AI

The AI Industry Drew Its Own Lines This Week, From Filings to Open Weights and Coding Rules

By Crypt0 · Nov 14, 2026

The AI industry spent this week drawing lines around itself, and every line landed in public. Anthropic told investors it has crossed its own threshold for AI speeding up the building of AI. Baidu promised open weights under Apache 2.0 while keeping a restrictive license on the model beside it. Perplexity shipped decision models with open weights and a free window. The coding tools that millions of developers use every day got new budgets, new retirements, and new shutdown dates. The common thread is plain. The frontier is now governed by paperwork as much as by parameters, and the paperwork is getting specific.

Anthropic filed Amendment 2 to its public offering filing on November 10, and the wording changed in a way that matters. The filing says the company made a determination that it has met its Responsible Scaling Policy threshold for dramatically accelerated progress attributable to AI, based on what it expects from future models and its overall rate of progress, rather than on what it has observed from current models. The threshold means doubling the rate of progress in aggregate AI capabilities, plausibly caused by automated research. Eight days earlier, the November 2 amendment had said only that Anthropic may imminently meet the line. The November 9 system card for Claude Fable 5.2 and Mythos 5.2 judged that an individual model stays short of the threshold on its own, and the filing keeps its determination at the company level, about expected progress. The same week, Anthropic rewritten Supported Regions Policy took effect on November 12, closing Claude to people in unsupported regions and to entities majority owned or controlled from them. The lab is drawing a map of where its models may go, and a line for when its own progress changes gear.

Baidu said November 12 that it has fully opened access to its Harness engine and Agent API, and it says the ERNIE 5.5 Flash Preview weights will be open sourced before November 25. Its product page pledges an Apache 2.0 license for those weights soon, while the press release leaves the license unnamed. The Preview is an intermediate checkpoint with 280 billion total parameters, 14 billion active per token, and a 1 million token context. Beside it sits the ERNIE 5.5 Flash Base model, published November 5 under a custom license that excludes the European Union, the United Kingdom and South Korea, requires written permission above 50 million dollars in revenue, and treats models trained on its synthetic data as derivatives. ERNIE 4.5 shipped under Apache 2.0 in 2025, so the new terms are narrower than the last open release. Until weights and a license file appear in a public repository, the Apache pledge is a promise, and the Base release shows Baidu has used a restrictive custom license on a comparable model. Developers weighing either release need the terms before they commit.

Perplexity shipped Decider v1.3 and Flash v1.3 on November 12, successors to the October Decider launch, with a locate route for extracting text spans and Apache weights on Hugging Face and OpenRouter. Decision models return probabilities, and the new versions add refreshed training. The Artificial Analysis Decision Index lists Decider v1.3 at 70, sixth of 22 on that board, and Flash v1.3 at 64, thirteenth of 22, up four points from Flash v1.2. Pricing matches OpenRouter paid routes at 0.02 dollars per million input tokens for the full model and 0.01 dollars for Flash, and the free endpoint runs through November 19. Perplexity also published an internal security audit anecdote from 7,000 agent sessions and about 1 billion tokens read in a day, leading to 14 rule fixes, six shipped within two days. Small models that decide instead of chat are finding their lane, and the price tag is part of the pitch.

The month belonged to control, and the coding tools proved it. GitHub retired 10 Copilot models across October, with Claude Opus 4.8, Claude Opus 5 and GPT 5.6 Terra scheduled to go on November 16, and its model policy API went generally available on October 19, letting admins manage model allow lists as code. Claude Fable 5.2 shipped November 9 at 8 dollars per million input tokens and 40 dollars per million output tokens, and Claude Code added pooled group budgets, a cost report API, and permission settings that deny instead of prompting. Devin added budgets and self serve sign on, and Cursor shipped its Agent SDK and live app hosting. Google Jules became read only on November 11 and shuts down on December 31, with Google pointing users toward Antigravity. The message for engineering teams is direct. Grep for pinned model names before November 16, put model policies under version control, set budgets on every agent platform, and replace any Jules automation now, while December still leaves room to move.

What it means for readers. The labs are writing their own rulebooks in public filings, the open weight releases come with license text worth reading before you build, and the coding tools now ship with budgets and off switches as first class features. Learn the routing between planners and workers, read the license before you ship, and set a budget before the agent sets one for you. The frontier keeps moving, and the paperwork keeps up.

Sources

• CellCog, Anthropic S-1 AI Acceleration Threshold Met
• TechTimes, Baidu Pledges Apache License for ERNIE 5.5 AI Weights
• ExplainX, Perplexity Decider v1.3 Locate Extraction
• BeevR, AI Coding Agent Updates November 2026

New to crypto? Read the crypto glossary, browse frequent questions, read our story, or explore the story archive.