The Speed Race: Ultrafast AI and Millisecond Decision Models

James Turner

Welcome to 48 Hour AI. I'm James Turner, and this covers October 10 to 12. Last time, we talked about OpenAI's Intelligent UI and Anthropic dropping Claude Haiku 5.5 to ten cents per million input tokens. It turns out those price cuts were phase one. Phase two is about speed, and it's showing up on almost every front at once.

James Turner

Start with OpenAI's new Ultrafast tier for GPT 6.1 Sol. It's a dedicated service tier across the API, Codex, and ChatGPT Work, built for latency critical agent workflows. OpenAI says it delivers an eight to fourteen times throughput boost while keeping reasoning near Astra, its top model. Think about what an agent actually does. It calls a model, waits, calls a tool, waits, calls the model again. Dozens of those waits stack up. Cut each one by a big factor, and a task that felt like watching a progress bar starts to feel like working with a colleague.

James Turner

OpenAI also halved top tier inference API pricing and made automated code review tooling free for developers. And in Codex and ChatGPT Work, the default context for GPT 6.1 Sol went up roughly eighty percent, to 500,000 tokens, with overrides up to a million. Here's why that matters. In long sessions, the system used to hit a limit and compact, summarizing earlier work to make room. Compaction loses detail. A bigger default window means fewer of those lossy summaries in the middle of a task.

James Turner

Then there's the Decisions API, now generally available. It handles classification and workflow branching in under 25 milliseconds using GPT 6 Luna, and it does it without autoregressive token generation. Hold onto that phrase, because it's the thread running through the rest of this episode.

James Turner

Now the reality checks. Reports say OpenAI told investors its annualized revenue is approaching 50 billion dollars, roughly 20 billion below figures that had been circulating. Its IPO has reportedly slipped to early 2027. And USA Today Co. and affiliated newspapers sued OpenAI in Manhattan federal court, alleging hundreds of thousands of articles were copied without authorization, and seeking more than 250 million dollars. The tech industry continues to argue that training is fair use. So the speed and price push is happening while the company has to justify very large spending. That last connection is my interpretation, not something the reports claim, but the timing is hard to ignore.

James Turner

Back to that phrase, non autoregressive. A normal language model generates an answer one token at a time, and each token depends on the last. That's great for writing an essay. It's wasteful when all you need is yes or no. Should this request go to the big model? Is this message safe? A decision model takes the question and the state, runs a single forward pass, and returns a structured answer, like a calibrated probability, with zero output tokens.

James Turner

Liquid AI released open weights for d1 3B and an experimental d1 omni 600M. The 3B answers a question in eight milliseconds on an RTX 4090. Hugging Face released JEV 27B VL, which takes text and images and outputs calibrated probabilities. Microsoft launched Microsoft Decision 1, a fine tuned Qwen 3.5 9B. And llama.cpp added native server support, so you can run routing, moderation, and guardrail checks on local hardware. Money is following. TypeSafe AI closed 870 million dollars at a 7.5 billion dollar valuation for its Jev framework. One case study says handling 100,000 compliance alerts dropped from about 290,000 dollars to under 26,000. That's a case study, so treat it as an illustration, not a benchmark.

James Turner

Google's move this week was a universal Gemini agent for work. Google Cloud CEO Thomas Kurian showed it spanning Google Workspace, Microsoft 365, Slack, and developer environments. The standout is persistent coworker agents, each with its own agents.company.com email address, calendar, and Drive storage, plus four kinds of memory: session, semantic, procedural, and episodic. It even routes work between Gemini and Anthropic's Claude models. But analysts at Forkast point to a delivery gap. It's still in private preview, and wide availability is only listed as coming soon for select Business and Enterprise plans. If you're outside that circle, it's an idea, not a tool.

James Turner

Protocols tell a similar story. Meta and Sierra are pushing the Personal Agent Protocol, with Shopify, Stripe, and Walmart among the launch partners. A working draft called Poppy followed, with 35 more design partners, including OpenAI, Visa, Mastercard, Okta, and Cloudflare. It covers discovery through a standard endpoint, OAuth sign in, sessions that persist across channels, and task execution. But version 0.1 of the spec isn't published, and the community A2A Agent Registry spec is delayed too. Forkast's warning is that when standards slip, proprietary vendor extensions rush into the vacuum, and interoperability suffers. Their advice is to treat vendor specific pieces as temporary bridges.

James Turner

Anthropic had a busy defensive week. Its Cyber Mission includes a Critical Infrastructure Defense Program for operational technology that can't easily be taken offline, plus OSS Scanner, which offers participating open source projects free audits by Claude models, including Mythos, with reproducers and sometimes proposed patches. The Cyber Verification Program now has three access tiers. Its revised usage policy, effective November 12, bans sustained, purposeless cruelty toward Claude, which formalizes the conversation ending behavior added in August. It also bars undermining democratic processes and guidance on arming drones. Claude for Google Workspace entered public beta, and Anthropic committed 150 million dollars over three years to the federal Genesis Mission.

James Turner

Quickly, on research and creative tools. Reflection introduced Beam, a mixture of experts model with 501 billion total parameters and 23 billion active, scoring 80.9 percent on SWE bench Verified, with Apache weights promised for late October. Meta Superintelligence Labs described agentic meta reasoning, a separate reasoning stream that lets agents monitor their own progress, reaching 71.5 percent on ProgramBench with GPT 5.5. Google's Nano Banana 2.1 does 4K conversational editing with up to 14 reference images, at 1328 Elo on LMArena. And Black Forest Labs' FLUX 3 Image adds bounding box controls for composition.

James Turner

So the thread is faster, cheaper, narrower models, with the plumbing around them still unfinished. Watch whether Poppy's version 0.1 and Gemini's wider rollout actually land this month. I'm James Turner, and I'll see you in 48 hours.