AI developer tools · What shipped
For a senior engineer who already reads HN. Real changes in AI developer tools today: releases with version numbers, papers with benchmarks, repos that crossed a threshold worth knowing. Skip hype threads, pre-announcement leaks, and recycled summaries. Always link primary sources.
← See today's latestClaude Fable 5.1 deepens agent lead; Mythos 5.1 unlocks biotech; benchmarks scale
Claude Fable 5.1 & Mythos 5.1
Agentic workloads just got 45% cheaper.
Continuing yesterday's launch, Fable 5.1 cuts costs through 75% cheaper cache reads ($0.25 per million tokens) while lifting Terminal-Bench 4.0 to 60.9% agentic coding and CursorBench 3.2.0 to 73.4% on multi-file refactoring [Quelle: Anthropic]. Mythos 5.1—the identical model with 60% fewer false-positive safety checks—now solves protein binder design 10x better than competition winners and yields GPU kernel speedups up to 2.5x, marking the first model usable for real biotech workflows. Production IDE agents and research labs have a new baseline.
Watch enterprise code migration schedules this week.
GPT-6 Astra extends math breakthrough
Unsolved Erdős problems now have AI solutions.
Epoch AI's FrontierMath Erdős benchmark—68 genuinely open math problems where systems write Lean proofs—saw GPT-6 Astra solve 2 of 68 (3%), the only model to solve any [Quelle: Epoch AI]. The benchmarking hub now covers 391 models across 85 tasks with published evaluation code and per-model output logs via the Inspect framework, turning frontier claims auditable. Reproducibility infrastructure just became table stakes.
Expect competing labs to publish their own Erdős solve rates this week.
Epoch AI: 391 models, 85 benchmarks
Frontier leaderboards just got reproducible.
Epoch AI now tracks 391 models across 85 benchmarks spanning mathematics, coding, continual learning, and game-puzzle domains, each with published task definitions and accessible model output logs [Quelle: Epoch AI]. Evaluations run through the Inspect framework with full transparency—researchers can inspect individual model responses, aggregate scores, and regenerate rankings. The infrastructure shift from opaque leaderboards to auditable benchmarks raises the bar for any vendor claim.
This is what scale and transparency look like together.
Hermes Agent: self-improving framework
Open-source agents just got a learning loop.
Nous Research released Hermes Agent with built-in skill generation from experience, full-text search across session memory, and multi-platform deployment (CLI, Telegram, Discord, Slack, WhatsApp, Signal) [Quelle: GitHub]. The framework runs on any LLM provider (Nous Portal, OpenRouter, OpenAI, custom) and infrastructure from $5 VPS to serverless, with batch trajectory generation and compression for tool-calling model training. One-command setup lands on Linux, macOS, Windows, and Android/Termux.
Self-improving agents hit the DIY tier.
Introducing Claude Fable 5.1 and Claude Mythos 5.1 - Anthropic3 hours ago ... ... AI experts, testing whether the models could match human specialists' performance. Mythos 5.1's capabilities are greater than those of Mythos 5. However ...anthropic.com
AI Benchmarks & Capabilities - Epoch AI2 hours ago ... Our hub for benchmark results, featuring the performance of leading AI models on challenging tasks. It includes results from benchmarks administered ...epoch.ai