Opus 5.5, GPT-6 Sol and Luna

Start here ↓

Anthropic released Claude Opus 5.5, the first model in the Claude 5.5 family. It performs roughly at Claude Fable 5.1-level while costing 40% less than Opus 5.

Opus 5.5 generates output more than 30% faster than Opus 5, improves on agentic coding and long multi-step tasks, and is priced lower across input, output, and especially cache reads.

Opus 5.5 also ships with higher five-hour limits for Pro, Max, Team, and seat-based Enterprise users, plus a saved rate-limit reset for subscribers.

Also, icymi, do watch the launch video too!

Anthropic announcement | Claude on X

🚀 Shipped

OpenAI shipped GPT-6 Sol and Luna for faster, cheaper GPT-6 workloads. They bring much of GPT-6 Astra’s strengths to workloads that need scale, with API prices 50% lower than GPT-5.6 promotional pricing. This isn’t simply a budget version of Astra: in Artificial Analysis’ independent testing, GPT-6 Sol max scored 48 on its Intelligence Index, placing it among the leading models while costing $2/M input and $10/M output tokens.
OpenAI announcement | Artificial Analysis

Google put a Jev-compatible model on Cloud Run for about $3 an hour. It’s a one-command Cloud Run setup for DiffusionGemma-Jev, or djev, with a Jev-compatible API and no GPU for you to manage. Google reports 35–60ms latency for one decision and 100–123 requests per second at batch 32.
Google Gemma announcement | Deployment repository

CopilotKit released an open-source alternative to Meta Muse. OpenMuse is a self-hostable personal agent in the mould of Meta Muse that can run on any agent harness. It comes with computer use across a browser, terminal, and files; connectors for personal apps; goal and progress tracking; and clients for mobile and web.
OpenMuse repository | Atai Barkai’s announcement

Firecrawl raised $75M and built the library agents keep trying to assemble themselves. They launched Alexandria, one place for agents to search the live web, official data providers, custom connectors, and Firecrawl’s own indexes. Ask for jobs above $150K and apartments below $4K, for example, and the agent can compare entire job and housing datasets instead of scraping the first few links. Alexandria already spans 100+ providers.
Firecrawl announcement | Launch thread

Jev-style decisions can now see images, hear audio, and watch video. A developer released Jev-Omni, an open-weight model that makes bounded decisions over text, images, audio, and video. Give it a question and fixed options, and it returns a probability for each instead of writing an explanation. It is built on Gemma 4 12B.
Jev-Omni model card | Akhila’s announcement

Cloudflare launched production-like Worker previews for every Git branch. It gives every Git branch its own production-like environment with separate code, configuration, URL, observability, and state. Durable Objects and Containers are isolated per branch too, so teams and coding agents can test larger changes without sharing state or colliding with one another.
Cloudflare Worker Previews | Cloudflare on X

Unreal Agent harness matches Codex + Astra for 39% less on agentic coding. Unreal Labs released this open-source coding-agent harness that keeps sessions, tool operations, and state alive independently of model turns. The team reports a score of 62 on Terminal-Bench 4.0 at $3.60 per task, 39% cheaper than its Codex plus GPT-6 Astra comparison while maintaining performance.
Unreal Agent repository | Unreal Labs announcement

Jev finally has a public leaderboard, and the open models are close. Benchmark Heaven published JevBench, testing Jev and its open alternatives across 534 English decisions for intelligence, calibration, speed, and cost. Jev leads, but SemIf and djev sit within a point on the benchmark’s combined score.
JevBench leaderboard | JevBench repository | HN discussion

🧠 Worth Knowing

Jev opened its doors, then demand forced them shut a day later. TypeSafe AI temporarily paused new Jev signups after what it described as an immense surge in demand. Existing accounts continue to work, but new users cannot currently register. This comes just one day after TypeSafe removed Jev’s waitlist and opened signups to everyone.
TypeSafe AI announcement

Anthropic says to stop telling Opus 5.5 to “think carefully”. They published an Opus 5.5 playbook for Claude and Claude Code. Its main recommendation is to provide the complete task, define exactly what “done” means, and say when the model should stop and ask. You can also remove “think carefully” instructions because Opus 5.5 already decides how much thinking each reply needs.
ClaudeDevs post | Opus 5.5 playbook

Changing Opus 5.5’s effort mid-session no longer breaks the prompt cache. Lydia Hallie from Anthropic posted that changing the effort level during an Opus 5.5 session does not invalidate the prompt cache, provided you are using Claude Code v2.1.280 or newer. That lets long sessions move between lighter and heavier reasoning without paying to process the entire cached context again.
Lydia Hallie on X

Opus 5.5 turned a formal-verification experiment into 16 bug-fix PRs. Boris Cherny used Opus 5.5 and Lean to formally verify the Claude Agent SDK, producing 16 PRs that fixed bugs and race conditions. He also combines Lean with TLA+ to look for problems in data flow, concurrency, and state management, despite not knowing either language well. Formal verification is becoming usable without first becoming a formal-methods expert.
Boris Cherny on X

Udemy founder announced The Horowitz Andreessen Academy, a selective San Francisco school for ambitious young builders after high school or early college. Instead of traditional grades and homework, students work on self-directed projects, short courses, and company co-ops. Its founding partners include Anthropic, OpenAI, Google, Meta, NVIDIA, Stripe, Replit, Coinbase, Palantir, and Anduril.
Gagan Biyani’s announcement

Build a local Jev-style scorer without training a new model. Avi Chawla published a walkthrough for turning an open model into a local Jev-style decision engine. The application supplies the allowed answers, SGLang scores them directly, and the system returns a probability distribution without generating JSON or prose. It recreates the inference pattern, not Jev’s training, calibration, or complete evaluation stack.
Avi Chawla’s walkthrough

Perplexity launched a Research Fellowship for early-career researchers, engineers, and analysts from any technical or quantitative discipline. The unusually broad eligibility means the program is not limited to candidates already following a conventional machine-learning research path.
Aravind Srinivas’s announcement | Fellowship page

🔧 Clone and Run

Clone & Run of the Day

Cloudflare open-sourced a security-audit skill that coordinates parallel agents across reconnaissance, vulnerability hunting, adversarial validation, reporting, structured output, and independent verification. It only reports findings with concrete attack scenarios.
Cloudflare security-audit skill

Build Jev workflows by connecting decision blocks in a visual editor. The Jev workflow builder lets you connect choices, scores, booleans, and LLM prompts inside a multiplayer visual canvas, test the resulting workflow, and call it through a REST API. It requires credentials for Liveblocks, TypeSafe, and Vercel AI Gateway.
Jev workflow builder | Chris Nicholas’s demo

Anthropic published an open-source Claude Code Migration Kit with prompts, templates, dependency-map scripts, rulebooks, agent boundaries, and behavioral-parity checks for large language migrations. It’s a reference code that is not actively maintained, and the included prompts are reviewed reconstructions rather than transcripts from its production migrations.
Claude Code Migration Kit

Give a personal AI memory without summarizing away the original conversation. Lossless Memory is a small open-source local memory system that preserves every conversation line, timestamps every record, and searches by time before semantic similarity.
Lossless Memory repository

Read the instruction files used by 100 major open-source projects. ossrules.md collects real AGENTS.md, CLAUDE.md, and skill files from 100 open-source projects, covering more than 1,200 skills. You can inspect how projects such as Hermes, Codex, PyTorch, Playwright, and Next.js teach agents their conventions.
ossrules.md

Run StepFun’s coding agent from your terminal. Step Code is an open-source terminal coding agent that reads code, edits files, runs tests, coordinates subagents, loads MCP servers and Agent Skills, and publishes static sites through StepPage. Its /goal mode handles longer autonomous tasks.
Step Code repository

Give agents one endpoint for more than 3,000 tools. treg is like OpenRouter for agent tools. It lets agents call more than 3,000 catalogued endpoints and team-owned tools through one base URL and token, while credentials stay on the server. You can self-host it.
treg repository

Parse PDFs and Office files locally without an LLM. LiteParse is an open-source local document parser with Rust, Python, Node, and WASM packages. It extracts spatial text and bounding boxes from PDFs, converts Office files through LibreOffice, supports images and OCR, and can identify pages that need heavier processing. Complex tables, handwriting, and difficult scanned documents may still need a larger parser.
LiteParse repository

Awesome LLM Apps is a curated collection of 100+ AI Agents, Agent skills, and RAG apps. It covers models from OpenAI, Anthropic, Google, and open-source models like GLM, DeepSeek, and Qwen that you can run locally on your computer. (Now accepting GitHub sponsorships)

That's all for today. Come back tomorrow for the next batch of AI tools, model drops, agent repos, and weird benchmarks worth your time.

If you found one thing to try, share the issue with someone who ships.

Unwind AI - X | LinkedIn | Threads

Awesome LLM Apps | Sponsor Us

Most founders are one system away from turning LinkedIn into their best sales channel.

Engagement is easy to mistake for pipeline. On Sep 30, watch how a founder turns LinkedIn content into real outreach. Live. You'll walk away with a repeatable system: what to post, who to reach out to, and how to sequence it. 

Eligible startups also get the LinkedIn-to-Leads Toolkit: ad credits, Apollo, Captions, and HubSpot's Prospecting Agent.

Claim my spot