Eng Coffee Sips EP.004

New Model, Who Dis?

Eng Coffee Sips is a short factual based newsletter for articles that might interest engineers or engineering leaders, not just AI news. Short and to the point to skim over your morning coffee.

In this edition

By Pete Shima · September 7, 2026

Today’s edition is 1502 words, a 7-minute read.

1. 1/

   OpenAI released GPT-6 Astra Astra shipped late last week and there have been a lot of impressive results from this so far. Definitely worth trying out.

   My Take

   The harness and model wars continue. I have had a pretty crappy time with codex. Most recently I left astra to work on something, it span out of control didn't do anything and ate all my weekly tokens. Frontier Model.

   Snapshot from X

   GPT-6 Astra is here.

   We hope it will begin to enable a new generation of entrepreneurship, scientific discovery, and building.

   We believe it is the best model in the world for computer use, professional work, science, coding, cybersecurity, and more.

   It took us some extra time to ensure that we could meet the safety and alignment standards required for this capability level, but we think you’ll find it worth the wait.

   It scores 98% on FrontierMath Tier 4, 99.9% on ARC-AGI 3, and 100% on ExploitBench.

   ( OpenAI / VentureBeat / @sama )

2. 2/

   Anthropic shipped Fable 5.1 two days before Astra Fable 5.1 and Mythos 5.1 landed September 1.

   My Take

   Fable continues to be pretty good, but dang can it eat through tokens like no ones business. Maxed my plan in a day but luckily got the free token reset on Friday.

   Snapshot from X

   Fable 5.1 is our best model yet for coding, data analysis, computer use, design, presentations, Tag, and the hardest long-running agentic work.

   This model is a pleasure to work with, and I've been using it for everything.

   ( SiliconANGLE / Vellum / @bcherny )

3. 3/

   Anthropic reset everyone’s Claude usage meters the day Fable 5.1 shipped On September 1, timed to the Fable 5.1 launch, Anthropic flushed 5-hour and weekly limits for all users. Max subscribers opened usage settings to find the meters at zero.

   My Take

   The token usage drama is pretty funny to me.

   ( Tech Bytes / explainx.ai / @austinshandle )

4. 4/

   Five Fable 5.1 optimizations you can use now 1. Set effort to 'low' 2. Run '/claude-api cost-optimize' 3. Run 'claude-api prompt-audit' 4. Change effort mid-conversation w/o cache hit 5. Update Fable 5.1 API config w/ 'claude-api migrate'

   My Take

   If you find yourself with a lot of skills or a lot of custom items in Claude, try doing a cost optimization or a prompt audit to "reset" your setup.

   ( @daniel_mac8 )

5. 5/

   Is Human Code Review Dead? Gergely Orosz asked engineers on September 6 whether they had stopped reviewing code, drawing over 300 replies. Duckbill Group had already answered: CEO Mike Julian says the firm mostly dropped review after hitting 60 open PRs across a team of five. Merged PRs then went from 353 to 684, and same-hour merges from 28% to 45%.

   My Take

   No one likes doing code reviews. Humans simply won't be able to keep up with the increased rate of AI changes so there needs to be a solution. There are a lot of interesting ideas and approaches in the replies. Intercom (now Fin) also posted a while back about how they are doing it which is similar to some of the mentions.

   Snapshot from X

   If you’ve stopped doing human code reviews: how is it going?

   If you’re still doing them: how are you able to keep up with the explosion of code + PRs, and how do you prioritise what to review “properly”?

   ( @GergelyOrosz / @mikejulian / Ref )

6. 6/

   A two-agent “Manager Loop” gets Astra through long-horizon builds Matt Shumer built a recreation of Manhattan with GPT6 Astra which is pretty impressive.

   My Take

   Big or challenging projects are still hard to one shot but some of these prompt techniques really do help a lot. I am still using a modified version of the claud of duty prompt which works surprisingly well in claude but terrible in codex.

   ( @mattshumer_ )

7. 7/

   Spotify cut Claude Code token usage 90% by blocking the expensive model from reading files Two cheap worker modes on Spotify’s Portal platform handle bulk file reads and repetitive code generation, both on Gemini 2.5 Flash. A hook blocks any Read over 350 lines and reroutes it; a second catches attempts to dodge it with cat, head or tail. Written instructions were ignored until the block was enforced.

   ( Techzine / @undefinedKi )

8. 8/

   RSA-260 was factored after 35 years, and nobody has said how Cognition engineer Eric Lu posted a 130-digit factor on September 3 with nothing but the words “divides RSA-260”. The 862-bit number displaces RSA-250, the record since 2020, as the largest factored with a general-purpose algorithm. As of September 4 Lu had disclosed no algorithm, software, hardware or running time. RSA-2048 is unaffected.

   ( John D. Cook / Scientific American / @AI_in_the_AM )

9. 9/

   Directing video models through 3D instead of re-prompting them Daniel Vogel has Astra write Python that generates and animates a .usda scene, then renders it in Blender. Words direct; the 3D layer handles geometry, camera, motion and composition, and locks them in rather than re-rolling every generation. He expects much agent-driven video to pass through 3D whether users see that layer or not.

   My Take

   Vogel with another banger of a short post. I need to try usd/usda specs.

   ( @EpicVogel )

10. 10/

    GitHub Next built a multiplayer planning environment called Chopin Chopin is a real-time, shared space for getting a team aligned on a plan before handing it to agents. GitHub Next says the planning tools it uses today feel cumbersome, static and lonely, and drown people in walls of text. It is an early prototype and open source.

    My Take

    Did someone say software factory? Well not quite, it's just planning but these multiplayer agent environments keep popping up, and for a reason.

    ( GitHub Next / @GitHubNext )

11. 11/

    Headroom compresses agent context before it reaches the model Headroom sits between an agent and the LLM, reversibly compressing tool output, logs, JSON and RAG chunks while the originals stay local. The README claims 20% fewer tokens for coding agents and 60–95% for JSON. Creator Tejas Chopra has left Netflix to work on it full time.

    My Take

    I haven't used this yet, and don't have a solid reason to yet but I think this approach is interesting particularly when it comes to SRE type of work.

    ( GitHub / Open Source For You / @BharukaShraddha )

12. 12/

    NVIDIA’s Personal AI Router pools the spare machines on your network PAIR discovers participating machines on a local network and routes each inference request to whichever node has capacity, behind Ollama- and OpenAI-compatible endpoints. It does not shard a model or split one request. It distributes separate requests, which suits multi-agent work rather than running one oversized model.

    My Take

    There are a lot of multi agent managers on the scene these days and a lot of ways to manage remote or cloud but this feels like it might be a big improvement to utilizing hardware at home.

    ( GitHub / @techNmak )

13. 13/

    OpenDesign is an open-source answer to Claude Design It turns coding agents into a design engine, reading brand rules from a DESIGN.md and producing prototypes, landing pages, dashboards, decks, images and video as real exportable files. More than 25 agents are supported, it runs local-first, and it is Apache 2.0. The README pitches it directly against Anthropic’s closed tool.

    My Take

    Claude Design is pretty incredible. With my effort to get off of Claude and onto Codex last week (which failed) the question is what do you do use as a Claude Design replacement? This looks promising, but for now I'm sticking with Claude.

    ( GitHub )

14. 14/

    Learn Inference is a free interactive course on serving models in production Seven chapters run from inference fundamentals through the model, hardware and software layers to production deployment, with interactive simulators, a glossary and per-topic reading lists. It is a companion to Philip Kiely’s Inference Engineering, published by Baseten.

    My Take

    I have been seeing a lot of learning sites pop up and I love it! Using AI to generate better ways to learn and consume information I think is a great use of the tech.

    ( Learn Inference )

15. 15/

    Diiverge is a persistent, infinite point-and-click adventure Charlie Clark launched it on September 2. Every picture is a fork: click something in it, decide what happens, and the world grows in that direction.

    My Take

    This is a pretty neat JIT choose your own adventure game. The neat thing is also the graph you can view for each generated timeline behind it. Kind of takes choose your own adventure to the next level.

    ( Diiverge / @charliie )