Google shipped Gemini 3.7 Flash while 3.5 Pro stays late, and OpenAI previewed Ultrafast running GPT-5.6 Sol at up to 14 times normal speed. Latency, not parameter count, is the pitch now.
Google released Gemini 3.7 Flash on Aug 13 promising fewer prompts per app and lower token prices, while 3.5 Pro still has no release date. OpenAI answered with an Ultrafast preview running GPT-5.6 Sol at up to 14 times normal speed.
Nemotron 4's largest variant is expected to reach at least one trillion parameters, but training is unfinished and Nvidia has confirmed no release date. Alongside it, the open-source NeMo Switchyard router targets where agent cost control actually happens.
The Trump administration is preparing 100% tariffs on certain imported drones and components, hitting a hardware category still dominated by Chinese suppliers. Commercial operators absorb the bill first.
Koda publishes a fact-checked AI intelligence briefing every morning: site, email, podcast, and video. Founding sponsors lock today's rates for 6 months as the audience compounds.
View the media kit →Google released Gemini 3.7 Flash on Aug 13, claiming better debugging, more deployable production-ready code on the first attempt, app builds in fewer prompts, and lower token prices than the prior Flash release. The company also moved its Gemini Spark productivity agent onto 3.7 Flash the same day and added...
Continue reading arrow_outwardGoogle released Gemini 3.7 Flash on Aug 13 with cheaper tokens, better debugging and more first-attempt production code, while the promised 3.5 Pro remains late.
The Trump administration is preparing 100% tariffs on certain imported drones and components, a category still dominated by Chinese suppliers, which would reprice fleets for agriculture, surveying, inspection and film operators.
Mood reads Fear, with crude sliding about 2% to near $81 even as Iran insists no vessel transits Hormuz without its authorization.
Nvidia gave away NeMo Switchyard, the routing layer that decides which model answers, on the same news cycle it teased a trillion-parameter Nemotron 4: own the cheap plumbing, then sell the expensive endpoint.
Google's pitch for Gemini 3.7 Flash is fewer prompts per app and lower token prices; OpenAI's Ultrafast preview runs GPT-5.6 Sol at up to 14 times normal speed. With GPT-6 still unannounced and Gemini 3.5 Pro still late, both labs are competing on responsiveness and cost per task rather than a new flagship. That is what a market does when the next capability jump is not ready to ship.
Nemotron 4's largest variant is expected to reach at least one trillion parameters, but training is unfinished and Nvidia has confirmed no release date, so this is roadmap signal, not artifact. Paired with the open-source NeMo Switchyard router, the strategy is to occupy both ends of the agent stack: the weights developers download free and the routing library that dispatches their requests.
Hiring Dali Rajic as Chief Revenue Officer puts an enterprise sales veteran over a product with more than a billion weekly ChatGPT users, in the same week as enterprise deployment case studies and days after ads testing began inside ChatGPT. The research narrative is being fitted with a revenue engine. Expect pricing, packaging and ad inventory to shape the next model release as much as benchmarks do.
14x Faster, Zero New Intelligence
Read full analysis arrow_outwardIt is a gift finder, sure, but it is also the cheapest five-minute lab for learning how much context a recommendation model actually needs before it stops guessing.
For anyone stuck on a birthday present tonight, this is a no-account, no-cost brainstorm that beats staring at a blank search bar. For builders, the real value is diagnostic: run the same recipient with two sentences and then with six and you will feel exactly where prompt specificity starts paying off. As a product it is thin, with generic and repetitive output flagged in hands-on testing, so treat it as an idea generator rather than a shopping engine.
“Reviewers aggregated on a directory page say it removes the stress of gift hunting and surfaces creative ideas they would not have thought of themselves.”Autonoly directory review aggregation
Ahrefs launched Letaido on August 12 as an AI agent workspace aimed at the recurring work marketing teams never finish: competitor monitoring, keyword and SERP research, and the weekly reports that eat a day. Start it on one report you already produce manually so you can diff the agent's output against a known-good version before you trust it with client deliverables.
Try it arrow_outwardBuilt by an ex-OpenAI researcher, Energy automates multi-step computer work by first pulling context from email, local files, the browser, and connected apps, then executing. That ordering matters: the failure mode of most desktop agents is confidently acting on a stale or partial picture, so test it on a task where you can verify every input it claims to have read.
Try it arrow_outwardxAI's Grok Bot is pitched as an always-on teammate: it receives a dedicated cloud machine, signs into your existing tools with real credentials, and finishes multi-step jobs unsupervised. Before delegating anything, decide which accounts it gets and create scoped logins rather than sharing your own, because an unsupervised agent with your session is an audit problem waiting to happen.
Try it arrow_outwardgetimg.ai is a generation and editing suite that sits in the top tier of TAAFT's image category by usage, with tens of thousands of saves logged on its directory page. It earns its keep on volume tasks like variant testing and background replacement across a batch; for a single flagship asset you will still spend the time in a real editor.
Try it arrow_outwardBenchmark Flash against your own prompt chain. Take one production workflow and rerun it on Gemini 3.7 Flash, counting prompts to a working build and total token cost rather than trusting the launch claims. If it needs fewer round trips, that is your real win.
Put a router in front of your agents. Try Nvidia's open-source NeMo Switchyard to direct requests across models and tools, then log which calls actually needed the expensive model. Cost control lives at the routing layer, not in the model choice.
Automate one reporting loop this week. Point Letaido, launched August 12, at a single recurring marketing report, or test Energy's desktop agent on a task where reading local context matters more than raw model power.
Two of the three biggest labs spent today selling speed and cost instead of capability, and that tells you where the margin pressure sits. Nvidia's trillion-parameter Nemotron 4 is the counter-narrative, but with training unfinished and no release date it is a roadmap, not a product. Meanwhile OpenAI hiring Dali Rajic as Chief Revenue Officer signals that a billion weekly ChatGPT users are about to be monetized in earnest. Build for the cheap fast model you can actually ship on, and treat routing as the place your bill gets decided.
AI analysis, world news, markets, and tools. One briefing, delivered free.