Engineering brief
Ultrafast is incredible and unusable. That's the point.
At a glance
The brief is enough
- Relevance
- Practical value
- Radar
- Time horizon
- This year
- Warnings
- None
Ultrafast mode delivers 300+ tokens per second. It also costs $600 to review two pull requests.
Ultrafast reveals a new interaction paradigm, but costs make it strictly experimental for now.
Watch if you want to understand the interaction potential of sub-second inference, not just the current pricing.
Summary
The cost problem. Ultrafast mode delivers astonishing speed — 300+ tokens per second on Anthropic's Astra model — but at a price that makes it impractical for almost every engineering team. The reviewer burned $600 reviewing two small pull requests and $46 in under a minute of light work on the $500/month plan. The 6x multiplier over standard pricing pushes output tokens to $300 per million, with cache writes at $75 per million.
What makes it special. The real innovation is the interaction pattern, not the speed alone. When inference drops from minutes to seconds, developers can stay in an active feedback loop with the model. The reviewer built a full analytics dashboard called Slopytics in under two minutes for the first draft, then iterated through 50+ rapid prompts — most under 30 seconds each — to refine UI details that would never emerge from a batch-and-wait workflow.
The hidden workflow shift. Fast inference changes the tool use ratio entirely. When generation took 10 minutes and shell commands took 30 seconds, nobody optimized the tool calls. But when generation drops to 30 seconds, those same tool calls now double the runtime. Teams will need different system prompts and agent designs specifically tuned to minimize non-inference overhead.
What this means for pricing. The $500/month plan gives roughly $1,000 of usage credit per week, which burns through in about 2.1 hours of continuous streaming. The reviewer's $500 plan dropped 3% in under a minute of work. OpenAI employees get unlimited inference, which tells you everything about who this is actually designed for. The product is a technology demonstration, not a production tool.
The real value proposition. The reviewer could build better UIs with Astra on Ultrafast than with superior models like Opus or Fable, because rapid iteration beat model quality. This is the strongest argument for the technology: speed enables a fundamentally different development workflow where you can try, reject, and refine ideas faster than you can list them.
The future signal. The reviewer confirms that 6.1 Sonnet Ultrafast is coming soon, likely on Cerebras hardware, and at potentially 10x cheaper pricing than Astra. That combination could make the $500 plan suddenly compelling. For now, Ultrafast is a glimpse of what's possible, not something teams should build on.
My take
I agree with the fundamental premise: speed changes the interaction pattern, not just the throughput. The Slopytics demo is the real signal, better output through iteration beats better models through batch work. But I'd push back hard on the framing that this is relevant for teams today. $500 for 2 hours of streaming inference is not a product, it's a research grant. What matters is the pricing trajectory. If 6.1 Sonnet Ultrafast lands at reasonable cost, the entire calculus shifts. Until then, this is a technology demonstration that engineers should watch closely but not touch. The workflow lessons about prompt granularity and tool-call optimization are worth adopting today, though.
What to do with this
- Audit your current AI spend per task to establish a baseline for comparing future Ultrafast pricing.
- Experiment with rapid iteration workflows using standard-speed models first to evaluate if the pattern works for your team.
- Ask your cloud provider about Cerebras or equivalent hardware availability for when cost-effective ultrafast options arrive.
Watch the video
This video is blocked due to your privacy settings. To watch this video, please accept YouTube marketing cookies.
Related breakdowns
GPT 6.1 Soul: cheaper and smarter, but not a replacement for Opus
OpenAI's GPT 6.1 Soul is aggressively priced and surprisingly capable, but teams should keep Opus 5.5 for complex autonomous tasks.
Agent Sandboxes Are the Real Product in Google's New API
Google DeepMind's managed agent sandbox turns infrastructure into an API call. The demo is slick; the cost, isolation, and control questions are unanswered.
AI agents are flooding large codebases with code faster than teams can
AI agents are flooding large codebases with code faster than teams can manage. Dan Adler argues the bottleneck is not model quality but cross-repo visibility…
Get TL;DW
Too Long; Didn't Watch.
A concise breakdowns of the AI and devtools videos that actually matter for engineering leaders.
Free. Weekly. No hype.
Video and thumbnails remain the property of their respective creators. tldw.news provides editorial analysis, commentary, and discovery links to original content.