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 latestFable 5.1 stays, Tencent Hy4 debuts, Apple Silicon speedup ships
Claude Fable 5.1
Continuing previous issue from yesterday, the pricing edge widens further.
Anthropic confirmed Fable 5.1 is now live on all platforms—AWS, Google Cloud, Azure, and the Claude API—with input at $10 per million tokens and cache reads locked at $0.25 [Source: Anthropic]. Mythos 5.1, the unrestricted sibling, routes through trusted-access programs for verified researchers in cybersecurity and life sciences, using identical internals but relaxed guardrails. Terminal-Bench 4.0 holds at 55.8% on agentic coding; Humanity's Last Exam hits 65.0% with tools.
Expect enterprise agents to default to Fable 5.1 now.
Tencent Hy4 open-source
Tencent's 770B mixture-of-experts model enters preview for coding and research.
Hy4 activates only 49 billion parameters per request despite its full size, designed for software engineering, financial analysis, and research workflows [Source: The Daily Star]. Integration into Tencent's CodeBuddy and WorkBuddy tools is underway; the early version trades complexity latency and answer verification costs for scale. Open-source release follows, expanding the field beyond closed-API incumbents.
Watch for latency benchmarks against Fable and Muse on long-context tasks.
Rapid-MLX 0.13.4 on Apple Silicon
Local inference on M-series Macs got measurably faster this week.
Rapid-MLX 0.13.4 ships verified speedups across Qwen3.8-27B: decode throughput jumps 1.43× to 2.34× depending on prompt length, with time-to-first-token improvements from 0.54s to 0.51s on short prompts and 25.04s to 24.66s on 8K contexts [Source: GitHub]. New multi-token prediction qualification lands alongside maintained tier-1 compatibility across Claude Code, Aider, DeepSeek Harness, and four other production agents tested end-to-end. The model catalog expanded to 253 aliases covering text, image, video, and audio, all running as an OpenAI-compatible backend with no external API calls.
Local agent infrastructure just shed another layer of friction.
Introducing Claude Fable 5.1 and Claude Mythos 5.1 - Anthropic2 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
PUBG parent Tencent releases open-source AI model | The Daily Star9 hours ago ... Chinese technology group Tencent, maker of hit mobile game PUBG, has released a preview of a new open-source artificial intelligence model aimed at software ...thedailystar.net