Skip to main content
Promptway
Signal vs. Noise··3 min read

What Actually Happened This Week in AI, Stripped of the Hype

A ranked cut for people with jobs. Model wave residue, coding-agent pricing, export-control comebacks, and a noise list you can skip.

Iris Tanaka-Bell

Reviewed by Agnel Nieves

View as Markdown
Engraved illustration of a compass atop an open ledger with quill and magnifying glass, gold needle and rays on green background, framed by oak, fern and thistle.
Hero illustration and animation generated with Grok.

Weekly cut. Ranked by what changes an operator's next quarter, not what won the timeline.

1. The July model wave is becoming defaults, not demos

Claude Sonnet 5, GPT-5.6 (Sol / Terra / Luna), and Grok 4.5 are no longer "did you see." They are showing up in model pickers and procurement threads as ordinary options. The signal is not a crown. It is tiering and routing: cheap Luna-class work, mid Terra/Sonnet work, flagship Sol or Opus-class work. Teams that still have one corporate default with no eval will thrash quietly until September.

[The interesting story is configuration, not coronation.]

2. Coding-agent cost is a first-class product feature

Grok 4.5's Cursor co-training and aggressive token pricing force a rethink of "always use the smartest model." Claude Code versus Cursor remains a posture fight (delegate vs edit), not a pure IQ fight. If you buy tools this week, buy for the job shape. See our longer Stack notes rather than a launch thread.

3. Export control is a product dependency

Anthropic's Fable 5 / Mythos-class weights returning after the June control pause is a reminder with a calendar. If your roadmap assumes uninterrupted access to a single frontier tier, you own a policy risk. Put it on the risk register or stop pretending.

4. AEO still runs on earned media

Muck Rack's Generative Pulse line, about 84% of AI citations from earned media, paid near zero, has not been refuted by a clever schema trick. Owned-site hygiene remains necessary. It is not sufficient. Operators who only invest in on-site AEO are optimizing the warehouse and ignoring the trucks.

5. SaaSpocalypse hangover, not sequel

Six months after the February software scare, equities look less apocalyptic. Buyer language about agents and seat consolidation did not fully revert. Watch renewals and packaging, not just the sector ETF.

Noise kill list

  • "Best model of the week" listicles with forty logos and no task.
  • Fake MRR screenshots in reply-guy threads.
  • Benchmark charts without a task definition or variance.
  • "SaaS is dead" posts with no cancelled contract attached.
  • Prompt packs sold as secrets that are just roleplay preambles.

One action if you only have twenty minutes

Write down the three AI jobs you actually run. Assign each a model on purpose. If two jobs share a model only because "that is what we clicked first," fix one assignment this week.

Looking at next week

I will track: any forced change in GPT-5.6 access tiers, Cursor or Claude Code pricing moves, and whether another lab tries an export-shaped surprise. Predictions without a date are decoration. These have a date: next Monday's edition.

Sources

View as Markdown

Read next

Engraved illustration of a golden balance scale with a quill and key in its pans amid a pocket watch, envelope, hourglass, and map on a blue background with white botanical borders.

The Stack · 5 min read

Claude Code vs Cursor: I Gave Both the Same Client Task for Five Days

One assumes you are delegating. One assumes you are editing. I ran the same brief through both and named a winner for each job, including where Grok 4.5 changes the Cursor side of the ledger.