Your dependencies moved. So did the deadline.

ayo denk n nathan

OpenAI cut Luna 80%. GitHub and AWS both closed products the same day. And the August 2 deadline you have been racing quietly moved to 2027 while nobody was looking. (about 6 minutes)

Here's the signal today: 🛸

🏷️ GPT-5.6 Luna drops 80%, and most pipelines will not re-test

🏷️ Thinking Machines open-weights a model that beats its own flagship

🔌 GitHub Models shut down with no grandfathering

🔌 AWS closes Bedrock Agents Classic to new accounts

🔌 Zuckerberg still will not say what Meta capacity to sell

🛡️ The August 2 high-risk deadline moved to 2027 and 2028

🛡️ FTC sues Hims and Hers over health-data trackers

🛡️ German court rules against Suno on training data

🏷️ Agent Watch

GPT-5.6 Luna drops 80%. On July 30, OpenAI cut GPT-5.6 API prices. Luna now runs $0.20 per million input tokens and $1.20 per million output, down 80%. Terra fell 20% to $2/$12. Sol picked up a Fast mode: up to 2.5x the speed for 2x the price, same intelligence. Simon Willison checked the numbers against the pricing pages rather than the announcement, and notes OpenAI credits GPT-5.6 Sol with helping optimize its own load balancing and inference kernels. Latent Space puts the four-month move at roughly 13x more intelligence per dollar. The trap: a price cut is only a discount if you can take it. Teams that hardcoded a model string six months ago will read this, feel good, and change nothing. 💰 Money play. Point staging at Luna today, rerun your evals, and save the score sheet before you touch production.

A 276B open-weight model that beats its own 975B flagship. On July 30, Thinking Machines Lab published Inkling-Small under Apache 2.0 with full weights on Hugging Face: 276B total parameters, 12B active, against the flagship Inkling's 975B and 41B. The vendor's own card puts it ahead of that flagship on Humanity's Last Exam (31.6% against 29.7%) and SWE-bench Verified (80.2% against 77.6%). Treat those as self-reported, one day old, and likely run at maximum reasoning effort; Artificial Analysis scored the pair 40 and 41 on its own index, which is a much narrower gap. The trap: "12B active" reads like a laptop model. Serving is sized by TOTAL parameters, and the model card asks roughly 180GB for the NVFP4 checkpoint. That is one B300, not your MacBook. 💰 Money play. If you already rent a B300 or a pair of H200s, pull the NVFP4 weights and run your own eval set against your current hosted tier before you believe either number.

AWS closed the door on Bedrock Agents Classic. As of July 30, 2026, Amazon Bedrock Agents (the 2023 product, now "Classic") sits in maintenance mode and is no longer open to new customers. The model catalog is frozen at that date, so anything released after it appears only in AgentCore. The trap is a quiet one: existing infrastructure-as-code keeps working for allowlisted accounts, so your pipeline stays green while the identical template fails in any account without prior usage. 💰 Money play. Try provisioning your Bedrock stack in a fresh AWS account this week. If it fails, you found your migration deadline before a client found it for you.

🔌 The Build

GitHub Models is gone. The playground, model catalog, inference API and bring-your-own-key all stopped on July 30, with no grandfathering: GitHub's changelog says it is "no longer available to any customer, including existing customers with active usage." The shutdown was announced July 1 with brownouts on July 16 and July 23. Microsoft Foundry takes the catalog, Copilot takes the GitHub-native work. The trap: this one was pre-announced for thirty days, so if it broke something of yours overnight, the deprecation is not the failure. Nothing in your stack told you a dependency had a shutdown date. 💰 Money play. Repoint anything calling that endpoint at a provider API directly, then go find the other three services you depend on that have already posted a sunset notice you have not read.

Zuck's capacity dilemma. Zuckerberg laid out the question investors keep asking: sell AI capacity to outsiders, or keep it for Meta's own products. Bloomberg notes he still has not filled in the details of an actual AI cloud business. If you design on open-weight Meta models, that unanswered question is really a supplier question: it decides whether your inference has a second source or exactly one. The trap: building a roadmap around SKUs that do not exist yet, then finding your architecture assumed a price nobody ever quoted. 💰 Money play. Until Meta publishes SKUs and SLAs, price your compute against AWS, Azure, and GCP, and treat open-weight Meta models as the thing you can actually ship on.

🛡️ The Shield

FTC sues Hims and Hers. Shares fell about 10% after the FTC sued over alleged sharing of patient health data with Meta and Snap through website trackers, plus billing and cancellation friction. TechCrunch puts the case on sexual wellness and mental health prescriptions. The trap: teams audit the checkout page and forget the tracker fires on every authenticated route behind it. 💰 Money play. Inventory every Meta, Snap, and Google tag on logged-in pages, kill the ones firing before consent, then click your own cancel flow end to end.

The August 2 deadline you are racing is mostly not the deadline. Regulation (EU) 2026/1744, the Digital Omnibus on AI, entered into force on July 27 and amends Article 113 of the AI Act. The Chapter III high-risk obligations everyone has been sprinting at moved: standalone high-risk systems under Annex III now apply from December 2, 2027, and embedded-product systems under Annex I from August 2, 2028. What still lands on August 2, 2026 is Article 50 transparency, the Article 5 prohibitions, and AI literacy in softened form. The trap runs in both directions. August 2 is not cancelled, and the duty that survived is the disclosure one most teams deprioritized because it looked like the easy half. 💰 Money play. If you ship generative output into the EU, your Sunday obligation is disclosure and marking, not a conformity assessment. Providers already on the market before August 2 get until December 2, 2026 for machine-readable marking specifically.

Suno loses in Germany. The Munich Regional Court ruled for GEMA on July 31 that Suno's training and German-hosted models infringed copyright, ordering it to stop training on the works, disclose infringement revenue, and pay damages still to be set. Case 42 O 763/25, not yet enforceable. 💰 Money play. If you ship generated audio or media, write your training-data provenance down now, while it is still a document and not a disclosure order.

🎓 Level Up

Four of today's stories are one story: something you build on moved, and you did not pick the timing. Run a portability audit this week.

1. grep -rn your codebase for hardcoded model strings (gpt-, claude-, gemini-). Every hit outside a config file is a migration ticket with no due date yet.

2. Same grep for vendor SDK imports. Count the files that would change if you switched providers. That count is your switching cost, and you have never priced it.

3. Move every model id into one config module behind an environment variable.

4. Write one line each for two questions: what you do if your current model doubles in price, and what you do if it is deprecated with 90 days notice.

5. Provision your stack into a clean cloud account. Whatever fails is a dependency you did not know you had.

You are not doing this because you plan to switch. You are doing it so the next price cut is a config change instead of a sprint.

Reply and tell us what your grep turned up. Forward this to someone still running a model string from six months ago: https://nomadsignal.ai

See you tomorrow,
Chase & Kobe 🐕

One tap. Feedback goes straight to the desk.