and the fallback /approve notice also failed, nothing was recorded. Once you run a build with it, a stalled agent waiting on approval leaves a warning in the Gateway log with the approval ID and channel.

Skill Card: box by Simular

The box by Simular plugin page on ClawHub describes it as "a cloud Linux desktop with Chrome and a shell that it drives over MCP," so the agent works there instead of on your own machine. Simular promoted the listing on X at 3:04 PM Thursday. When we checked, ClawHub showed version v0.1.1, a security audit result of "Pass," and a last update one week ago. Install with openclaw plugins install @simular-ai/openclaw-box; on 2026.10.4 or older, use the clawhub: prefix.

  • What it is. The listing says the plugin declares box's remote MCP server at mcp.box.simular.ai with OAuth sign-in, ships a box skill, and "runs no code of its own." You sign in from Settings, Plugins in the Control UI, and you can watch the desktop live or take over from Simular's site. The source is in the simular-ai/box-plugin repository on GitHub, which is MIT-licensed.
  • The sign-in detail to respect. If you connect through the CLI rather than the plugin page, the box approval page shows "Unverified - this agent registered itself." The listing's advice is to "allow it only when you just started the sign-in yourself." An older CLI setup also adds a box entry to mcp.servers that overrides the plugin, so remove one of them.
  • Our caveats. box is a hosted service from a third party. Anything the agent opens there, including sites it logs into, lives on Simular's machine. We read the listing and the repository; we did not install it or check its pricing.

Also on Main: Automation Cards and Quieter Refusals

Pull request #202126, which shows planned automations and their controls in chat, merged at 6:05 PM Thursday. When an agent or template sets up scheduled jobs, chat now shows one card listing each planned automation, with a switch and settings once the job is saved, and a red "Not created" row when one fails. Installing the three-job Weekly Planner template produced one card instead of three notices. Later, pull request #202252, which records a refused queued request for the owner without replying in the channel, merged at 10:24 PM. It fixes a case where a literal NO_REPLY could be sent to a conversation whose access had just been revoked. The owner now sees "This request was not started because the requester no longer has access to this channel." The PR says only the X plugin reaches that path today.

Community: A Weekend Hackathon on OpenClaw

Cerebras is running a weekend event, and OpenClaw reposted the invitation. The Cerebras x Kimi K3.1 Pro hackathon page on Luma calls it the "Cerebras, Moonshot AI's Kimi K3.1 Pro and OpenClaw Weekend Hackathon" with $5,000 in prize money. It runs on Discord from 10 AM Saturday, November 21, to 10 AM Sunday, November 22. Cerebras says participants get early access to Kimi K3.1 on its hardware.

What We Left Out

One developer said on X that Google banned an account over third-party OAuth use. We could not verify that, so we quoted only Google's terms. Speculation that OpenClaw will offer Google AI subscription sign-in is a rumor, and we did not read anything into a cryptic "December 2nd" post that OpenClaw reposted. We also skipped a report of an embedded-harness model fix that we could not tie to a specific pull request, the MyClaw 3.0 and Agent Barn launches, which are third-party products, and an OpenSSL advisory we have not traced to OpenClaw's dependencies.

What to Watch Next

Watch for a 2026.11.10 beta that carries the GPT-6.2 caching and pricing fixes, Neon's first week of real cost data once turns stop recording $0, and the formatted docs page for the 2026.11.10 beta, which still returned a 404 this morning.

Rolling Out 2026.11.6 or Trialing Gemini 4 Neon?

SEN-X helps teams plan OpenClaw upgrades around stricter approvals, move Gemini access onto supported API keys, and measure real model costs before switching defaults.

Contact SEN-X →