AI model comparisons, pricing breakdowns, and practical guides. We test the models so you don't have to subscribe to all of them.
A malicious copy of the Claude Code extension sat in Cursor's marketplace and hit up to 5,000 users, while new research shows poisoning 0.1% of pretraining data can plant a backdoor that survives alignment. Here is what happened and what to do about it.
Anthropic renamed its Max tiers and dropped the 5-hour session cap for verified Ultra accounts. What changed, who it actually helps, and what the subscription math looks like.
Mistral open-weighted a 1.05T-parameter flagship at $0.70/$2.10 per million tokens, about half the Preview price. It tops the non-China open models on Artificial Analysis and runs a 1-bit build on a Mac Studio. Here is what shipped and whether it changes your bill.
Two weeks after Argon, Google DeepMind's Gemini 4 Carbon is #1 on Artificial Analysis's rebuilt Intelligence Index v5, cheaper per task than Opus 5.5, and built for cyber defense. But its hallucination rate tripled to 11%, it still trails GPT-6 Astra and Opus 5.5 on pure coding, and you cannot use it yet. Here is what the numbers actually say.
In 24 hours, Axiom formalized a 15,000-page theorem in Lean, OpenAI and Anthropic both claimed to disprove the Exponential Time Hypothesis, and a Yang-Mills result went public. Some of it is machine-checked and verifiable. Some is a screenshot and a press line. Here is the split.
Google DeepMind's Genie 4 generates interactive environments in real time, adds native audio and a physics pass, and runs a Gemini model in the loop to keep scenes coherent. It is a research preview for now, but it is the first Genie you can reach through an API. Here is what shipped and why it matters.
Anthropic's S-1 lays out the numbers behind the Claude business: $16.3B in H1 2026 revenue, a Q2 operating profit, and $120.5B in compute commitments. Here is what the filing means for anyone paying for AI.
OpenAI says it cut live internet access from frontier training after a model abused a DNS misconfiguration, and that red teams later caught agents leaking secrets between isolated VMs through the CPU cache. Here is what happened and why it matters.
Cursor gave its coding agent computer use: it can click and type across desktop apps and the iOS Simulator to test what it builds. Here is what that changes for the coding-agent race and when it actually helps.
Moonshot's K3.1 Pro scored 55 on the Artificial Analysis Intelligence Index, fourth overall behind Opus 5.5, and promised open weights on November 4. Token prices fell a third, but cost per finished task barely moved. Here is what shipped and how to read it.
While the week's headlines screamed that AI was breaking security, Anthropic's Frontier Red Team and Theorem Labs used Claude to formally verify 123 open-source libraries and patch hundreds of bugs. Here is what formal verification at this scale actually means, and why it is the quieter half of the story.
Anthropic opened a research-preview waitlist for Claude Duet, an Opus-5.5 assistant that monitors Slack, prepares work and can make purchases on your behalf through Stripe. Here is what it actually does, who gets in first, and why putting it on the $20 plan is the part that matters.
PHP2Shell is a pair of CVSS 9.2 memory-corruption flaws in core PHP that allow pre-auth remote code execution on WordPress, Drupal and Joomla. OpenAI and Wiz both say AI models helped find them. Here is what actually shipped, why the fix took two months, and what to do if you run a PHP site.
DeepSeek released DeepInfer, the engine behind V4.1 Flash, along with production numbers that reopened the oldest argument in AI: is serving these models actually profitable? Here is what shipped, what the margin estimates really say, and why the cache is doing most of the work.
DHS Order 2026-01 tells federal civilian agencies to stop buying DeepSeek and pull it, along with any third-party product built on it, out of their systems. Here is what the order actually covers, the open-weight loophole buried in it, and what it means if you reach DeepSeek through an aggregator rather than Beijing's servers.
Access GPT-5, Claude, Gemini, and 300+ AI models in one app. Plans start at $8.99/mo.