🗞 OpenAI's Cost Cuts, Acquisition Talks, and YC President's AI Safety Fund

🔎 The Latest on the AI Frontier:

• OpenAI Cuts API Costs in Half with New Flex Processing Tier

• OpenAI Pursued Cursor Before $3B Windsurf Acquisition Talks

• Former Y Combinator President Launches $100K AI Safety Fund

• Google's Gemini 2.5 Flash Introduces Cost-Saving "Thinking Budgets"

• OpenAI's Deep Research Outperforms Humans at Complex Searches But Still Fails 48.5% of the Time

• Other news you might find interesting

💰 OpenAI introduces Flex processing in API, cutting developer costs by 50% while extending response times

• The new service tier, currently in beta for o3 and o4-mini models, reduces API costs from $10 to $5 per million input tokens and $40 to $20 per million output tokens for o3 model.

• Flex processing comes with trade-offs including slower response times, occasional resource unavailability errors, and potential timeout issues for complex prompts, making it ideal for non-production tasks like model evaluations and data enrichment.

• Developers can mitigate timeout errors by increasing the default 10-minute timeout duration, while the "429 Resource Unavailable" error can be managed through exponential backoff retries or switching to standard service tier when timely completion is necessary.

💻 OpenAI pursued Cursor maker before entering talks to acquire AI coding company Windsurf for $3 billion

• OpenAI approached Cursor creator Anysphere in 2024 and again this year about a potential acquisition, but talks fell through as Anysphere is now reportedly raising capital at a $10 billion valuation.

• Windsurf generates approximately $40 million in annualized recurring revenue (ARR), while Cursor reportedly makes about $200 million on an ARR basis.

• Despite launching its own Codex CLI "agent" for code writing and editing, OpenAI's acquisition strategy indicates urgency in capturing market share in the rapidly growing AI code generation sector.

💼 Former Y Combinator president Geoff Ralston launches new venture fund exclusively for startups focused on AI safety and responsible deployment

• The Safe Artificial Intelligence Fund (SAIF) will provide $100,000 initial investments via SAFE agreements with a $10 million cap to early-stage companies developing tools for AI alignment, security infrastructure, governance, and misinformation prevention.

• Portfolio companies will receive weekly office hours with Ralston himself, coaching on Y Combinator applications, and access to his extensive network of investors—applying the successful YC mentorship model to this specialized fund.

• "There is a lot of disagreement over our future with AI, although we can surely all agree we want to get there safely," Ralston wrote, positioning SAIF as "a bet on a future that is both accelerated and aligned—and on the founders brave enough to build it."

💡 Google launches Gemini 2.5 Flash with innovative "thinking budgets" that reduce AI costs by 600% when reasoning capabilities are turned down

• The new model introduces unprecedented control over computational resources, charging $0.60 per million output tokens with thinking turned off versus $3.50 with reasoning enabled, allowing businesses to optimize costs based on task complexity.

• In benchmark testing, Gemini 2.5 Flash scored 12.1% on Humanity's Last Exam, outperforming Anthropic's Claude 3.7 Sonnet (8.9%) and DeepSeek R1 (8.6%), while posting strong results on technical benchmarks like GPQA diamond (78.3%) and AIME mathematics exams.

• "Companies should choose 2.5 Flash because it provides the best value for its cost and speed," said Tulsee Doshi, Product Director for Gemini Models at Google DeepMind, highlighting the model's strengths in "math, multimodal reasoning, long context, and several other key metrics."

🔍 OpenAI's Deep Research outperforms humans at complex web searches but still fails almost half the time despite superior stamina

• In OpenAI's new BrowseComp test designed for "challenging questions requiring searching through large spaces of potential answers," human researchers gave up on 70% of questions after two hours, while Deep Research successfully answered 51.5% of the 1,266 difficult questions that required finding "obscure, multi-hop facts."

• Researchers found that when Deep Research was allowed to generate multiple answers (up to 64) and then select the best one, accuracy improved significantly, suggesting "the model frequently 'knows' when it's right, even if it struggles to express that certainty."

• Performance scaled directly with computing power, reaching over 75% accuracy with more parallel tasks, but the technology still struggles with "confidence calibration" issues where models with browsing capabilities can become "overconfident about wrong answers."

🏭 TSMC accelerates US chip production with independent Arizona hub

• Taiwan Semiconductor Manufacturing Co. (TSMC) plans to produce 30% of the world's most advanced semiconductors in Arizona, aligning with the "Made in America" vision, according to CEO CC Wei.

• The company will accelerate construction of its second and third Arizona plants, with the third plant's construction expected to begin by the end of 2025.

• Despite these ambitious plans, TSMC's US facilities are reportedly lagging five years behind Taiwan's production capabilities, with the first plant operating at 4-nanometer technology while Taiwan advances to more cutting-edge processes.

More news you might find interesting:

• Meta removes Apple Intelligence writing tools from its major apps.

• OpenAI spends "tens of millions of dollars" on users saying "please" and "thank you" to ChatGPT.

• Google offers its premium Gemini Advanced subscription for free to all US college students until spring 2026.

• Salesforce CEO Marc Benioff envisions AI agents as business transformation tools.

• Enterprise risk management startup CTGT develops method to bypass censorship and bias in large language models like DeepSeek without sacrificing performance.

• DeepMind researchers propose "streams" approach to allow AI to learn from environmental experiences beyond human knowledge limitations.

• New Intel CEO Lip-Bu Tan dismantles hierarchy to revitalize innovation culture, bringing key engineering divisions under his direct control.

• Google plans to appeal the "adverse" portion of the U.S. court's decision finding it liable for monopolistic practices in digital advertising markets.

Have any feedback? Send us an email