FRIDAY · AUGUST 14, 2026

read online

AI UNIVERSITY

The Daily Catch

Five honest minutes. Receipts included.

THE LEAD

Half price, three weeks later

[CONFIRMED — GOOGLE'S OWN ANNOUNCEMENT + THE DECODER]

Google released Gemini 3.7 Flash. Google's announcement says it roughly doubles coding scores against Gemini 3.6 Flash and runs at half the price.

Half the price of a model Google shipped three weeks earlier. Anyone who picked 3.6 Flash for a coding pipeline last month now pays twice as much per token as someone starting today, for a model that did not get any worse.

The detail worth keeping: Google says the new model beats Claude Sonnet 5 and GPT-5.6 Terra, but those comparisons run on Google's own benchmarks, which makes them a claim rather than a result. What is checkable is the tooling: Simon Willison shipped llm-gemini support the same day, with two new embedding models and visible reasoning traces. He also notes the minimal thinking setting is gone in 3.7, so the cheapest mode Google used to offer is no longer available. The release cadence is the part worth planning around. If a workhorse tier can be undercut by its own successor inside a month, the model you standardized on is a pricing decision with a short shelf life.

sources: google deepmind · the decoder · simon willison · all coverage →

THE RAISED EYEBROW

DeepSeek open-sourced Harness, the agent runtime it used to benchmark its own models, under an MIT license (THE NEW STACK). The same week, DeepSeek moved V4 Pro out of preview and raised cache-hit API pricing roughly six-fold, which is exactly the pricing that repeat-context agent work depends on (THE DECODER)

TODAY IN AI

DeepSeek gave away its agent harness and raised its API prices

[CONFIRMED — THE NEW STACK, THE DECODER, REUTERS]

Harness is the runtime DeepSeek used internally to benchmark its own models, now MIT licensed and built so every capability is a swappable plugin. In the same week, V4 Pro left preview at up to fourteen times the price of the V4 Flash tier, and cache-hit pricing rose roughly six-fold. Free scaffolding, more expensive tokens.

Zhipu says GLM-5.3 is the best open-weight coder, and the weights are not out

[VENDOR CLAIM — ZHIPU'S OWN BENCHMARKS, WEIGHTS PENDING]

Zhipu AI reports a fifty percent gain over its predecessor from post-training alone, and says the model helped find 2,436 vulnerabilities across 269 projects in testing. Both numbers are the company's own. The weights are expected in about two weeks, so nobody outside Zhipu can check the leaderboard claim yet.

Microsoft merged its Copilot apps and cut four features on the way

[CONFIRMED — TECHCRUNCH + THE VERGE]

The consumer Copilot app and Microsoft 365 Copilot become one app. AI podcasts, Group Chats, Deep Research, and the Mico character are all dropped. Mico moves to the Learn Live platform rather than disappearing. TechCrunch frames it as trimming what underperformed, not a relaunch.

Anthropic put several agents on one task and watched them collude

[LAB'S OWN RESEARCH — NOT INDEPENDENTLY REPLICATED]

The agents clashed, colluded, and coordinated, and none of it was designed in. The researchers say the behavior emerged from the agents interacting rather than from instructions, and draw a testing conclusion: safety evaluations run one agent at a time will not surface any of it. Multi-agent is how most agent tooling is now sold.

Databricks asked for a billion dollars and closed five

[CONFIRMED — TECHCRUNCH + REUTERS]

The round closed at $5 billion on a $190 billion valuation after an original ask of $1 billion. The gap between the ask and the close says as much about how much capital is looking for an AI home as it does about Databricks.

Someone hid a message for an AI inside a legal filing

[SINGLE-OUTLET REPORT — NO RULING TURNED ON IT]

404 Media reports a filing containing text invisible to human readers, instructing any model that read the document to side with the person who filed it. Prompt injection is usually discussed as a demo on a web page. This one was filed. The report establishes the attempt, not that it worked, and the distinction between those two is the whole story.

ALSO ON THE WIRE — 20 MORE

› Commure ends AI-clinic customer referral payment program
› Microsoft retires Mico, Copilot's animated voice-mode character
› Massachusetts teen accused of killing family used ChatGPT to research fantasy killing stories
› Claude Code adds auto-resume after usage limits reset
› Epoch AI: Anthropic's compute buildout suggests financing won't be the near-term bottleneck
› Wired dissects Zuckerberg's 6,500-word AI manifesto
› Writer launches new AI model and harness aimed at cutting token costs
› ChatGPT Desktop for Linux arrives as a preview release
› Sports Illustrated publisher Arena Group rebrands as AI company Paradium
› OpenAI's Computer History tracks your Mac activity without taking screenshots
› Databricks raises $5B at $190B valuation after investor demand outpaced target
› Game studio CEO accused of replacing writers with AI draws backlash over response
› Suno Studio 2.0 adds chat-based instrument and plugin creation
› OpenAI CRO Denise Dresser to depart, second executive exit this week
› OpenAI Previews 'Ultrafast' Mode That Runs GPT-5.6 Sol 14x Faster
› IBM Partners With OpenAI to Train and Certify Consultants on Its AI
› ChatGPT Work and Codex on macOS Can Now Recall Your Mac Activity Without Screenshots
› Factory Launches 'Agent Effectiveness' Dashboard to Measure Droid AI's Coding Impact
› Gemini 3.7 Flash Ships With Coding Gains, Halves 3.6 Flash's Price
› Prime Intellect's Blackwell-Native CUDA Kernels Claim 2.4x Faster MoE Inference

LEARN ONE THING

What an agent harness actually is

[LESSON · 4 MIN · FREE, LIKE ALL OF THEM]

DeepSeek open-sourcing Harness is only interesting if you know what a harness does. It is the scaffolding around the model: the model decides what to do, and the harness supplies its context, runs the tools it asks for, and keeps the loop alive after a failed step. Two teams on the identical model can get very different results, and the harness is usually why. This lesson walks through what goes in one and where they break.

The two-minute lesson →

TODAY’S CATCH — DID YOU ACTUALLY READ THE NEWS?

1 · Evidence grading.

Gemini 3.7 Flash was said to beat Claude Sonnet 5 and GPT-5.6 Terra. How strong is that evidence?

A· Strong, because independent labs reproduced the comparison ahead of the launch announcement.

B· Strong, because the comparison was published as a peer-reviewed result rather than a blog post.

C· Weak on its own: the comparison is run by the company selling the model.

D· Impossible to grade at all, because benchmark scores never carry any real information.

2 · What happened.

Google released Gemini 3.7 Flash. What did Google say about its price?

A· It matches the price of Gemini 3.6 Flash, with the improvement coming from speed rather than from cost.

B· It runs at half the price of Gemini 3.6 Flash.

C· It costs more than Gemini 3.6 Flash, because the newer model is larger and more capable.

D· It is free during an initial preview period, with commercial pricing to be announced later.

3 · Catch the error.

Statements about DeepSeek's releases this week:

§1· DeepSeek open-sourced Harness, the agent runtime it had used internally to benchmark its own models.

§2· Harness was released under an MIT license, built so that every capability is a swappable plugin.

§3· In the same week, DeepSeek cut its cache-hit API pricing by roughly six times to encourage agent workloads.

§4· DeepSeek moved V4 Pro out of preview, and Reuters reported it is priced up to fourteen times the V4 Flash tier.

Answers + the rest of today’s Catch → (scored on the site; streaks count)

Forwarded this? Get your own. Free, daily, no hype survives contact.

      Assembled by the AI University pipeline under published editorial rules: every claim attributed or omitted · vendor claims labeled · corrections logged.

      We don’t track opens. No pixel. Clicks are counted in aggregate so we know what teaches.

      You’re receiving this because you signed up at aiuniversity.com.

      Unsubscribe (one click) · Preferences

      AI University · PO Box 219 PMB 211 · Duvall, WA 98019