August 26, and it is not a rename

Welcome back technologists🫡 

Here's the signal today: 🛸 
👨‍💻🐕️ 

• 🗓️ OpenAI's Assistants API stops answering August 26

• 🔧 The MCP spec deleted sessions entirely

• 💸 Sonnet 5's intro pricing expires August 31

• 🚨 Six critical SQLite CVEs were AI slop

• 🛡️ EU chatbot disclosure went enforceable August 2

🗓️ August 26, and it is not a rename

OpenAI's deprecation page puts a hard date on it: August 26, 2026. After that, /v1/assistants, /v1/threads and /v1/threads/runs return errors. No degraded mode, no read-only grace, no quiet extension promised.

Here is the part that will cost people their weekend. Most are filing this as an endpoint rename. It is an architecture change: Assistants become Prompts, Threads become Conversations, Runs become Responses, and the state model moves with them. A migration that swaps the URL and redeploys does not compile.

This bites a solo operator harder than a platform team, because you do not own all the code that calls it. Client work from eighteen months ago is still running on beta.threads somewhere, and nobody has told that client.

💰 Money play. Tonight, run grep -rE "beta\.assistants|/v1/threads|beta\.threads" across every repo you have shipped, including the archived client ones. Every hit is a migration you invoice this week instead of an emergency you eat on August 25.

🔧 Agent Watch

The MCP spec deleted sessions. The 2026-07-28 revision removed the Mcp-Session-Id header and the initialize handshake, and deprecated Roots, Sampling, Logging and HTTP+SSE transport. The trap is reading that list and deciding you can wait. Going stateless is not a deprecation, it is structural: a server holding per-call state behind a session id breaks all at once, because cross-call state now has to be explicit server-minted handles passed as tool arguments. 💰 Money play. If you ship an MCP server, grep it for Mcp-Session-Id, sampling/createMessage and roots/list today. While you are in there, make tools/list return a deterministic order. The spec now says outright that it lifts prompt cache hit rates.

Claude Code shipped permission-bypass fixes. Version 2.1.221 patches multiple permission-check bypasses in Bash and PowerShell handling. Read past the new Focus view: the buried lede is that your allowlist had holes, which matters most to the person running an agent semi-autonomously while doing something else. 💰 Money play. Update before your next session, then run the new prompt-audit subcommand in the claude-api skill over your committed agent prompts.

Your agent can read your inbox now. Cursor's August 3 changelog shipped Google Workspace plugins, handing agents Gmail, Drive and Calendar. That is a prompt-injection surface pointed straight at the least trusted text you own, which is mail other people send you. 💰 Money play. If you turn it on, do not give that same agent code execution in the same session.

💸 The Build

Sonnet 5's intro pricing ends August 31. It goes from $2/$10 per million tokens to $3/$15 on September 1, batch in step. The compounding is what people miss: the same page notes the 4.7-and-later tokenizer emits roughly 30 percent more tokens for identical text, so a September bill absorbs the rate rise and a token-count rise at once. 💰 Money play. Pull your August token counts from the console, multiply by 1.5, and decide per route before September 1 whether each lane stays on Sonnet, drops to Haiku, or moves to Batch.

Cursor deleted the receipt. On July 31 Cursor removed per-request dollar amounts from the self-serve usage page; the CSV cost column now reads zero and the API stopped returning cost. A Cursor staffer called it "deliberate design," noting the displayed dollar amounts "were often higher amounts than the user's plan cost." Three days later Cursor said its cloud agents are 20 to 30 percent more token efficient. The claim may be true. The instrument you would check it with left the same week. 💰 Money play. Export your usage CSVs today while the old exports still carry data, then move budget tracking to Dashboard, then Spending, and set a hard on-demand cap tonight.

DeepSeek V4-Flash is cheap, with a clock attached. It runs $0.14 per million in against $0.28 out. The pricing page also pre-announces 2x peak-hour rates for 09:00 to 12:00 and 14:00 to 18:00 Beijing time, date pending. Convert before you celebrate: that second window is roughly 11pm to 3am Pacific, exactly when North American overnight crons run. 💰 Money play. A/B it on your ugliest high-volume, low-stakes lane this week, and write your cron times next to those windows now rather than discovering the overlap on an invoice.

🚨 Six CVEs that never existed

JFrog took apart six "critical" SQLite CVEs rated 7.5 to 9.8, all filed from a freshly created GitHub account, and found AI-generated advisories citing functions and line numbers absent from the versions they name. The proof-of-concept payloads did not run. Red Hat had already cut one from a 10.0 to a 7.6, and a wider audit of 55 advisories from that same account found 54 completely fabricated and exactly one real bug.

This lands on your desk rather than a security team's because your scanner cannot tell the difference. A solo builder who sees SQLITE CVSS 9.8 CRITICAL in a Dependabot alert at 11pm does the emergency bump and breaks something real to fix something that was never there. The slop is not in your code. It is in the feed you trust to tell you your code is broken.

💰 Money play. Before your next emergency dependency upgrade, spend ninety seconds on who filed the CVE and whether the proof of concept runs in a sandbox. A new account plus a PoC that does not execute means park it until a maintainer speaks.

🛡️ The Shield

EU chatbot disclosure went enforceable on August 2. Article 50 now binds: users must be told they are talking to an AI, and synthetic audio, image and video must be labelled. Fines reach 15 million euros or 3 percent of worldwide turnover. Both careless reads are circulating. One panics that high-risk obligations are live; those moved to 2027 and 2028. The other relaxes about the December grace period, which covers only machine-readable marking, only for systems already on the market. Chatbot disclosure got no grace period at all. 💰 Money play. Open your production chat widget and read its first message. If an EU user can reach it and it does not say they are talking to an AI, ship that line today.

The safety layer became evidence. Tennessee plaintiffs are suing xAI over sexualized images they say Grok made from ordinary photos; the July 7 amended complaint added Stability AI. Read past the generation to the reporting: the filing cites an NCMEC finding that 90 percent of xAI's CyberTipline reports were not actionable, arriving without user information, without IP addresses, sometimes without the images. These are allegations in a civil complaint and nothing is proven. Running the other way, CDT argues the FTC's AI policy statement would treat ordinary safety work as "deceptive steering," chilling the safeguards the lawsuit says were missing. 💰 Money play. File one test abuse report through your own product end to end this week and look at what lands. A step that files something nobody can act on is worse than none: it creates a record saying you knew.

An entire ad agency in the palm of your hand.

Your next campaign needs a dozen fresh ad variations by Friday. Your agency quotes two weeks and a five-figure invoice. Your in-house designers are already buried under this quarter's requests.

Hightouch Ad Studio fixes that. It reads your brand guidelines, your best-performing creative, and your product catalog, then generates on-brand ads your team can ship the same afternoon. You review and approve every asset before it goes live, so quality holds.

Growth teams use it to build variations for every audience, test more of them, and stop rationing creative because production got expensive. The work that once needed a full agency retainer now runs inside your own workflow, at your pace and under your direction.

You direct the work while Ad Studio handles production, and your designers get their week back.

Start your first campaign today

🎓 Level Up

The deprecation grep

Four of today's stories are the same chore in different hats. The whole week in one paste:

# every dying identifier in this issue, one sweep
grep -rEn \
  "beta\.assistants|/v1/threads|beta\.threads" \      # dies Aug 26
  --include="*.{py,ts,js,tsx,jsx,rb,go}" .

grep -rEn "imagen-4\.0" .                              # dies Aug 17
grep -rEn "Mcp-Session-Id|sampling/createMessage|roots/list" .
grep -rEn "temperature|top_p|top_k" .                  # 400s on current models

That last line is the one people forget. Anthropic removed sampling parameters on its newer models, and Google deprecated temperature, top_p and top_k on the latest Gemini models on July 21. Sending them is not a stylistic choice now, it is a 400. A wrapper that passes them through unconditionally is already broken against half the models you might route to next.

I cut two stories from this issue today because their only source was an aggregator, which is the same discipline in a different costume: the cheap check now beats the expensive surprise later.

Reply and tell me which of these four greps came back dirty. I read every one.

Forward this to one person still running on beta.threads: nomadsignal.ai

See you tomorrow,
Chase & Kobe 👨‍💻🐕