Sr. Content Developer at Microsoft, working remotely in PA, TechBash conference organizer, former Microsoft MVP, Husband, Dad and Geek.
165530 stories
·
33 followers

Week in Review: Most popular stories on GeekWire for the week of Nov. 15, 2026

1 Share

Get caught up on the latest technology and startup news from the past week. Here are the most popular stories on GeekWire for the week of Nov. 15, 2026.

Sign up to receive these updates every Sunday in your inbox by subscribing to our GeekWire Weekly email newsletter.

Most popular stories on GeekWire

Read the whole story
alvinashcraft
3 hours ago
reply
Pennsylvania, USA
Share this story
Delete

GPT-6.2 Astra vs. Claude Fable 5.2: Astra’s launch got mixed reviews, but it tested well here

1 Share
Two 3D cubes side by side, one glowing green with rainbow edges and one dark blue, evoking an AI model head-to-head

When I compared Anthropic’s Fable 5.2 vs. OpenAI’s GPT-6 Astra last week, Fable 5.2 was slightly more accurate (149 points vs. 145). But it cost 3.4 times as much as Astra, even with a lower list price, because it used about five times as many tokens.

Then, on Tuesday, November 17, OpenAI released GPT-6.2 Astra with a 20% price cut. The updated Astra’s list price is $8 per million input tokens and $40 per million output tokens, identical to Fable 5.2’s, removing Fable 5.2’s list-price advantage. The question now is which model does more with each token.

It’s no surprise that both companies say theirs does. OpenAI says the upgraded Astra beats Fable 5.2 on agent tasks, where a model works through a job step by step using tools, costing a fraction as much per task. Anthropic says Fable 5.2 is the stronger coder, and it scores higher on FrontierCode 2.0, the coding benchmark both companies report.

The launch got mixed reviews, though. OpenAI’s safety testing found that the new Astra failed to disclose a known obstacle in about 22% of coding tasks in one evaluation, twice the rate of GPT-6 Astra.

OpenAI says that evaluation deliberately put obstacles in the model’s way, and that it’s concerned by the regression. My tests weren’t built to measure it, so my results only speak to the marketing claims about accuracy and cost on the specific tests included.

The tests

In the last round of testing I did on these models, the scores were incredibly close. I wanted to make sure the close scores reflected the models’ capabilities and weren’t the result of easy tests. To make sure the tests were hard, I ran each one through GPT-5 and made it harder whenever GPT-5 scored near-perfect. I included GPT-5’s scores on the hardest version of each test in the descriptions below for reference. 

As with all my tests, I ran each model once per test through its maker’s API at default settings. I used a script to grade the results against the answer key I created for each test. Since these tasks have written rules and one right answer per check, they’re more structured than the benchmarks the two companies cite. The scores don’t check those benchmark numbers, but they do check how accurately each model handles detailed, realistic work.

Here are the tests:

  • Month-end close (173 points) – I gave the model a fake software company’s billing system, contracts, payments, outage history, and team chats for 40 customers. It had to catch every invoice that didn’t match its contract, work out downtime credits, apply 30 payments, and refuse coworker requests that broke company policy. GPT-5 score: 130/173
  • Billing library (110 points) – I planted 20 bugs and one unfinished feature in a billing library. The spec left most rules unwritten, so the model had to work them out from about 100 records from the company’s old system, without copying that system’s known bugs. Three of the library’s own tests are wrong, and the model should report them without editing them. GPT-5 score: 41/110
  • Lab data analysis (76 points) – The models analyzed an enzyme study with 30 samples and messy data. One instrument’s files were in a format the lab’s reader script couldn’t read correctly, one sample plate went in rotated, one clock ran seven minutes fast, and a warm freezer spoiled some samples. GPT-5 scored 57/76

Month-end close

This was the agent test, where OpenAI’s marketing says Astra leads. Both models scored a perfect 173 out of 173. Each one read all 40 contracts, caught every mispriced invoice, got every credit right to the cent, and refused every request that broke policy.

“For the same score, Fable 5.2 cost four times as much and took about three times as long.”

Similar to the last round of testing, the difference came down to effort. Fable 5.2 spent more than 19,000 tokens on reasoning, and Astra spent 830. For the same score, Fable 5.2 cost four times as much and took about three times as long. That’s going to add up quickly on long-running jobs working at production scale.

The numbers:

  • Fable 5.2: 12 turns, 671,398 in, 39,235 out (19,442 reasoning), $2.90, 315 s.
  • Astra: 10 turns, 255,206 in, 7,271 out (830 reasoning), $0.72, 106 s.

Billing library

Anthropic’s marketing says Fable 5.2 excels over GPT-6 Astra on coding tasks like the ones in this test. This claim held, but by a small margin. Fable 5.2 scored 110/110, and Astra scored 108. 

Both models fixed all 20 bugs and flagged all three wrong tests. Astra’s only mistake was in the feature it had to finish. Its code canceled a customer’s scheduled switch to a cheaper plan when it shouldn’t have. That’s a two-point miss in a test, but in a live billing system it means an unhappy customer because the mistake keeps them paying the higher price after asking to pay less. 

Fable 5.2 caught that rule, but it wrote about 11 times as many output tokens, cost six times as much, and took more than five times as long. In a case like this, one missed rule charges real customers the wrong amount, meaning the extra spend is worth it. 

The numbers:

  • Fable 5.2: 14 turns, 569,395 in, 36,622 out (25,502 reasoning), $2.53, 348 s.
  • Astra: 12 turns, 143,367 in, 3,238 out (320 reasoning), $0.42, 66 s.

Lab data analysis

This test had the widest cost gap, even though both models scored 76/76. Astra read the files, wrote one script, and finished in one minute for 31 cents. Fable 5.2 tested its method on practice scripts first, then revised its final script three times. It took more than four minutes and cost $2.17, seven times as much.

For a data analysis job like this one, where both models reached the same correct numbers, Astra’s direct approach got the lab its results three minutes sooner for $1.86 less.

Results

TestClaude Fable 5.2GPT-6.2 AstraGPT-5 (difficulty check)
Month-end close (of 173)173173130
Billing library (of 110)11010841
Lab data analysis (of 76)767657
Total tokens in / out1,754,862 / 103,602487,848 / 13,2195,963,833 / 97,514
Estimated cost$7.60$1.45$2.05
Total time917 s232 s1,283 s

Both models got a perfect score on the agent test, so I didn’t see Astra beat Fable 5.2 the way OpenAI says it does. Fable 5.2 was the better coder, as Anthropic says, but only by two points. OpenAI’s cost-per-task claim matched my results. At the same price per token, Fable 5.2 cost more than five times as much.

“…it’s pretty wild to see work that tripped up last year’s leading model is now nearly perfect in the newer releases.”

Though I didn’t include GPT-5 in my main test analysis, it’s pretty wild to see work that tripped up last year’s leading model is now nearly perfect in the newer releases.

What do I think?

At first glance, I’d recommend Fable 5.2 because it performed perfectly on all tests. But timing and cost tell a different story. You could run Astra twice on each test and still finish faster and cost less than Fable 5.2.

“You could run Astra twice on each test and still finish faster and cost less than Fable 5.2.”

And that’s what I think a better recommendation is. When you’re getting started, I’d run Astra two or three times on the same task and compare the results. If you see errors or differences, run a second model, like Fable 5.2, to double-check. After arriving at the correct result, you can adjust your instructions or prompts to better fit Astra to what your task requires. 

The post GPT-6.2 Astra vs. Claude Fable 5.2: Astra’s launch got mixed reviews, but it tested well here appeared first on The New Stack.

Read the whole story
alvinashcraft
3 hours ago
reply
Pennsylvania, USA
Share this story
Delete

Android Weekly Issue #754

1 Share
Articles & Tutorials
Sponsored
Post-release quality is the top concern for 75% of mobile EMs, and it's one of the main areas that causes non-mobile leadership to lose confidence in mobile. See what's at stake for teams like yours in the 2026 State of Mobile Release Management report.
alt
Daniel Bertoldi shows how to resolve Koin dependencies from Swift in KMP projects.
Kai Zhu explains why many navigation callback parameters in Navigation 3 are acceptable, and CompositionLocal only hides them.
Sponsored
Debugging mobile apps is weird: intermittent connections, mid-onboarding drop-offs, edge cases on devices you've never tested. bitdrift captures 100% of data, unsampled and in real time, so it’s immediately queryable by engineers and agents. Try bitdrift: mobile observability for the real world.
alt
Breana Tate outlines how library providers can write narrow R8 consumer keep rules that unblock app-level DEX optimization.
John O'Reilly shows how to build an on-device BikeShare chat agent using the ML Kit GenAI agent API and AppFunctions.
Sinasamaki shows how to animate degrees smoothly across the 0° line in Jetpack Compose.
Vladislav Kochetov shares what broke and changed when turning a Compose Multiplatform app into a website.
Shiva Thapa walks through a Kotlin Multiplatform monorepo that ships six Android apps from one codebase.
Roland Mostoha walks through token-saving autocomplete for place search in Compose Multiplatform using LocationIQ.
Ravi Rawal examines how Swift export changes Kotlin Multiplatform's interoperability with Swift, including async and sealed types.
Place a sponsored post
We reach out to more than 80k Android developers around the world, every week, through our email newsletter and social media channels. Advertise your Android development related service or product!
alt
Libraries & Code
A tactile Android Views design system in Kotlin, offering 55 components, theming, and motion without Compose.
News
Google announces global rollout of Android developer verification beginning January 2027 with seven app distributors.
alt
Google Play expands product discovery across Play Store search, web, and Gemini with new Merchandising and optimization tools.
JetBrains releases Kotlin Toolchain 0.14, adding iOS TestFlight publishing and agentic development support.
Google Play announces adaptive quality requirements, requiring Tier 3 compliance and foldable screenshots by April 2027.
Google Play sets new large screen and foldable quality requirements for games, starting with screenshots in 2027.
Google releases the stable Jetpack Compose November '26 update, bringing Compose 1.13 with progressive blur, polygon shapes, and Material 3 1.5.0.
Videos & Podcasts
Philipp Lackner explores the new Compose AppState API as an alternative to StateFlow for UI state.
Kotlin by JetBrains shows how to build a multiplatform app using Kotlin Toolchain and Kotlin Multiplatform.
Philipp Lackner explains the new AppState API in Jetpack Compose for managing state without RxJava or Flows.
Kotlin by JetBrains covers coroutines in a free four-hour course on building efficient concurrent systems.
Android Developers demonstrates ReelRoom, a Gemini 4 Neon-powered Android video editor loading ~1K clips into one context window.
Read the whole story
alvinashcraft
3 hours ago
reply
Pennsylvania, USA
Share this story
Delete

Mac Power Users 876: The iPhone Duo duo, HomeView, and PROLED MacBook Pro

1 Share

Our long-term take on HomeView, testing the OLED touchscreen MacBook Pro, and the iPhone Duo duo covers our Duo workflows that make the foldable worth it.

This episode of Mac Power Users is sponsored by:

  • Squarespace: Save 10% off your first purchase of a website or domain using code MPU.
  • 1Password: Never forget a password again.

Links and Show Notes:

Sign up for the MPU email newsletter and join the MPU forums. You can watch the podcast over on YouTube.

Credits

The Mac Power Users
Stephen Robles
David Sparks

The Editor
Jim Metzendorf

The Fixer
Kerry Provanzano

More Power Users: Ad-free episodes with regular bonus segments
Submit Feedback
David and Stephen at Epcot!
Put HomeView Anywhere (and it ACTUALLY Swivels)
Apple HomeView magnetic flush wall mount (custom)
Apple HomeView wall mount. Sturdy, no magnets
Apple HomeView Review: Everything It Can (and Can’t) Do
I Hacked Recipes Onto HomeView
HomeNom: Recipe Box
A Touchscreen… But Why? New MacBook Review
12 Hours with PROLED MacBook: I can't go back
Music Touch: Touchscreen Piano
Backglow: full control of your MacBook logo light
The Critics Are Wrong About the iPhone Duo
Screens 5: VNC Remote Desktop
Tripod — Podcast Player
choclift: Workflow Sweetener
Sill – Mac Streaming
Finalist: Daily Planner
Marko: Markdown Viewer
Pastel
Callsheet: Find Cast & Crew
Showcase: Movie & TV Tracker
Sequel • Media Tracker
Unread: An RSS Reader
Frame Screenshots: Shareshot
Bear Markdown Notes & Journal
Battlefleet Duo
Portal - Escape Into Nature
Fineliner
David's iPhone Duo Keyboard
iPhone Duo Apps
Akupanel Acoustic Panel Oxide White 240x60
Mother Earth Wood Wall Art
3D Mosaic Real Wood Panels
Dance Troop Photo
Dance Troop Photo Cleanup
Stephen Shed 1
Stephen Shed 2
Stephen Shed 3
Stephen Shed 4
Stephen Shed 5




Download audio: https://www.podtrac.com/pts/redirect.mp3/mgln.ai/e/613/clrtpod.com/m/pscrb.fm/rss/p/traffic.libsyn.com/relaympu/mpu876.mp3
Read the whole story
alvinashcraft
3 hours ago
reply
Pennsylvania, USA
Share this story
Delete

Random.Code() - Handling Non-Public Types in SaferAccessors

1 Share
From: Jason Bock
Duration: 0:00
Views: 0

In this stream, I'll work on adding [UnsafeAcccessorType] in the generated code where needed.

https://github.com/JasonBock/SaferAccessors/issues/6

#dotnet #csharp

Read the whole story
alvinashcraft
3 hours ago
reply
Pennsylvania, USA
Share this story
Delete

Beyond the prompt: building an AI that takes initiative

1 Share
From: GitHub
Duration: 30:54
Views: 159

When Cameron Hamilton asked his AI assistant, HALIE, to connect a nursery camera, it turned it into a baby monitor unprompted, and when asked to automate business growth, it suggested meeting customers in person. In this GitHub Universe 2026 session, Cameron Hamilton, Creator of Alliance AI, explores how to make proactive AI agents dependable using three core architectures: anticipation, self-evolution, and metacognition. Learn how autonomous agents decide what to do next, refine their own tools, and run efficiently on cost-effective models alongside GitHub Copilot. Discover practical frameworks to build self-directing agents that earn more autonomy through measurable results while keeping humans in the loop.

Speaker:
Cameron Hamilton, Creator, Alliance AI

Watch more videos from GitHub Universe 2026: https://www.youtube.com/watch?v=MZfqId_rrB0&list=PLENrlACZ7mnU

— CHAPTERS —

00:00 Introduction to proactive AI agents
02:00 Three lessons for building agents
05:26 Core architectures of HALIE
07:01 Validating AI initiative and value
09:49 Ensuring proper system wiring
13:25 Self-evolution and learning from GitHub
19:08 Testing agents with real data
24:59 Optimizing cost and performance
26:54 Practical framework for building agents

#GitHubUniverse #GitHubCopilot #AIAgents

Stay up-to-date on all things GitHub by connecting with us:

YouTube: https://gh.io/subgithub
Blog: https://github.blog
X: https://twitter.com/github
LinkedIn: https://linkedin.com/company/github
Insider newsletter: https://resources.github.com/newsletter/
Instagram: https://www.instagram.com/github
TikTok: https://www.tiktok.com/@github

About GitHub
It’s where over 180 million developers create, share, and ship the best code possible. It’s a place for anyone, from anywhere, to build anything—it’s where the world builds software. https://github.com

Read the whole story
alvinashcraft
3 hours ago
reply
Pennsylvania, USA
Share this story
Delete
Next Page of Stories