▶ YouTube Summaries

• YT :: https://www.youtube.com/watch?v=pJljViiUEPw
• Original title :: I love Ultrafast (it's unusable)

Theo tests OpenAI's Ultrafast mode for Astra (~300+ tokens/sec) and finds it far too expensive for almost anyone — $600 to review two sub-100-line PRs — yet it gave him a coding experience he hadn't felt in years: a tight, real-time back-and-forth with the model. He built Slopalytics, a model-benchmark dashboard, with it and argues that the value isn't saving time but finishing things you otherwise wouldn't build. The real hope is the upcoming 6.1 Soul Ultrafast, which could bring these speeds at sane prices.

Pricing is brutal

• Astra is already $50/M output tokens; Ultrafast is 6x — $300/M out, $450/M with long context. Cache reads cost $6/M and cache writes $75/M.
• Reviewing two small PRs (under 100 lines each) cost about $600.
• The $500/month plan gives roughly $1,000 of usage per week, which Ultrafast burns in about 2.1 hours. Two quick prompts (38s and 16s) cost 1% of the weekly limit; a handful of small UI changes cost 3%.
• One session priced out at $46.32 on Ultrafast, against $7.72 on standard Astra and about $1.10 on 6.1 Soul — over 40x more for the same work.
• Even OpenAI employees (who get free inference) now need per-task approval for Ultrafast. ThePrimeagen burned half his limit in an hour after buying the $500 plan.
• Theo's view: when you give users a knob they turn it all the way up — as with max reasoning — and Ultrafast shouldn't be a casual toggle without warnings.

What it feels like to use

• Measured speeds: regular about 30 TPS, fast about 60, Ultrafast 320–340.
• Changes show up live in the dev server as he types, so it makes sense to split the screen between app and model again rather than send a task and walk away.
• Tool-call latency becomes the bottleneck: when inference drops from 10 minutes to 30 seconds, a 30-second build or git call doubles the runtime. He wants a different system prompt for Ultrafast that steers the model away from slow tool calls.
• Astra handles steering well, so he can keep firing small corrections while it works and it folds them all in.
• It brings back the old flow state of jumping between code and output, now at the level of whole features.

Slopalytics: building in the loop

• He built it with Astra Ultrafast and Opus 5.5 in parallel. Opus took about 15–20 minutes to a first draft that looked better; Ultrafast took about 1.5 minutes, and by the time Opus finished, Ultrafast had been iterated past it.
• The thread had 50+ short messages ("not acceptable", "averages are useless, sort by highest", "top-right is forbidden, I need it for my face"), each answered in 10–60 seconds — about five prompts in four minutes.
• UX ideas came from staring at the live result: grouping models by family, reasoning-level toggles (low/mid/high/max), a click-to-pin comparison panel showing cost/token/score multipliers, and highlighting a line on hover.
• Astra is much worse at design than Anthropic's models, yet he ended up with a better design because being in the loop let him iterate on how things feel. Models are still bad at naming — he picked "Slopalytics" himself.
• It also highlights problems with Artificial Analysis: it defaults to max reasoning, so token and cost comparisons inside a model family are misleading.
• Cost: $36 for the main thread plus $250 for the follow-up. He puts the Ultrafast total at about $550, against about $90 on standard Astra or about $12 on 6.1 Soul.

The real comparison

• Measured as "is saving time worth $540?", the answer is no; never use it on that basis.
• But the honest comparison is that without the speed he wouldn't have built it at all. Waiting between steps loses context, so he would batch 20 changes and wait 5 minutes to 2 hours, as he does with Opus.
• With Ultrafast he wrote smaller requests and stayed engaged — ideal for refining UIs.

Where it's heading

• If 6.1 Soul Ultrafast keeps the 6x multiplier, it would cost about a tenth of Astra Ultrafast — around standard Astra pricing. Tibo has publicly confirmed 6.1 Soul Ultrafast is coming soon.
• That could make the $500 plan good value and Ultrafast his default. He already uses 6.1 Soul high fast as the default for managing machines in his fleet.
• Ultrafast runs on Nvidia, not Cerebras; the Soul version might be either.

When to use it today

• Barely ever — perhaps high-severity incident response, if you work at OpenAI.
• His own use: burning limits right before a reset. He has nine manual resets, and OpenAI promises a ship or a full reset every day for 28 days.
• Don't upgrade to the $500 tier for this yet. Whether the $200 tier is worth it is the subject of an upcoming video.
• Sponsor: Greptile's T-Rex, review agents that actually run the code on a real machine.