|
|
|
A new default workhorse at $1.50/M input, a budget tier at $0.30/M, and a classified security model that outperformed Claude Opus 4.6 on V8 vulnerability discovery, per Google's blog.
|
|
JULY 22ND, 2026 · BY JONATHAN HILDEBRANDT
|
|
Google and DeepMind shipped three Gemini models on July 21, per the Google Blog. Gemini 3.6 Flash runs $1.50/M input and $7.50/M output and becomes the default workhorse. A cheaper 3.5 Flash-Lite lands at $0.30/$2.50 for high-volume jobs. And a government-only 3.5 Flash Cyber, wired into Google's CodeMender agent, found 55 confirmed V8 JavaScript vulnerabilities versus 36 for Claude Opus 4.6, per Google DeepMind. Gemini 3.5 Pro slipped a third time; Google teased Gemini 4.
Why it matters. The $1.20/M input gap between 3.6 Flash and Flash-Lite is wide enough to benchmark before you route high-volume pipelines, and 3.6 Flash's fewer-tokens-per-task claim means real per-request cost sits below the headline price, per Google. Run a few production-representative calls before you lock a tier.
|
|
See Google's launch post →
|
|
|
|
•
|
Kimi K3 open weights land July 27. The model is 2.8 trillion parameters and 594 GB in BF16, needing 4 to 8 H100s to run locally, per CryptoBriefing; the hosted API paused new subscriptions July 19 to 20 when GPU demand outran supply. Self-hosting the one open-weight model leading agentic coding benchmarks beats renting a rate-limited API. Read our review →
|
|
•
|
A Codex allowlist entry became remote code execution. Pillar Security's GitPwned writeup shows OpenAI's Codex CLI classified "git show" as read-only, so a write-path option injected a malicious externalDiff hook and escaped the sandbox. OpenAI confirmed CVSS 8.6 and patched in Codex v0.95.0. Update your CLI now. Read the writeup →
|
|
•
|
EU AI Act transparency rules hit August 2. Article 50 mandates chatbot disclosure and machine-readable watermarks on AI-generated text, with penalties up to EUR 15M or 3% of global annual turnover. Robust text watermarking that meets the spec does not exist as deployed technology. See the EC guidelines →
|
|
•
|
Anthropic outspent Nvidia on federal lobbying. Anthropic reported $1.97M in Q2, up 26% from Q1, versus Nvidia's $1.9M and OpenAI's $1.2M; the AI sector set a $4.3M single-quarter record. The spend targets export controls and cybersecurity rules that decide which models you can legally deploy and where. Read the breakdown →
|
|
•
|
Beehiiv. Newsletter platform built for creators who want a paid list and monetization baked in, not bolted on. Try Beehiiv →
|
|
•
|
Make. Visual automation for AI pipelines, and the monetizable alternative when n8n or Zapier lock you out of the affiliate math. Try Make →
|
|
•
|
Cloudways. Managed cloud hosting to stand up self-hosted models and agents (Open WebUI, n8n, a Kimi K3 endpoint) in minutes. Try Cloudways →
|
Kimi K3 is the open model to switch to before July 27
JULY 17TH · PONDERO REVIEW
Our review runs the sourced pricing math: Kimi K3 leads agentic coding benchmarks and undercuts Fable 5 on long-context loops. It maps where it wins, where it loses, and which teams should move before the weights drop.
|
Cursor vs Copilot, updated for July 2026
JULY 18TH · PONDERO COMPARISON
Refreshed with the SpaceX case study, the Kimi K2.7 model option, and the new browser-agent capabilities, plus a clear pick for solo devs and for teams.
|
|
|
|
|
Jonathan Hildebrandt
Co-founder and primary operator of Pondero. Writes the Pondero Brief.
|
|
|
Affiliate disclosure
·
Unsubscribe
·
Manage preferences
Pondero earns commissions on some links. This does not affect our editorial picks.
|
|
