Farseen

haha, so apparently OpenAI are about to start training on your Daybreak Blue sessions, with no way to opt out. Literally stealing people's private work, as they do it, is cartoonishly evil. It also will presumably kill Daybreak usage by security researchers. Or at least, it should. The thing OAI have had to swear they are not doing to math researchers, they are now explicitly about to start doing to security researchers. Yet another reminder of the importance of open models.

Your coding agent pays every session to relearn the same repo. Microsoft Research and UC Berkeley measured a fix: 18.4–30.9% lower mean cost on SWE-Bench tasks. TokenEconomizer is an open-source plugin for Claude Code, Codex, Copilot CLI and Pi. It doesn't touch the model or the agent: • Tool pruning: a tool unused for 10 prompts across 2+ sessions (or for 5 sessions) is left out through the agent's own settings. One command brings it back. • Loop guard: when a turn keeps hitting the same failures, a judge model sends one redirect. It never stops the turn; it tells you how. • Micro-skills: learns repo facts like "python -m pytest, not pytest". Off by default, since it adds model calls. • 277 SWE-Bench tasks, 5 agent-model combos: cheaper in every repo for all 5 combos, and first-try solve rate up in 4 of 5 (team-reported). • Apache 2.0. No API key of its own; what it learns stays on your machine. My take: tool pruning is the part I'd turn on today. Every loaded tool definition is billed on every request, and most of us carry tools we forgot we installed. Caveat: it's tested on Claude Code 2.1.300 to 2.1.310. The current release is 2.1.314, not tested yet. And with Opus 4.8, the first-try solve rate dropped 1.0 point. Would you let a plugin edit your agent's tool list?

Attended a fun session at a16z with @illscience and @joshelman yesterday. As usual they were dropping knowledge bombs: 1. There are three economic miracles. The old one is zero marginal cost software distribution. But recently there have been two more: 1) Coding going from the most precious asset in the world to one of the cheapest and 2) zero marginal cost digital labor with agents. 2. It's hard to appreciate what it feels like to be on the exponential. Just 10 to 20 million people account for $150 to 200 billion of token spend. "Some of the best companies that we work with have higher token spend than payroll." We are still early. 3. The AI workspace, like ChatGPT and the Claude desktop app, is going to be the container for all kinds of AI interactions and it's going to subsume the browser. Personally, these apps are already the browser for me. Chrome is just the middle layer. 4. The way to tell if you're an AI-native company is to ask: If the model gets twice as good, do your customers get twice as much value? Are you happy or sad that the model got better? If you're sad, you're probably on the wrong side of the exponential. 5. "I would much rather have AI be the boss of me than vice versa." For example, Anish's agent is responsible for a 30 minute block in his calendar every week to pick someone he should reconnect with. 6. Improving individual productivity does not always improve a team's productivity. Coding has taught us the wrong lesson on this. For many functions, having individuals be 10x more productive doesn't mean the team will be 10x. e.g., If everyone goes in a different direction. 7. It's actually easier to start a hard company than an easy company... because they're all hard. At least the hard companies are interesting to work on and people want to work with you on them. Excited to do a podcast recording with Josh next week to go deeper on some of this.

We identified a security incident affecting a modified Claude Code VS Code extension accessed through our Cursor IDE extension marketplace, impacting at most 5,000 users. This malicious version of the extension was not developed by Anthropic. It does not impact other extensions downloaded from the cursor.com/marketplace, Grok Bot, Grok or our other SpaceXSI products. In response, we took down our extension marketplace this afternoon, worked with relevant providers to neutralize the malware, and have reinstated the marketplace after confirming there is no further risk. If affected, please review instructions in this blog post to understand how to find the malicious package and establish a path to mitigate any potential impact. We are asking all customers update to Claude Code for VS Code version 2.1.314: cursor.com/blog/claude-code-…

ChatGPT web app now has a new hidden shared usage limit checkout disclaimer - "Starting in December, plan usage will have a shared limit across all of ChatGPT and Codex"

When we launched Jules, it was an experiment to learn how asynchronous coding agents can help developers ship faster. Seeing what you built taught us so much about where agentic development is headed, directly shaping several products across @Google. We’re working on bringing our learnings from Jules into @antigravity so you get the best of both worlds: autonomous remote background task execution and an agent-first development environment. As we focus on this unified experience, Jules will be winding down. Whether you used Jules to crush bugs, write tests, or launch features— thank you for shaping the future of agentic coding! 💙🦑

We just shipped our biggest improvement to our provider router since the founding of OpenRouter. Before, we prioritized mostly based on price. The new router now rewards the best quality inference for the actual cost of a request, so requests stay cheap but are also fast and reliable. Providers that deliver more value automatically win more traffic, so every provider has a reason to keep improving. And every time a provider gets better, so does your inference on OpenRouter.

We’ve announced the Artificial Analysis Progress Index: a single capability scale for every language model since 2022, launched alongside Intelligence Index v5 The Intelligence Index is still our primary measure of current model intelligence. It uses the evaluations most relevant to developers today, so its mix changes as evaluations are added and retired, and scores from different versions can’t be compared. The Progress Index is built for comparisons across time. It uses item response theory, a statistical method used to score standardized tests, to estimate one general capability score for each model from the underlying tasks in all our evaluations, including those that are retired or excluded from the Intelligence Index. Use the Intelligence Index to choose a model today. Use the Progress Index to see how capability has changed over time. The Progress Index builds on prior work placing models on a single capability scale, in particular @EpochAIResearch's Epoch Capabilities Index and the capability scale @AnthropicAI published in its Claude Mythos Preview system card. Key takeaways: ➤ Since reasoning models arrived, the frontier has risen ~65 to 70 points a year, about double the 2023 pace: the highest score among released models rose 35 points in 2023 (GPT-3.5 Turbo 125 to GPT-4 Turbo 160), 67 in 2024 (to o1 at 228), 68 in 2025 (to GPT-5.2 at 296) and 56 so far in 2026 (to Gemini 4 Carbon at 352). ➤ Open weights models generally trail proprietary models by about 4 months: since January 2025, each new highest open weights score came 2-8 months (median ~4) after a proprietary model first reached it. DeepSeek V4.1 Flash 1017 is the highest-scoring model with released weights at 324, a level Claude Fable 5 reached in June 2026. Kimi K3.1 Pro (weights not yet released) scores 341, about two months after the closed frontier ➤ New frontier scores are being set more often: the highest score among released models was raised on 2 release dates in 2023, 4 in 2024, 9 in 2025 and 11 so far in 2026. That is about once every four weeks this year. ➤ OpenAI has held the top score most of the time since November 2022, but the frontier has become a close contest: OpenAI held the number one position for the first 27 months measured, but in 2026 especially there have been rapid changes in the leader between OpenAI, Anthropic, and now Google

mistral large 4 (1t params) goes open-weight. per mistral, it leads cognition's frontiercode 2.0 at 19.7% (deepseek v4.1 flash 19.3%, glm-5.3 17.8%) and effectively ties glm-5.3 on the aa coding agent index at 48%, with 47% fewer tokens per task. cheaper agents, not just smarter

long agent rollouts can get expensive fast together post-training uses per-gpu pricing, so running more rollouts in parallel brings the cost of each one down together.ai/post-training