Speed, safety, and the real world now move together
GPT-5.3-Codex-Spark, 6th-generation Waymo Driver and Gemini 3 Deep Think, and The Deploy Log for this week. Every price and claim read off the publisher's own page, with a call on each.
OpenAI
GPT-5.3-Codex-Spark
DEPLOYif you pay for ChatGPT Pro
WHAT SHIPPEDOpenAI released a research preview of GPT-5.3-Codex-Spark, a smaller GPT-5.3-Codex model that delivers more than 1000 tokens per second on Cerebras hardware, with a 128k context window and text-only input.
THE CATCHThe speed comes from a latency-first serving tier on Cerebras Wafer Scale Engine 3, and the same pipeline work cut time-to-first-token by 50% for all models, not just Spark.
IF YOU RUNa Codex workflow where you interrupt and redirect the model while it edits code
SOURCEOpenAI, Introducing GPT-5.3-Codex-Spark
Waymo
6th-generation Waymo Driver
WATCHif you operate a fleet
WHAT SHIPPEDWaymo began fully autonomous operations with its 6th-generation Driver on the Ojai, built on nearly 200 million fully autonomous miles across 10+ major cities, with a 17 megapixel imager and fewer than half the cameras of the 5th-generation system.
THE CATCHThe public cannot ride yet: Waymo is still giving trips to employees and their guests while it refines the rider experience before opening to the public.
IF YOU RUNa ride-hailing or delivery service in a city with extreme winter weather
SOURCEWaymo, Beginning fully autonomous operations with the 6th-generation Waymo Driver
Google DeepMind
Gemini 3 Deep Think
DEPLOYif you subscribe to Google AI Ultra
WHAT SHIPPEDGoogle DeepMind released a major upgrade to Gemini 3 Deep Think, now available in the Gemini app for Google AI Ultra subscribers, with 48.4% on Humanity's Last Exam without tools, 84.6% on ARC-AGI-2, and an Elo of 3455 on Codeforces.
THE CATCHThe API route is not open access: researchers, engineers, and enterprises must express interest in an early access program rather than sign up directly.
IF YOU RUNa research team working on messy scientific data with no single correct answer
SOURCEGoogle DeepMind, Gemini 3 Deep Think: Advancing science, research and engineering
The Deploy Log
Models and APIs
| WATCHif you work in theoretical physics GPT-5.2 physics result. GPT-5.2 Pro conjectured a formula for single-minus gluon tree amplitudes, and an internal scaffolded GPT-5.2 spent roughly 12 hours producing a formal proof, with the preprint available on arXiv. OpenAI |
| WATCHif you build for defense agencies ChatGPT on GenAI.mil. OpenAI deployed a custom ChatGPT on GenAI.mil, the Department of War's secure enterprise AI platform used by 3 million civilian and military personnel, for unclassified work in authorized government cloud infrastructure. OpenAI |
| DEPLOYif you run models in the browser Transformers.js v4. Transformers.js v4 is now available on NPM with a C++ WebGPU runtime, a 10x build time drop from 2 seconds to 200 milliseconds, and a default export that is 53% smaller. Hugging Face |
| DEPLOYif you optimize PyTorch inference on H100 CUDA kernel agent skill. Hugging Face released an agent skill that teaches Codex and Claude to write production CUDA kernels, producing an RMSNorm kernel for Qwen3-8B with an average 1.94x speedup on H100. Hugging Face |
| WATCHif you evaluate tool-using agents OpenEnv Calendar Gym. Turing contributed a production-grade calendar management environment to OpenEnv, where agents achieved close to 90% success with explicit calendar identifiers but dropped to roughly 40% with natural language descriptions. Hugging Face |
Tools and products
| DEPLOYif you use Notion for ambitious projects Claude Opus 4.6 in Notion. Notion added Claude Opus 4.6, the strongest model Anthropic has shipped, which breaks complicated requests into concrete steps and produces polished work. Notion |
| DEPLOYif you ship Next.js or SvelteKit apps Vercel Flags. Vercel Flags entered public beta with targeting rules, user segments, and environment controls in the Vercel Dashboard, priced at $30 per 1 million flag requests. Vercel |
| DEPLOYif you run origin servers that fail intermittently stale-if-error support. Vercel CDN now supports the stale-if-error Cache-Control directive, serving previously cached responses during 500 errors, network failures, or DNS errors. Vercel |
| DEPLOYif you route multiple models through one API MiniMax M2.5 on AI Gateway. MiniMax M2.5 is now available on AI Gateway, planning before it builds and handling full-stack projects across Web, Android, iOS, Windows, and Mac. Vercel |
| DEPLOYif you extract data from contracts GLM-5 on AI Gateway. GLM-5 from Z.AI is now available on AI Gateway with no other provider accounts required, adding multiple thinking modes and improved long-range planning over GLM-4.7. Vercel |
Breakthroughs
| WATCHif you schedule batch jobs on cloud Time-varying scheduling. Google Research introduced the first constant-factor approximation algorithms for non-preemptive throughput maximization under time-varying capacity, with a 1/2-approximation offline and a 1/11 competitive ratio online with common deadlines. Google Research |
| WATCHif you design multi-party AI conversations DialogLab. Google Research introduced DialogLab, an open-source prototyping framework for authoring, simulating, and testing dynamic human-AI group conversations, evaluated with 14 participants who rated human control mode significantly more engaging. Google Research |
Left out
- Runtime logs via MCP: a figure in the draft could not be found on the publisher's own page
- next-mdx-remote block: a figure in the draft could not be found on the publisher's own page
- GABRIEL toolkit: it did not carry a number that moved
- vercel logs command: it did not carry a number that moved
- Sandbox egress firewall: it did not carry a number that moved
- Token secret scanning: it did not carry a number that moved
- PostHog on Marketplace: it did not carry a number that moved
- Sign in with Apple: it did not carry a number that moved
- Browserbase on Marketplace: it did not carry a number that moved
- Lockdown Mode: it did not carry a number that moved
The pattern
The week's strongest items all trade one scarce resource for another: OpenAI trades model size for 1000 tokens per second on Cerebras, Waymo trades camera count for a 17 megapixel imager, and Deep Think trades broad access for benchmark scores that reset the frontier.
The quieter thread is control. Lockdown Mode, Vercel's egress firewall, and secret scanning all assume the model or the code will be attacked, and they ship deterministic off switches rather than better prompts. The deployment decision is no longer just which model to run, but which network paths, tools, and credentials to leave open while it runs.