Skip to content
NLEN

Radar, November 18, 2026: ten signals from X on GPT-6.2, Perplexity and Claude Code

All daily signals · Markdown version

In short: The key signal of November 18: OpenAI released GPT-6.2 Astra and Sol, with a cheaper Astra and discounts for background work. Alongside it, ten short signals on Perplexity's Orchestrator API, Code Mode and version 2.1.327 of Claude Code, Codex 0.170, agents in companies and rumours, each with its source on X.

Illustration: a radar screen on a monitor on a desk, with two glowing dots, three robots under a conductor's baton, a server with a honeycomb between robots and a terminal window.

Illustration: AI-generated (OpenAI image model via Codex): a radar screen with ten signals.

1.OpenAI releases GPT-6.2 Astra and Sol; Astra 20 percent cheaper

Confidence: high for models and windows (changelog read); prices via The Decoder

The November 17 API changelog lists gpt-6.2-astra and gpt-6.2-sol; tool calling requires the Responses API. According to The Decoder, Astra costs 8 and 40 dollars per million tokens and Sol 2 and 10 dollars. Also new: a two- or eight-hour background window at 30 or 40 percent off. Independent benchmarks put Astra second, behind Claude Fable 5.2.

Source on X: @OpenAI · OpenAI, API-changelog

Context in the briefing: GPT-6.2 in the November 18 briefing

Read also: GPT-6.2 Astra and Sol · AWS sees AI agents hunting API keys

2.Perplexity opens an Orchestrator API with 45 models behind one key

Confidence: high (documentation read)

The documentation lists three formats (Chat Completions, Responses and Anthropic Messages), per-token billing with no per-request fee, and a retry on another deployment when a call fails. The model list counts 45 models from seven providers. Your own OpenAI or Anthropic keys are not accepted, and responses carry no cost field.

Source on X: @perplexity_ai · Perplexity, Orchestrator FAQ

Context in the briefing: Perplexity in the November 18 briefing

Read also: Perplexity Orchestrator API

3.Claude Code gets an experimental Code Mode and linked billing in beta

Confidence: medium (official account, documentation not read)

Claude's developer account announces Code Mode: Claude writes scripts that chain tools instead of calling each tool separately, enabled with an environment variable. Linked billing lets users of Linear, Warp and others pay through their own Claude Platform organisation without sharing an API key. We did not read the documentation page.

Source on X: @ClaudeDevs · @ClaudeDevs

Context in the briefing: Code Mode in the November 18 briefing

Read also: Claude Code 2.1.326

4.Claude Code 2.1.327: mostly fixes, and tokens off the command line

Confidence: high (release notes read)

The November 17 release (23:17 UTC) adds claude mcp add-json reading from stdin, so a token stays off the command line, and makes /status list broken skills, hooks and permissions. Fixed: worktree mode wiped git-ignored files such as .env, and mcp add/remove deleted other keys from .mcp.json. The '173 changes' from a post on X are not in the release.

Source on X: @ClaudeCodeLog · GitHub, release v2.1.327

Read also: Claude Code 2.1.326 · AWS sees AI agents hunting API keys

5.Grok Bots in accounting and Microsoft Autopilot on OpenClaw

Confidence: low to medium (companies' own accounts, no primary source read)

The @bot account says five Grok Bots have a permanent role in SpaceXAI's accounting team. The OpenClaw project says Microsoft Autopilot is built on OpenClaw and in preview for Frontier programme customers, and that Sign in with ChatGPT is out of beta. We did not read a blog post or Microsoft documentation.

Source on X: @bot · @elonmusk

Context in the briefing: Enterprise agents in the November 18 briefing

Read also: OpenClaw 2026.11.4 · Agentic AI

6.Codex rust-v0.170.0 is stable: WebRTC for voice in the TUI

Confidence: high (release notes read)

The November 18 release (03:40 UTC) adds WebRTC for realtime voice calls in the terminal UI, auto-refreshed authentication headers for OpenTelemetry and experimental support for identically named MCP servers across environments. It also fixes MCP OAuth races and cloud conversation reconnects. Posts on X first reported only maintenance release 0.169.1.

Source on X: @CodexReleases · GitHub, release rust-v0.170.0

Read also: GPT-6.2 Astra and Sol · Agentic AI

7.OpenCode temporarily offers 200 dollars of GLM 5.3 usage for 10 dollars

Confidence: low (X only, promotion)

OpenCode's account announces a second phase of a promotion: 200 dollars of GLM 5.3 usage for 10 dollars, later raised to 500 dollars until the end of the week. The posts drew over a million views combined. We have not read the terms or duration outside X; check them before paying.

Source on X: @opencode · @opencode

Read also: Perplexity Orchestrator API

8.Hermes Agent: a plugin to train your own model, canary knows GPT-6.2

Confidence: medium (canary via our feed; plugin X only)

A post says Adaption is now a verified partner plugin in Hermes: gather data, train and evaluate a model from the same agent. According to our feed, the November 18 canary adds GPT-6.2 Astra and Sol to the model catalogue; the stable version stays 0.23.2. Anyone trying the canary should back up the session database first.

Source on X: @IBuzovskyi · @adaption_ai

Read also: Hermes Agent 0.23.2

9.SGLang v1.0 and consortium reported; Mistral Vibe free for three months in Europe

Confidence: low (SGLang tag contradicts the posts; Mistral X only)

Several accounts report SGLang v1.0 and a consortium including NVIDIA, AMD, Google and Meta. The latest GitHub release was still v0.5.23 from November 2 when we checked, so v1.0 is unconfirmed. Mistral says Mistral Vibe and the Le Chonk model are free for three months in Europe; we did not read the terms.

Source on X: @DataChaz · @MistralAI

Context in the briefing: Open-source inference in the November 18 briefing

Read also: Mistral Large 4 Le Chonk

10.Unconfirmed: Grok 4.8, a Hermes harness in Telegram and a roadshow slide

Confidence: low (X only)

One account reconstructs confusion over Grok 4.8; the rollout claim comes from the @grok reply bot and is not official. Another account claims an agent harness in the Telegram macOS beta runs on Hermes Agent. A slide said to come from an Anthropic roadshow is second-hand. Do not repeat such claims without a primary source.

Source on X: @Marko_Poly · @SenditHighor

Context in the briefing: Unconfirmed in the November 18 briefing