Juan Roa (@jroadev) header
Juan Roa (@jroadev) profile picture

MSc CS. Sr Software Engineer. Film shooter 📷. Open source guy. Organizer of @rgcalioficial | CEO of @rodarlibre , also building @HoshiraApp 🔭

Bogotá, Colombia
Joined March 2010
Photo by Juan Roa (@jroadev)Photo by Juan Roa (@jroadev)Photo by Juan Roa (@jroadev)Photo by Juan Roa (@jroadev)Photo by Juan Roa (@jroadev)Photo by Juan Roa (@jroadev)Photo by Juan Roa (@jroadev)Photo by Juan Roa (@jroadev)Photo by Juan Roa (@jroadev)Photo by Juan Roa (@jroadev)
Latest versions just dropped, and we have a new best contender: 6.2 Astra. At xhigh or medium, it does really well (accuracy, cost, and speed).
Four new versions have been benchmarked in Agents on Rails: @OpenAI's GPT-6.2 Astra (xhigh and medium) and GPT-6.2 Sol (xhigh), and @deepseek_ai V4.1 Flash 1017 (max). Here’s what we found: New first place: GPT-6.2 Astra. At xhigh effort it scores 65.8%, narrowly ahead of Claude Opus 5.5 (65.3%), at $4.42 per run versus $10.53. It’s also faster than the previous Astra. At medium effort it lands at 4th place and averages under 11 minutes per run, the fastest configuration we’ve tested. GPT-6.2 Sol is nearly twice as fast as GPT-6.1 Sol (about 12½ minutes per run versus 23½) at roughly the same cost. Its accuracy dipped this round, from 60.8% to 54.1%, though some of that may be run-to-run variation. DeepSeek V4.1 Flash 1017 is a big efficiency gain. It scores within a few points of the previous version while running about 1.5× faster at roughly one-sixth the cost. At $0.16 per run, it’s the cheapest model we’ve tested to date, but you get more accuracy and speed with Luna or Sol for just a little more cost. Thanks once again to the team at @evilmartians for their work. Explore the full results at: rubyonrails.org/ai
3
30
1,084
Juan Roa retweeted
Four new versions have been benchmarked in Agents on Rails: @OpenAI's GPT-6.2 Astra (xhigh and medium) and GPT-6.2 Sol (xhigh), and @deepseek_ai V4.1 Flash 1017 (max). Here’s what we found: New first place: GPT-6.2 Astra. At xhigh effort it scores 65.8%, narrowly ahead of Claude Opus 5.5 (65.3%), at $4.42 per run versus $10.53. It’s also faster than the previous Astra. At medium effort it lands at 4th place and averages under 11 minutes per run, the fastest configuration we’ve tested. GPT-6.2 Sol is nearly twice as fast as GPT-6.1 Sol (about 12½ minutes per run versus 23½) at roughly the same cost. Its accuracy dipped this round, from 60.8% to 54.1%, though some of that may be run-to-run variation. DeepSeek V4.1 Flash 1017 is a big efficiency gain. It scores within a few points of the previous version while running about 1.5× faster at roughly one-sixth the cost. At $0.16 per run, it’s the cheapest model we’ve tested to date, but you get more accuracy and speed with Luna or Sol for just a little more cost. Thanks once again to the team at @evilmartians for their work. Explore the full results at: rubyonrails.org/ai
10
12
5
93
9,203
Hey everyone, Ruby 4.0.10 is out! It contains a critical security fix. Update immediately! ruby-lang.org/en/news/2026/1…
12
180
11
507
56,564
Juan Roa retweeted
Join vercel.com/x, our 6-week AI incubation program. You'll get to work directly with the people building internet-scale infrastructure, devtools and agents. I look forward to working with you.
Introducing Vercel X, our AI accelerator. For top builders starting at day -1. Six weeks in San Francisco, January to March 2027. Apply by November 29 ↓ vercel.com/x
112
72
12
1,692
121,867
Juan Roa retweeted
Attackers can take over nginx servers that have HTTP/2 turned on, and every release since 2015 is affected. nginx 1.31.9 and 1.30.7, released today, fix 10 CVEs. The worst, CVE-2026-123007 (CVSS 9.2), is a memory corruption bug in HTTP/2 flow control. Wiz, Xint, Rapid7 and others found it independently. If you can't patch yet, turn HTTP/2 off. Xint called a pre-auth nginx RCE "a holy grail for security researchers." It chained this one with a Linux kernel SCTP bug and got root on the host from inside a Docker container. That kernel bug has no patch yet, so Xint says to block the SCTP module where you don't need it.
11
39
4
191
10,543
Tenemos 2 nuevos CVE en Nginx con ejecución de codigo remota asi que salgan ya mismo a actualizar sus NGINX a la ultima versión. El exploit es por medio de HTTP/2 que básicamente lo van a tener habilitados todos. Nginx ya lo subió arregló y puso el changelog del mismo. No hay paz.
Today we disclose CVE-2026-123633 and CVE-2026-123007 (CVSS 4.0: 9.2) - exploitable memory corruptions in NGINX We achieved reliable remote code execution (ASLR defeated) on NGINX serving static files - no backend or third-party modules 🤯 Both require HTTP/2 Read more 🧵⬇️
33
78
9
792
74,018
Juan Roa retweeted
According to Artificial Analysis, Opus 5.5 is no longer on the pareto line (best value per intelligence) Love AA and all, but they really need some behavioral benches to smooth out this gap. Opus 5.5 is still the best value available for real long running code work.
167
54
17
3,082
363,965
Juan Roa retweeted
A milestone worth celebrating: In the last month, Rails quietly passed 100,000 commits (and counting*)! That's over 100,000 big ideas, small fixes, and countless improvements. And behind every commit is someone who helped make Rails what it is today. No matter how big or small your PR, if you've contributed to Rails, thank you! (*101,179 as of today.)
7
65
7
418
14,703
Juan Roa retweeted
Anthropic's current dominance compared to other labs is insane. IMO this is the biggest lead any lab has had since GPT-3.
Big news: Claude Fable 5.2 (Max) by @AnthropicAI takes #1 in Agent Arena with an unprecedented lead: +30.4% net improvement, about 1.6x the #2 model Opus-5.5, across 4.7k real-world agentic sessions from our global community. It's also #1 in Code, Chat and Work. A major leap over Claude Fable 5.1 (Max) - Rank: #5 → #1, Net improvement: +14.8% → +30.4% - Median cost per task: $7.7 → $7.3 (6% lower). The API pricing $10/$50 → $8/$40 per million input/output tokens (20% lower) vs. The Frontiers - #1 Claude Fable 5.1 (Max): +30.4% | $7.3/task - #2 Claude Opus 5.5 (Max):+19.5%) | $5.5/task - #3 Gemini 4 Carbon (High):+17.0%) | $1.5/task - #6 GPT-6-Astra (Max): +14.5% | $6.0/task Breaking down performance across signals: Fable 5.2 improves over 5.1 across all four reported signals, with particularly large gains in user satisfaction and steerability. - #1 Praise vs. Complaint , +73%, biggest jump vs +41% the next model. - #1 Steerability: +33% - #1 Confirmed Success: +22% - #3 Bash Reliability: +14% #1 across all major categories - #1 Code: +34% - #1 Work: +31% - #1 Chat: +27% - Versus Fable 5.1: Code #6 → #1 (+19.8 %), Work #4 → #1 (+15.8 %), Chat #6 → #1 (+12.0%). Huge congrats to the @AnthropicAI team on this incredible release!
105
55
8
2,227
96,922
Juan Roa retweeted
A thread of early experiments with Claude Fable 5.2: A 3D model of a cell, built entirely from code one organelle at a time. Built by @kevin_t_ngo.
483
1,331
295
17,640
1,440,095
“Code is cheap. Show me your uptime” @rosapolis on the new reality at #SFRuby
1
7
54
1,838
"I think Ruby's best years might be ahead of it." Garry Tan (@garrytan) of @ycombinator opened SF Ruby Conference with words of appreciation for the community and @evilmartians. To Garry, agents need readable, consistent code that does things the same way every time, and that's what makes Rails conventions perfect for agentic coding. Watch the full opening:
2
24
3
139
6,352
Juan Roa retweeted
This Week in Rails is out! There were many updates to the framework this week. Here are a few: - Rails 8.0.5.2 and 8.1.4.1 were released to address 21 security issues, so upgrade if you haven't already. - You can now rotate your master key without downtime, because RAILS_MASTER_KEY accepts several comma-separated keys and Rails tries each until one works. - PostgreSQL null byte errors are now wrapped in ActiveRecord::NullByteError instead of a generic ArgumentError to make invalid inputs easier to handle. If you rescue ArgumentError around queries for this, you will need to update that code. - A series of fixes makes default_order apply correctly in preloading and eager loading, and keeps it out of counts, aggregates and collection cache keys where it caused invalid SQL. There's a lot more to read about, so check this week's wrap-up here: rubyonrails.org/2026/11/6/th…
4
15
3
97
9,031
We are bringing two amazing investors Amir Rustamzadeh (@amirrustam) and Chad Metcalf (@metcalfc), for a facilitated networking sessions with smaller teams. They'll talk about 2026 trends, best practices, and common mistakes in early-stage startups. Victoria Melnikova (@vmelnikova_en), head of new business at @evilmartians, will lead the conversation on Wednesday at 1:00 PM in the Joe Henderson Lab (JHL) on the top floor.
1
4
1
10
739
Juan Roa retweeted
Expo Modules 2.0 on Android needed much cheaper JNI calls, so we built our own JNI library. It’s called Kolibri, and calls into C++ objects are up to 51x faster than with 𝚏𝚋𝚓𝚗𝚒 👏 Full blog post right here from Łukasz Kosmaty: expo.dev/blog/expo-modules-2…
21
34
8
397
19,419
Hermes Agent v0.23.0: The Hearth Release Highlights: - Desktop in a browser tab - One gateway for every conversation - /Rewind - Group Chats with automatic routing - Spend limits for paid tools - Lighter and faster and more Changelog below
221
267
97
4,160
595,394
¡ChatGPT ahora puede crear GIFs! Y quedan de lujo:
20
25
2
677
23,927
Juan Roa retweeted
How @stedapp went from DR 0 to 12 in a week, for $0: 📝 save it and please do it • Google Search Console: verify your site + add your social profiles • Bing Webmaster Tools: import your site from GSC in 1 click • buildhop.io: list your product (DR 52 backlink) @Buildhop_io • ibuildinpublic.com: sign up (DR 29 backlink) @acadictive • indiemakerlist.com: get listed (DR 13 backlink) • macOS app? get listed in macbase.app/ @attacomsian • Your other sites: link them all to each other (dantegaleazzi.com & finikslabs.com in my case) • Add your startup to as many directories and launch platforms as possible • Every bio (X, LinkedIn, GitHub, Product Hunt): your website link • Ahrefs free plan or @openseo OSS: verify your site to see which links actually count bonus: Ship one free tool people want to link to most founders pay a SEO agency for half of this.
13
4
17
650
Cloudflare tiene de código abierto su Skill para encontrar vulnerabilidades en tu proyecto. Lo que hace paso a paso: ① Revisa tu código ② Encuentra todos los bugs ③ Envía lo que encuentra a un agente de IA ④ Crea el informe, verifica y escribe un reporte final Pasos para instalarlo: github.com/cloudflare/securi… Y luego le pides a la IA: "Audita la seguridad de este proyecto"
41
330
5
2,966
119,637