IMPLICATOR
From San Francisco
1 |
The Editorial |
Good morning.
Google priced Gemini 3.7 Flash at $0.75 per million input tokens, half what Gemini 3.6 Flash launched at. The rate doubles January 1. Alibaba published Qwen3.8's 2.4-trillion-parameter weights for anyone to download, and Nvidia's first serving test used a 72-GPU rack. Either way, you rent the rack.
Below: how long Google's discount lasts, what open weights actually cost to serve, and the tool that promises to strip Claude's watermark without proving it can.
Stay curious,
Marcus Schuler
2 |
The Big Story |
Google released Gemini 3.7 Flash on Thursday at $0.75 per million input tokens and $3.75 per million output tokens, half what Gemini 3.6 Flash launched at.
The introductory rates run through December 31. On January 1 they double.
Coding scores moved with the price. Gemini 3.7 Flash reached 65.3% on DeepSWE v1.1, up from 49.0%, and 43.6% on FrontierCode 1.1 Main against 34.4% for its predecessor. Google's own comparison table still puts GPT-5.6 Terra ahead on DeepSWE at 69.6%, on both Terminal-bench versions and on OSWorld-2.0.
Why This Matters:
- Teams testing Flash on production code get a five-month window at half rate before the January 1 increase changes their unit economics.
- A cheaper token does not lower cost per finished task if agents need more retries or human correction.
3 |
Also Today |
Alibaba published the Qwen3.8 checkpoint on Aug. 12 with 2.4 trillion parameters, free for anyone to download.
Nvidia's day-zero serving test ran it on one GB300 NVL72, a rack of 72 Blackwell Ultra GPUs. DeepSeek answered with a 1.6-trillion-parameter V4 Pro and an MIT-licensed agent harness. Its peak output price rises from $0.87 to $3.96 per million tokens at 16:00 UTC on Aug. 16, leaving developers to price their own hardware against the new managed rate.
4 |
The Outside Read |
Working Copy explains why an AI-detector score says little about whether a piece deserves a reader's trust.
Nick Hagar proposes judging writing by utility, prose and trust, then argues that a classifier cannot reveal where ideas came from, how judgment was exercised or who accepts responsibility for the claims.
5 |
The One Number |
6 |
Today's Headlines |
- Databricks closed a $5 billion round at a $190 billion valuation led by Coatue, with revenue run-rate crossing $7 billion on more than 80% second-quarter growth.
- IBM will build a dedicated OpenAI practice staffed by thousands of consultants and engineers, embedding GPT-5.6, Codex and ChatGPT Work into IBM Consulting Advantage.
- Anthropic investors are modeling an October IPO at $2 trillion or more, while Ramp's July sample puts Fable 5 at about 75% of GPT-5.6 Sol spending.
- Microsoft is retiring its weaker AI features and merging its separate Copilot apps, narrowing the roadmap for users and developers who built on them.
- Google cut the Pixel 11 Pro's base RAM to 12GB from 16GB, with devices chief Rick Osterloh blaming a memory shortage as producers shift capacity to AI servers.
7 |
The 5-Minute Skill |
Put denominators back into the board pack.
Board slides often report growth rates without the base, period, or population needed to interpret them. Ask the model to locate claims that look precise but cannot yet support a decision.
Your raw input: the board memo or slide text, including footnotes, chart labels, reporting periods, and source notes.
The prompt:
Why this works: the fixed fields make the model inspect the structure of each number instead of summarizing the deck. The ranking focuses discussion on omissions that could change a decision.
What to use: Claude Opus 4.7 handles long board packs well. GPT-5.6 is a strong fallback for shorter material.
8 |
What To Watch Next |
SAT 8/15 |
AI: IJCAI-ECAI 2026 starts three days of workshops and tutorials in Bremen before the main conference opens Aug. 18. |
SUN 8/16 |
Fair: Pebble Beach Concours organizers open the 75th show field to spectators at 5:30 a.m. PT, with judging starting at 8 a.m. |
TUE 8/18 |
Finance: The Bureau of Labor Statistics reports July U.S. import and export price indexes at 8:30 a.m. ET. |
TUE 8/18 |
Politics: Alaska and Florida hold statewide primary elections, with Florida polls open from 7 a.m. to 7 p.m. local time. |
WED 8/19 |
Policy: The Federal Reserve releases minutes from its July 28-29 FOMC meeting at 2 p.m. ET. |
9 |
AI Image of the Day |
10 |
The Rausschmeisser* |
Guillaume Meyer's watermarks-remover repository reached 4,102 GitHub stars by Aug. 13, two days after he created it. Its documentation states the limit plainly: "Until vendors ship public detectors and keys, no tool can honestly certify 'this fails the official check.'" (GitHub, August 13, 2026)
Our take: Meyer built two tools and shipped them under one name. The first strips zero-width characters and C2PA metadata, and a diff of the file proves it worked. The second asks a model to rewrite your prose until Claude's statistical mark is hopefully gone. Nothing proves that one worked at all.
Anthropic has not published a detector, so nobody can grade the homework, including Meyer, who says exactly that in his own README. Four thousand people starred a repository whose headline feature is untestable by design. What the rewrite reliably delivers is flatter prose that may still carry the mark.
Marcus Schuler
Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai