Artificial Intelligence: News, Business, Science

Mar 13, 2026

Anthropic is making Claude's extra-large context window a lot cheaper. The Opus 4.6 and Sonnet 4.6 models now offer a context window of one million tokens at the standard price. Until now, Anthropic charged a surcharge of up to 100 percent for requests exceeding 200,000 tokens. The context window determines how much text an AI model can process in a single request.

Opus 4.6 still costs $5/$25 per million tokens (input/output), and Sonnet 4.6 runs $3/$15. But whether a prompt contains 9,000 or 900,000 tokens no longer matters for pricing. On top of that, the media limit jumps from 100 to 600 images or PDF pages per request. The new pricing applies to Claude Code (Max, Team, and Enterprise) and is available through Amazon Bedrock (except for the media limit), Google Cloud Vertex AI, and Microsoft Foundry.

The GraphWalks BFS benchmark measures how well AI models handle logical reasoning across large amounts of text. Opus 4.6 reportedly shows almost no drop in performance even at full context length. | Image: Anthropic

According to Anthropic, both models achieve the highest accuracy among comparable models at full context length in benchmark tests. That said, the broader problem of declining precision as context windows fill up is still far from solved.

Comment Source: Anthropic

Matthias Bastian

Mar 13, 2026

Short News

Elon Musk's AI company xAI is going through a major shake-up. Musk acknowledged on X that the company "was not built right first time around" and is now being rebuilt from the ground up. Six of the twelve co-founders have left xAI since January, most recently Guodong Zhang and Zihang Dai. Only Manuel Kroiss and Ross Nordeen have stayed on alongside Musk.

At a recent conference, Musk admitted that Grok is falling behind competitors like Google, Anthropic, and OpenAI when it comes to coding - but said the company aims to close the gap by mid-2026. To get there, xAI has hired two senior executives from the AI coding startup Cursor: Andrew Milich and Jason Ginsberg, both reporting directly to Musk. According to the Financial Times, Musk has also brought in "problem solvers" from SpaceX and Tesla to help restructure xAI.

Comment Source: The Information | Financial Times | Musk via X

Matthias Bastian

Mar 13, 2026

AI in practice

Google explains the differences between its three Nano Banana image generation models

A new guide from Google breaks down the three Nano Banana image models and when to use each one. The cheaper Nano Banana 2 reportedly delivers 95 percent of Pro’s capabilities and can search the web for reference images on its own before generating output.

Read full article

Comment

Maximilian Schreiner

Mar 13, 2026

AI in practice

AI chips are pushing everything else off TSMC's most advanced production lines

By 2027, 86 percent of TSMC’s N3 capacity could go to AI accelerators, according to SemiAnalysis. Smartphones are becoming a buffer for overflow demand.

Read full article

Comment

Maximilian Schreiner

Mar 13, 2026

Short News

Ukraine opens its battlefield data to allies to train AI models for autonomous drones.

"Today, Ukraine has a unique array of battlefield data that is unmatched anywhere else in the world," Defense Minister Mykhailo Fedorov wrote on Telegram. "This includes millions of annotated images collected during tens of thousands of combat flights."

Fedorov had first announced the plan in January, shortly after taking office. Now, he says a platform has been created that provides allies and companies with constantly updating datasets and large quantities of photos and video footage. The goal is to accelerate the development of AI models that can guide drones to their targets without a pilot or quickly analyze vast pools of data.

Ukraine wants to increase the role played by autonomous systems in the war. Top commander Oleksandr Syrskyi said the war had "entered a new phase" with platoons of drone interceptors are now being created inside the Ukrainian armed forces.

Comment Source: Reuters

Maximilian Schreiner

Mar 13, 2026

AI and society

Bytedance secures access to Nvidia Blackwell cluster in Malaysia, circumventing US export ban on China

Jonathan Kemper

Mar 13, 2026

Short News

Perplexity AI's "Personal Computer" is an AI assistant that works around the clock - handling emails, presentations, and app control. It runs on a dedicated Mac Mini connected to the user's local apps and Perplexity's servers, controllable from any device. CEO Aravind Srinivas called it a "digital proxy" that never sleeps on X. The service builds on Perplexity Computer, which launched in February and bundles multiple AI models.

Security features include a kill switch and an activity log. Access requires the Max subscription at 200 dollars per month, with only a waiting list available for now. Perplexity is also launching an enterprise version that connects to over 400 tools like Salesforce and Snowflake - the company claims it completed 3.25 years' worth of work internally in four weeks. The concept draws comparisons to the controversial OpenClaw, whose developer now works at OpenAI. Agent-based AI systems dominate the current landscape but face sharp criticism around resource demands and security vulnerabilities.

Comment Source: Perplexity

Matthias Bastian

Mar 13, 2026

Short News

Meta has reportedly delayed its next AI model, codenamed "Avocado." Originally set for mid-March 2026, it won't ship until May at the earliest, reports the New York Times, citing three people familiar with the matter.

In internal tests, Avocado fell short of leading models from Google, OpenAI, and Anthropic in logical reasoning, programming, and writing. It beat Meta's previous model and Google's Gemini 2.5 but couldn't match Gemini 3.0. Meta's leadership even discussed temporarily licensing Gemini, though no decision was made. A next-gen model codenamed "Watermelon" is already planned. Meta is also building an image and video generator codenamed "Mango."

Meta says updates are coming "very soon," with more models planned this year. The company found early success with its open Llama models but lost momentum with Llama 4. CEO Mark Zuckerberg has since poured billions into AI, including $14.3 billion in Scale AI. Scale AI's CEO Alexandr Wang now runs Meta's frontier AI division, "TBD Lab," tasked with building superintelligent AI systems. Reports also suggest Meta may be moving away from its open-source strategy.

Comment Source: New York Times

Matthias Bastian

Mar 12, 2026

Short News

xAI's Grok 4.20 can't keep up with the top AI models in benchmarks but hallucinates less than any other model tested. According to Artificial Analysis, Grok 4.20 Beta scores 48 on the Intelligence Index with reasoning enabled, well behind Gemini 3.1 Pro Preview and GPT-5.4 at 57, but still a 6-point improvement over Grok 4.

Grok hängt den neuesten Modellen der großen KI-Labore hinterher. | Bild: Artificial Analysis — Grok trails the latest models from major AI labs in overall benchmark performance. | Image: Artificial Analysis

xAI shipped three API variants: with reasoning, without reasoning, and a multi-agent mode. The model supports a 2-million-token context window and costs 2 or 6 dollars per million tokens; cheaper than Grok 4 and competitively priced among Western models.

Where Grok 4.20 stands out, of all things, is factual reliability. On the AA Omniscience test, it hit a 78 percent non-hallucination rate, a record, according to Artificial Analysis. The test measures how often a model fabricates an answer instead of admitting it doesn't know, alongside factual recall. Grok 4.20 only got it wrong about one in five times when it didn't have the answer.

Comment Source: Artificial Analysis

Matthias Bastian

Mar 12, 2026

Short News

Emil Michael, the US Department of War's chief technology officer, made clear that classifying Anthropic as a supply chain risk is an ideologically motivated move. Claude models "pollute" the supply chain because they have a "different policy preference" baked into them, Michael told CNBC. He pointed to Anthropic's "constitution," a ruleset emphasizing ethics and safety, which he said could result in soldiers receiving "ineffective weapons, ineffective body armor, ineffective protection." The measure was "not meant to be punitive," he added.

Anthropic is the first US company to receive this classification, which is normally reserved for foreign adversaries. The AI company is suing over the designation and has drawn support from Microsoft, OpenAI, and Google employees, as well as former US military personnel. Anthropic has previously pushed back against its own AI models being used for US mass surveillance and autonomous weapons.

The administration has already signaled its intent to control AI along ideological lines by enacting regulations targeting so-called "woke AI," framed as a commitment to political neutrality. The approach echoes the Chinese government's own efforts to exert political control over AI models.

Comment Source: CNBC