Sponsored by

ai-fire-banner

In METR’s tests, top AI agents had a 50% shot at tasks that could take a human expert 27 hours. So how many people could still beat OpenAI’s GPT‑6 Astra and Anthropic’s Claude Opus 5.5 in a coding contest?

IN PARTNERSHIP WITH CHATFEATURED

"How did you hear about us?" “ChatGPT.”

Buyers ask ChatGPT who to hire, then contact the companies it names. ChatFeatured shows where AI recommends you and where it leaves you out, then writes and publishes the content that gets you cited; all while reporting AI traffic and the leads it brings.

AI INSIGHTS

ai-agents-task-horizon-doubling-every-110-days

METR just put OpenAI’s GPT‑6 Astra and Anthropic’s Claude Opus 5.5 through some of its hardest tests yet. The number that jumps out is 110 days. That’s roughly how long it took for the length of tasks top AI agents could finish with a 50% success rate to double. That length is measured in how long the work would take a skilled human.

  • Some tasks now stretch to 27 hours: On METR’s newer tests with clear scoring, the best agents had about a 50% shot at finishing tasks estimated to take a human expert up to 27 hours.

  • The best humans still took the top spots: In programming contests, the strongest models landed in the 96th to 98th percentile. In a typical contest, about four people still beat the best AI result.

  • AI got generous when grading AI: METR asked models to design new tests for its researchers. None made the cut. When Claude Opus 5.5 graded tests made by GPT‑6 Astra, it scored them about 20 points higher on a 100-point scale than METR’s human reviewer did.

That last bit matters if you’re betting on AI to speed up AI research. A coding test can tell an agent in seconds whether its fix worked. Picking which experiment is worth a week of compute takes judgment, and the feedback may come days later. How well agents make those calls could shape how quickly they change the way AI labs work.

PRESENTED BY HUBSPOT

Turn AI into Your Income Engine

Ready to transform artificial intelligence from a buzzword into your personal revenue generator

HubSpot’s groundbreaking guide "200+ AI-Powered Income Ideas" is your gateway to financial innovation in the digital age.

Inside you'll discover:

  • A curated collection of 200+ profitable opportunities spanning content creation, e-commerce, gaming, and emerging digital markets—each vetted for real-world potential

  • Step-by-step implementation guides designed for beginners, making AI accessible regardless of your technical background

  • Cutting-edge strategies aligned with current market trends, ensuring your ventures stay ahead of the curve

Download your guide today and unlock a future where artificial intelligence powers your success. Your next income stream is waiting.

AI SOURCES FROM AI FIRE

1. 5 UI Libraries That Make Your Claude Apps Look Polished. Pick the right library for a quick prototype, dashboard, or business app. Give Claude ready-made buttons, tables, and menus so you spend less time fixing the design.

2. Can Gemini 4 Neon Beat GPT-6.2 and Claude Opus 5.5? We break down where each model wins on the benchmarks and share the two prompts we used to build a 3D dragon from math and a playable survival game with Neon.

3. Turn Claude Code Into a 3D Animation Studio With One Free Skill. Follow five steps to make your first faceless YouTube Short, from script to final video. Keep the 3D characters and style guide so your next episode is easier to make.

FIRE RECAP: BIGGEST AI NEWS THIS WEEK

🚀 OpenAI just launched GPT‑6.2 Astra for tough projects and GPT‑6.2 Sol for everyday work. On Agents’ Last Exam, Astra scored 5.5 points ahead of Claude Fable 5.2 at just 5% of the cost.

🧮 Claude helped Anthropic mathematician Levent Alpöge produce a proof of the Riemann hypothesis, a 167-year-old problem about prime numbers. The Lean and Mathlib teams verified the proof. The Clay Mathematics Institute still lists the problem as unsolved.

📈 Anthropic just went public on Nasdaq under ANT. Its $93.5B IPO was the biggest on record, and a separate $26B private deal drew investment from Google, Microsoft, Nvidia, and others.

⚡ Google’s Gemini 4 Neon edged out Gemini 4 Carbon in Artificial Analysis’s tests, scoring 61 vs. 60 with more reasoning effort. Neon cost about 17% less per task compared with Carbon at standard pricing.

🔒 OpenAI is tightening defenses against model copying, which can pass on an AI’s skills without its safeguards. In OpenAI’s own tests, attempts to extract GPT‑6.2 Astra’s private reasoning succeeded less than 1% of the time.

TODAY IN AI

AI HIGHLIGHTS

🎞️ Google Gemini can now turn a prompt into a looping GIF. Think animated photos or reaction memes you make yourself. It runs on Gemini Omni 1.2 Flash and is free in 8 countries, including the U.S. and U.K.

🧩 ChatGPT Space is rolling out worldwide to Plus users 18 and older. You can build interactive pages and slides with your team, or set a project plan to update itself when new decisions land in Slack.

🛍️ Longtime Etsy sellers have closed their shops over new terms that would let Etsy and its partners use seller content to train AI. After the backlash, Etsy told The New York Times it would revise the policy.

🎮 Benj Edwards got an unofficial ChatGPT-style bot running on a real Super Nintendo. It uses the same chip Yoshi’s Island used for graphics and can even tell you who Mario is.

🩺 AI could turn your job into low-paid gig work. Yale’s Julián Posada writes in Rest of World that platforms like Scale AI’s Outlier draw in doctors, lawyers, and other experts with high rates that can drop without warning.

💰 AI Daily Fundraising: Former Alibaba exec Sisi Song’s Everful raised $48.5M in Series A funding with backing from Lightspeed. The startup uses AI agents to help small retailers buy directly from Chinese factories, and Song says it’s on track to hit $100M in annualized revenue by year-end.

NEW EMPOWERED AI TOOLS

  1. 🔎 Dageno AI 2.0 tracks how ChatGPT, Gemini, Perplexity, and 7+ other AI models recommend brands, revealing competitor rankings and cited sources.

  2. 🤝 AlterContacts is a voice-first AI relationship assistant that remembers client meetings, updates CRM records, and prepares personalized follow-ups for approval.

  3. 🧠 AI Context Flow MCP gives Claude, ChatGPT, Cursor, and other AI agents shared memory with access controls, version history, and rollback.

  4. ⚡ Codemany runs GitHub Actions on your own infrastructure with isolated virtual machines, claiming 3x faster CI without per-minute billing.

  5. 📧 Glassmail is a private Mac email app with on-device AI for Gmail, Outlook, and iCloud, helping you summarize emails and draft replies without subscriptions.

AI BREAKTHROUGH

laptop-model-beats-qwen-and-gemma-at-coding

Thinking Machines Lab put a MacBook Pro in airplane mode and gave its new Inkling 2 Mini a job. A few minutes later, it had built a playable flight simulator. The 26.4B open-weight model looks even more interesting on the coding charts.

  • DeepSWE v1.1 - Inkling resolved 48.2% of tasks, ahead of Alibaba’s Qwen3.8-27B at 42.2% and Google’s Gemma 4 31B at 0.3%.

  • SWE-bench Pro v2 Hard - Inkling scored 77.9% to Qwen’s 73.8%. Qwen still took the lead on Terminal Bench v2.1, scoring 73.0% to Inkling’s 71.6%.

The trick is a looped transformer. Inkling runs the same middle layers three times per token, giving it more room to work through a problem before it writes the next bit of text. That can help it solve coding tasks with fewer tokens.

The chart’s speed numbers are theoretical, and test setups vary by model. The offline demo ran on an M5 Pro MacBook Pro with 48GB of RAM. If your Mac has at least 32GB, you can try Inkling through Ollama and see what it can do with your own codebase.

Finally, a Monday that doesn’t feel like one.

Imagine a Saturday where you're not catching up on the CRM you didn't touch all week. 

HubSpot's new self-updating CRM doesn't wait for you to log back in. Friday's calls are already logged. The deal that needs a nudge Monday morning is already flagged. The automation you set up once keeps running without you thinking about it. None of it needed you over the weekend. 

Customers are already seeing it pay off: 45% more deals closed. Finally, you start Monday caught up, not behind.

We read your emails, comments, and poll replies daily

Hit reply and say Hello – we'd love to hear from you!
Like what you're reading? Forward it to friends, and they can sign up here.

Cheers,
The AI Fire Team

Reply

Avatar

or to participate