Skip to the page
Tecyfy

16 November 2026 to 22 November 2026

GPT-6.2 and Gemini 4 Neon ship; a LiteLLM flaw is exploited

AI and data news from 16 November 2026 to 22 November 2026: Claude Code starts in auto mode on Opus 4.6 and Sonnet 4.6 on clouds; Gemini 4 Neon is generally available at $3 and $15 per million; The Gemini API reads Word, PowerPoint and Excel files directly, and 11 more.

FeatureAnthropic

Claude Code starts in auto mode on Opus 4.6 and Sonnet 4.6 on clouds

From Claude Code 2.1.328 on 19 November, auto mode works with Claude Opus 4.6 and Sonnet 4.6 on Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry and the Claude apps gateway, and an interactive session with no permission mode starts in it. Since 2.1.326 on 16 November, a terminal or VS Code session with no mode set also starts in auto mode, not Manual, when the organization's policy hasn't loaded. If you want prompts for every action on those models, set disableAutoMode.

Where it came from

ReleaseGoogle

Gemini 4 Neon is generally available at $3 and $15 per million

Google made Gemini 4 Neon (gemini-4-neon) generally available on 18 November, at $3 per million input tokens and $15 per million output tokens on the paid tier. It has 4 thinking levels and a 1M-token context window, and it reached GitHub Copilot the same day. Check its price against the model you use now before you switch.

The idea behind it

FeatureGoogle

The Gemini API reads Word, PowerPoint and Excel files directly

Since 18 November the Gemini API takes Word, PowerPoint, Excel and OpenDocument files as they are, with no conversion step. You can also slice documents, spreadsheets, video and audio in a prompt or a function response, and you process and pay for only the part you select. A pipeline that turns Office files into PDF or text before sending them can drop that step.

Where it came from

IncidentOthers

Attackers are using a LiteLLM guardrail flaw, says CISA

CISA added CVE-2026-40217 in LiteLLM to its Known Exploited Vulnerabilities catalog on 18 November, which means it has been exploited. The flaw lets code sent to the proxy's POST /guardrails/test_custom_code endpoint escape its sandbox and run in the proxy; it affects litellm 1.81.8 up to but not including 1.83.10, and in a default setup the caller needs a proxy-admin credential. LiteLLM's advisory says to upgrade to 1.83.10 or later (its text names the fix as 1.83.11, so take that or newer), and to block the endpoint at your reverse proxy or API gateway until you can. Check which version your proxy runs.

The idea behind it

FeatureDatabricks

Databricks guardrails now check streamed answers as they go

Since 18 November, in beta, Databricks service policies on model output evaluate streaming responses as they stream, and this is the default for existing policies too. Databricks notes that text already released can't be recalled, so check what your app does with a partly sent answer. The same policies can now mask or deny sensitive data in embedding inputs, single or batch.

PriceSnowflake

Snowflake caches the repeated start of AI_COMPLETE prompts

Prompt caching for AI_COMPLETE became generally available on 18 November. Snowflake processes a prompt's shared opening once, such as a long system prompt, rules, a schema or few-shot examples, and bills the repeated tokens at a cached input rate up to 97.5% below the standard one. It needs no SQL change, so put the fixed part of a prompt first to benefit.

The idea behind it

FeatureOpenAI

Codex 0.170.0 stops permission updates dropping sandbox limits

Codex 0.170.0 on 18 November fixed a bug where permission updates and child threads lost existing sandbox restrictions. It also fixed races in MCP OAuth credentials during cancellation, login and logout. If you rely on Codex's sandbox, update to it.

ReleaseOpenAI

OpenAI releases GPT-6.2 Astra and GPT-6.2 Sol

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on 17 November, in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and in the API as gpt-6.2-astra and gpt-6.2-sol. Astra costs $8 per million input tokens, $0.40 cached and $40 output, 20% below GPT-6 Astra; Sol keeps GPT-6.1 Sol's $2, $0.10 and $10. Both take text and images and work with the Responses and Chat Completions APIs, and both are on Databricks Unity Gateway from the same day. If you call GPT-6 Astra by name, moving to 6.2 costs less per token; Sol is a drop-in at the same price.

FeatureAnthropic

Your app's users can pay for its Claude usage themselves

Anthropic launched Linked billing in beta on 17 November, behind the user-profiles-billing-2026-10-01 header. A customer links their own Claude organization to your app, and you set billing_policy on each /v1/messages or /v1/chat/completions request to choose who pays. The customer can cap what your app spends, or unlink it, in the Claude Console; it works for organizations that pay Anthropic directly for API usage.

Where it came from

FeatureAnthropic

Claude Desktop lets you pick a Stable or Latest release channel

Since 17 November you can choose which version of Claude Desktop the app updates to. Stable stays the default and gets new versions less often, after more testing; Latest gets one about once a day. Choose Latest only if you want fixes sooner and can live with more change.

Where it came from

FeatureHugging Face

Hugging Face sends language model calls through OpenRouter

Since 17 November, Hugging Face Inference Providers route language model requests through OpenRouter, which picks the provider by live price, speed and reliability and falls back to another if one fails. Your own provider keys no longer work for language models. To keep control, add :fastest, :cheapest or a provider name such as :novita to the model ID, or turn providers off in your settings.

The idea behind it

FeatureAnthropic

Anthropic's text watermark now covers 4 older Claude models

Since 16 November text from Claude Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 carries Anthropic's text watermark on the supported platforms, as text from the newest models already did. A request that sets temperature, top_p or top_k to a non-default value might return text without it. Staying on an older model no longer keeps generated text unmarked.

The idea behind it

Where it came from

ReleaseGitHub

GitHub Copilot drops Claude Opus 4.8, Opus 5 and GPT-5.6 Terra

GitHub deprecated Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra across most Copilot experiences on 16 November. It suggests Claude Opus 5.5 in place of both Opus models and GPT-6.1 Sol or GPT-6 Luna for GPT-5.6 Terra; Business and Enterprise plans with default model enablement get them automatically, and others enable them in model policies. MAI-Code-1.1-Flash follows on 18 December, replaced by the latest MAI Code Flash model.

The idea behind it