Gemini 3.7 Flash 가격 인하, 초고속 GPT-5.6 Sol, 에이전트 실행 환경 경쟁

Today we're introducing Gemini 3.7 Flash, our most intelligent workhorse model yet for coding and agents. This model brings substantial gains across software engineering, web development, and complex knowledge work. Now through the end of the year, Gemini 3.7 Flash is available at an introductory price* of $0.75/1M input and $3.75/1M output tokens — making it easier and more cost-effective for developers and customers to scale production-ready agents.

Our most intelligent workhorse model yet for coding and agents has arrived Meet Gemini 3.7 Flash. — Crush that seemingly endless to-do list. Gemini Spark in the @geminiapp now uses 3.7 Flash. The new model can equip your personal AI agent to work even smarter for you by seamlessly handling complex, multi-step tasks across your @GoogleWorkspace apps like @gmail , @googlecalendar and @googledocs — Enjoy a smoother build experience. The model thinks more diligently, putting more effort into multi-step planning and tool calls. A more disciplined execution means less manual oversight and fewer retries across engineering workflows — Build more, spend less. 3.7 Flash is available through the end of the year with an introductory price of half the original 3.6 Flash cost per million tokens ($0.75/1M input tokens and $3.75/1M output tokens)

Previewing Ultrafast mode: GPT-5.6 Sol at up to 14x the speed. Launching first in the OpenAI API to a select group of customers with expanded access to more businesses as capacity grows.

Powered by @Cerebras , Ultrafast generates up to 750 tokens per second, bringing our most intelligent model to products and workflows where every second counts. Ultrafast is designed for businesses where faster frontier intelligence creates a measurable advantage, including real-time voice and customer support, commerce, coding and design, financial research, and security response.

Cloud agents now start 3x faster so you can hand them ambitious, long-running tasks to execute from start to finish. This performance improvement comes from builds: ready-to-use development environments that Cursor prepares continuously in the background, at no additional cost.

Builds also make agents more resilient and easier to debug. When a new build fails, it never goes live. Agents keep working from the last successful build while you debug in the background.

An agent is three things: a harness, a model, and context. If you're serious about owning your intelligence, you probably want to own all three. @LangChain founder @hwchase17 joined us at our @sequoia Own Your Intelligence to talk about the piece that often gets the least attention: the harness. He offers a clear heuristic for when to build your own. The more out of distribution you are from what the models were trained on, the more you'll want to customize. And good technical content on how to actually measure performance with evals and langsmith. 00:00 Introduction 00:58 The three parts of an agent: harness, model, context 02:12 What a harness actually does 03:25 Customizing the core loop with middleware 04:41 Sandboxes, file systems, sub-agents, summarization 05:47 Cognitive architectures — and when you still need them 07:03 Build your own harness or use off the shelf? 08:24 In-distribution vs. out-of-distribution: the file-editing example 09:39 Why evals define what "good" means in an organization 11:04 Harbor: what an eval task actually looks like 12:11 Comparing harnesses and models on accuracy, latency, and cost 13:20 Why observability is underrated — it's usually the context 14:34 The data flywheel: traces → curation → experiments 15:42 Getting feedback through UX design and online evaluators 16:51 Demo: LangSmith Engine 19:23 Q&A: Running Engine on Engine, and "codex-ification" 20:44 Q&A: Will harnesses converge or diverge?

Very interesting new paper from Microsoft and colleagues. (bookmark it) Skill libraries are used in every major harness on the assumption that more guidance is free. This work measures what a bad skill actually costs you. They attribute 307 agent failures to specific loaded skills, 125 functional failures and 182 efficiency regressions, by comparing each skill-guided run against a matched reference run that solves the same task. The failures rarely come from irrelevant skills. Seemingly relevant skills push the agent to incorrectly implement or omit something the task required. Cost regressions are not explained by prompt length either. The largest source is excessive verification at 67 cases, followed by heavy implementation pipelines at 30 cases. It turns out that skills quietly turn validation checklists into mandatory work. Paper: https:// arxiv.org/abs/2608.11888 Track more trending AI papers in our academy: https:// academy.dair.ai

When SWE-bench was launched initially, people questioned whether it was feasible at all. We got very similar feedback for ProgramBench. Up until now, only 2/200 tasks were solved. But Opus 5 just made an enormous jump with 9 fully solved tasks instances!

Elon Musk

@elonmusk

Not bad

We're releasing Mistral OCR 4.1, a model update to OCR 4 that reads busy, marked-up pages more precisely: bounding boxes align to each element, with no drift and no nested images. More in

#1. Not just among open models. SOTA, full stop. H3 just took the crown on Video Edit Arena. Drop in your footage, prompt the impossible, and let H3 reimagine every frame. Show us your wildest edit.

This video was one of the very worst MPU has done. It’s entirely focused on farmers in the desert feeling entitled to take as much water from the Colorado River as they’d like to farm in the very most arid place in the country with no competition at all with any industry that also uses water. Their first example is a pecan farmer who demonstrates how he dumps huge amounts of water into the ground and then implies that anyone who’s okay with data centers using water must “know how to eat computer chips” because “there won’t be any food left.” This guy thinks if he has to pay a little more for water to farm in a desert, it’s like saying we shouldn’t have food at all. The video frames him as a hero. They also bring up the offsite water cost of generating electricity as if it’s this secret data centers are hiding by not reporting it in their water use. For any other building that uses water, we usually only expect them to report their internal power and water cost and infer the water cost of the electricity from the power plants. This is all out in the open. They interview a county water manager who says the data center really isn’t an issue and only uses 0.4% of the local water, and then immediately cut to a local activist saying they can’t have enough for the data center because water costs have gone up. These are obviously unrelated, because as the video notes the city’s raised water rates every year for 11 years. I could go on and on, every moment is insane.

Ali Ghodsi

@alighodsi

Today, we announced that we crossed $7B in revenue run-rate, growing over 80% year over year in Q2. We also shared: $100M+ revenue run-rate for Lakebase $1.5B+ revenue run-rate for Lakehouse, growing over 100% year over year Continued positive adjusted free cash flow And we raised $5B in our latest fundraise. We’ll use this capital to invest in: Lakebase, our serverless Postgres database built for AI agents Genie, our AI coworkers that actually understand your business data Unity AI Gateway, our multi-AI governance solution that helps control costs @iamVictorDey shares more in @Forbes :

Databricks wraps $5 billion funding round at $190 billion valuation