AI news, November 18, 2026: GPT-6.2, Claude Code Code Mode, Perplexity Orchestrator API, agents in business
Short answer: The main stories of November 17 to 18: OpenAI released GPT-6.2 Astra and Sol with lower prices and mixed independent measurements, Claude Code got an experimental Code Mode, Perplexity launched an Orchestrator API for dozens of models and agents are appearing in accounting and office software. Rumors about Grok 4.8 and paste-prompts are unconfirmed.

Illustration by Claude (AI), drawn with code: two models with a price scale (GPT-6.2), a terminal with code (Code Mode) and a router to many models (Orchestrator API).
OpenAI releases GPT-6.2 Astra and Sol: cheaper, with mixed reactionsConfidence: high for availability, prices and the independent measurements: a news site that cites the testing institutes was read; medium for OpenAI's benchmark claims (the maker's figures); the criticism on X is opinion.
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17 (6.8 million views). They are in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu and in the API as gpt-6.2-astra and gpt-6.2-sol. According to The Decoder's reporting, Astra costs 8 dollars per million input tokens and 40 dollars per million output tokens, 20 percent less than its predecessor, and Sol costs 2 and 10 dollars, the same as GPT-6.1 Sol. Ultrafast is 50 percent cheaper for three months (OpenAI developer account, 15 thousand views). OpenAI says Astra scores 5.5 points above Claude Fable 5.2 on Agents' Last Exam at about 5 percent of the cost; those are OpenAI's own figures.
Independent measurements are more mixed. Artificial Analysis ranks Astra second with 62 points behind Claude Fable 5.2 (65), reports that coding has plateaued and cites hallucination rates of 59 percent for Astra and 55 percent for Sol on a very hard question set. Epoch AI sees Astra level with Fable 5.2 on its index but weaker on real research tasks. On X, @kimmonismus calls the release "a catastrophe" (229 thousand views) and @Teknium points to contradictory Blender demos (114 thousand views). Perplexity reports 0.783 for Astra (about 12.54 dollars per task) and 0.746 for Sol (about 2.65 dollars) on its own WANDR test.
What this means: the price cut is factual and comes from the price list; the performance gain is mostly a claim by the maker that independent measurements partly confirm. OpenAI says both models reach the Critical threshold for cybersecurity and adds extra access rules. Anyone running agent work can compare Sol with the current choice in their own cost tracking before switching. An article on the previous generation is on our site: our article on GPT-6.1 Terra.
Sources: The Decoder over de lancering van GPT-6.2 (gelezen) · OpenAI: GPT-6.2 Astra en Sol zijn beschikbaar op X · OpenAI-ontwikkelaars over API-namen en Ultrafast op X · Perplexity over GPT-6.2 op WANDR op X · @kimmonismus over GPT-6.2 op X · @Teknium over tegenstrijdige demo's op X
Claude Code gets an experimental Code Mode; linked billing in betaConfidence: medium: official account, but the documentation page could not be read; the token saving is an expectation.
Claude's developer account announces a new, experimental way for Claude Code to work with tools: Code Mode (697 thousand views). Claude writes scripts that chain tools together instead of calling each tool separately. According to the post and secondary reports, you turn it on with the environment variable CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1. The matching documentation page is titled Run tool calls from code with Code Mode; that page could not be loaded today, so the details come from the post.
The same account reports linked billing in beta: users of Factory, Linear, Ramp and Warp, among others, pay for Claude usage through their own Claude Platform organization without sharing an API key (245 thousand views). Claude Desktop also has a Latest release channel with faster fixes but less reliable builds than Stable (174 thousand views).
What this means: anyone who connects many MCP servers may use fewer tokens with Code Mode; that is an expectation, not a measurement. Test it in a test repo first and compare usage with a normal run. Earlier news about the tool is in our article on Claude Code 2.1.326.
Sources: @ClaudeDevs over Code Mode op X · @ClaudeDevs over linked billing op X · @ClaudeDevs over het Latest-kanaal van Claude Desktop op X
Perplexity launches the Orchestrator API for dozens of modelsConfidence: high for the operation and price structure: the documentation was read directly; medium for the number of models, spend limits and allowed-model lists: only the post, not found in the documentation read; the revenue figure is a statement by the chief.
Perplexity introduces the Orchestrator API (209 thousand views): models from OpenAI, Anthropic, Google, xAI and open-weight providers behind one API with unified billing, where the provider is chosen automatically based on availability, capacity and latency. The documentation was read: the API works with the OpenAI Chat Completions, OpenAI Responses and Anthropic Messages formats, bills per token at the rates of the requested model without a per-request surcharge and resends a failed request to another deployment of the same model. Bring-your-own provider keys are not accepted and some models require a higher usage tier.
The post mentions more than 45 models and spend limits and allowed-model lists per key. The FAQ that was read gives no total count and no spend limit; access is restricted there through usage tier and organization settings. Perplexity also reports a new effort level, Multiagent, in Computer and Teammates (6.5 thousand views) and annual revenue above 1 billion dollars according to Aravind Srinivas (521 thousand views).
What this means: the API is a gateway for anyone who wants to switch models without routing of their own, with the limits that own keys are not possible and that the cost of the choosing step under auto-routing is not described separately. Read the prices and terms yourself before connecting anything. An earlier article covered the product: our article on Perplexity Teammates.
Sources: Perplexity: Introducing the Orchestrator API op X · Orchestrator API: veelgestelde vragen (gelezen) · Orchestrator: modellen en prijzen · @testingcatalog over het Multiagent-niveau op X · @AravSrinivas over 1 miljard dollar ARR op X
Agents in business: Grok Bot in accounting and Microsoft Autopilot on OpenClawConfidence: medium: own announcements by the accounts involved; the blog and a Microsoft source were not read; no figures on results.
The account @bot reports that five Grok Bots have a full-time role in SpaceXAI's accounting team, with a blog about the order-to-cash process (1.56 million views); Elon Musk shared the post saying Grok Bot can do your accounting (1.16 million views). A SpaceXAI employee writes that a bot becomes more capable the more context it gets through connectors (66 thousand views). The blog was not read; the description comes from the post.
The OpenClaw project reports that Microsoft Autopilot is built on OpenClaw and is in preview for customers of the Frontier program (17 thousand views); a Microsoft employee shared the architecture. The project also reports that Sign in with ChatGPT is out of beta, so a ChatGPT subscription, Business included, works in OpenClaw (26 thousand views). No Microsoft announcement or documentation was read.
What this means: both reports come from the companies' or projects' own accounts and show agents being used in accounting and office software, but without independent confirmation or figures on results. For anyone running OpenClaw: our article on OpenClaw 2026.11.4 describes an earlier release.
Sources: @bot over vijf Grok Bots in het accountingteam op X · @elonmusk over Grok Bot en boekhouding op X · @poteto over context via connectors op X · @openclaw over Microsoft Autopilot op X · @OmarShahine over de architectuur op X · @openclaw over Sign in with ChatGPT op X
SGLang v1.0 with a consortium of companies and a free Mistral Vibe in EuropeConfidence: low to medium: the Mistral post is on the official account; the consortium comes from posts by individual accounts; no primary source read.
Several accounts report SGLang v1.0 and an SGLang Consortium, with NVIDIA, AMD, Intel, Google, Meta, Thinking Machines, Cloudflare, Modal and RadixArk, among others, as founders (5.5 thousand and 11 thousand views). A primary source, such as a project post, was not read. Mistral's account reports that Mistral Vibe and the Le Chonk model are free for three months in Europe (252 thousand views); the terms were not read.
What this means: for anyone running models themselves, an open inference engine with broad backing is interesting, but check the founder list at the project itself. For developers in Europe the free period is a reason to try the coding agent; read the terms first. Earlier coverage of the model is in our article on Mistral Large 4 Le Chonk.
Sources: @DataChaz over SGLang v1.0 en het consortium op X · @MistralAI over Vibe en Le Chonk gratis in Europa op X
Unconfirmed: Grok 4.8, a Hermes harness in Telegram, a slide from Anthropic's roadshow and paste-promptsConfidence: low: one or a few posts per claim; no primary source read.
The following claims circulate on X and are unconfirmed. One account reconstructs confusion about Grok 4.8: Elon Musk reportedly waits for that version while the reply bot @grok told a user that a soft rollout to Bot users is already under way; that reply comes from a bot and is not official (135 thousand views). Another account claims that a built-in agent harness in the Telegram macOS beta runs on Hermes Agent (3.7 thousand views). An account shares a slide said to come from an Anthropic IPO roadshow; that is secondhand and not verified (200 thousand views). Fan accounts report updates to Grok Build.
Warning: many of the most viewed Grok posts of the past day are @grok-is-this-true posts about an alleged attack between the United States and Iran; they have no news value for AI and look like a disinformation pattern. Paste-prompts for trading bots and claims of profits with Grok Bots also circulate; do not use them and do not install anything from a list without checking the source.
What this means: do not adopt rumors, paste-prompts and secondhand slides without a primary source.
Sources: @Marko_Poly over de Grok 4.8-verwarring op X · @SenditHighor over een agentharnas in Telegram op X · @Brandondoyle over een slide uit een roadshow op X · @XFreeze over Grok Build op X
Primary sources: The Decoder: lancering van GPT-6.2 Astra en Sol · Perplexity: Orchestrator API, veelgestelde vragen · ClaudeDevs over Code Mode (X)


