🗞️ AI Daily Briefing — 2026-06-25

🔥 Top Story

OpenAI and Broadcom unveil Jalapeño — OpenAI’s first custom AI chip. Announced yesterday, the LLM-optimized inference accelerator went from design to tape-out in just nine months (with help from OpenAI’s own models). Early lab results show ~50% cost savings over current GPUs, with engineering samples already running GPT-5.3-Codex-Spark at production frequency. Deployment target: end of 2026. This is OpenAI’s clearest move yet to reduce its Nvidia dependency ahead of a potential IPO. (OpenAI) (TechCrunch) (Bloomberg)

🚀 Model & Research News

  • GPT-5.5-Cyber goes live as OpenAI’s most capable security model. Launched June 22 under the expanded Daybreak program, it scores 85.6% on CyberGym (vs. 81.8% for base GPT-5.5) and can trace attack paths, validate exploitability, generate patches, and produce remediation evidence in a single workflow. Access is gated to vetted defenders at Akamai, Cisco, CrowdStrike, Palo Alto Networks, and others. (OpenAI) (Axios)
  • ByteDance announces Seedance 2.5 at Volcano Engine FORCE. The next-gen video model generates 30-second native clips in a single pass (no stitching), accepts up to 50 multimodal reference inputs, and ships with an AI copyright system to address the Hollywood IP complaints that shelved Seedance 2.0’s international rollout. Early July launch via Volcano Engine. (The Next Web) (TechTimes)
  • Fable 5 / Mythos 5 — day 13 of the US export ban. No restoration timeline. Anthropic disagrees with the standard, saying it would “halt all new model deployments.” Claude Opus 4.8 remains unaffected and holds Arena #1. Meanwhile, Anthropic’s “AI for Science” virtual event on June 30 will be John Jumper’s first public appearance since joining from DeepMind. (Anthropic)
  • Gemini 3.5 Pro now likely slipping to July. Google has reportedly postponed the GA release, with prediction markets at ~50% for a June 30 launch. The model remains enterprise-only on Vertex AI. (CryptoBriefing)

🛠️ Tools & Developer Updates

  • Alteryx Agent Studio enters preview. Announced at Inspire 2026, Agent Studio lets analysts convert existing data workflows into autonomous AI agents without code rewrites. The companion Alteryx One MCP Server connects those agents to Slack, Teams, Claude, and OpenAI. The pitch: agents inherit years of business logic rather than hallucinating it from raw tables. (Enterprise DNA) (TechTarget)
  • Google reimagines Search with an AI-native input box. The biggest Search box upgrade in 25 years now accepts text, images, files, videos, and Chrome tabs as input — all processed by Gemini. (Google Blog)
  • OpenAI ships “Patch the Planet.” A new initiative funding open-source maintainers to remediate vulnerabilities found by GPT-5.5-Cyber. Part of the broader Daybreak cybersecurity push. (OpenAI)

💰 Funding & Business

  • SpaceX lands $6.3B compute deal with Reflection AI. The open-source AI startup (founded by ex-DeepMind researchers Misha Laskin and Ioannis Antonoglou) will pay $150M/month through 2029 for Nvidia GB300 access at Colossus 2 in Memphis. SpaceX is rapidly becoming a major neocloud — Anthropic, Google, Cursor, and now Reflection are all tenants. (CNBC) (Bloomberg)
  • Qualcomm circling Tenstorrent for $8–10B. The Qualcomm-Jim Keller deal would give Qualcomm serious RISC-V-based AI chip capabilities and a shot at Nvidia’s dominance. Talks ongoing, no deal confirmed. (Reuters via Yahoo Finance)
  • Amazon custom silicon hits $20B+ annual run rate. Now one of the top three datacenter chip businesses globally, growing >100% YoY. Trainium and Inferentia chips are finding enterprise traction. (Crescendo)

🐦 Notable from the Timeline

  • Anthropic’s June 30 “AI for Science” event will be the first public test of whether hiring Nobel laureate John Jumper and opening wet labs translates into a credible AI-for-science strategy. Worth watching.
  • Stellantis + Wayve + Uber signed an MoU for global Level 4 driverless robotaxis — Stellantis builds the hardware, Wayve provides the AI driving software, Uber deploys on its network. Builds on Wayve’s existing London/Tokyo autonomous rides. (Wayve)
  • Enterprise AI ROI reality check continues. 56% of CEOs still report no revenue or cost benefit from AI spending. Average enterprise AI spend hit $11.6M in 2026. The “tokenmaxxing is dead” narrative from last week keeps gaining steam.
  • Colorado’s AI Act (SB 189) replaced the original sweeping regulation with a lighter disclosure-based framework and pushed the effective date to January 1, 2027. The original June 30 deadline is moot. (Brownstein)

📊 Benchmark Watch

Arena leaderboard holds steady: Claude Opus 4.8 at #1 (~1510 ELO), GPT-5.5 Pro at #2, Gemini 3.1 Pro Preview at #3. Top 5 still within ~55 ELO — the tightest spread on record. All eyes on whether Gemini 3.5 Pro (whenever it actually ships) disrupts this order. 6.9M+ votes across 367 models. (Arena.ai)

🔗 Worth Reading

  • OpenAI’s Jalapeño chip deep-dive — The official blog post details the architecture philosophy, nine-month development timeline, and how OpenAI used its own models to accelerate chip design
  • “SpaceX is becoming the AWS of AI compute” — Medium analysis of how Colossus went from Grok’s playground to a multi-tenant neocloud with $1.4B+/month in committed compute revenue
  • Seedance 2.5 technical breakdown — How ByteDance solved the stitching problem for 30-second native video generation and what the 50-reference-input system means for professional workflows