Ultrafast love story: instant prototypes, brutal bills

Live prototyping: type a command, watch it change

The host asks the editor for a light theme with a funnier tone; before a screenshot finishes, the job is done and the interface updates in front of his eyes through a real-time stream . The pair behind this pace is the Astra stack with the model running in Ultrafast mode. This opening demo proves the speed claim on screen, not in words.

Then he opens the price table: while Standard Astra asks $50 per million output tokens, Ultrafast writes $300 for output, climbing to $450 in long context. Input lists at $60 per million tokens, with cached input at only $6. This price comparison is confirmed against the CodexHow reference and reads clearly next to the current OpenAI listing.

The everyday-work example hits hard: reviewing two pull requests under 100 lines costs roughly $600 . A figure that looks harmless in a one-off trial becomes a budget line once it spreads across every request. The lesson is simple: watch cost per completed task , not the model name.

At this point sponsor Greptile TREX takes the stage: it runs the pull-request branch in a sandboxed trial with services and browser agents, leaving logs, screenshots, and traces as evidence. It reportedly catches 20% more bugs than review alone; the product was announced by Daksh Gupta in June 2026. This sponsored detail is confirmed from the Greptile source and presented with its sponsored limits stated openly.

Slopalytics: a dashboard that bins averages

The host complains about the ArtificialAnalysis Intelligence Index v4.3.2 screen: among ten evaluations, the top maximum-reasoning figures and model-family relationships stay invisible. Claude Opus 5.5 and Sonnet 5.5 sit at the smartest end while Celeris-1 (1,739 tok/s) and Mercury 2.5 top the speed list. This gap is confirmed from the ArtificialAnalysis source, and the idea of building his own dashboard grows out of it.

Then the overnight live build begins: screenshots taken 41 seconds apart document progress with 1:13 a.m. and 1:14 a.m. timestamps. A human stays in the loop; vertical bars , lab-matching logo colors, and dropping a stray Opus thread play out frame by frame. The resulting board takes the name Slopalytics , with viewers watching every step live.

The design rule is blunt: averages are useless, promote the highest option. Average lines get removed and models cluster by family group . Instead of one number, viewers see who actually races whom, and the chart turns readable.

The figures land in a table: Sonnet at 200K, Opus at 120K, and Astra at 27K output tokens line up side by side at maximum reasoning. Once family grouping opens up, the scale pattern turns clear; small-output models shine where they do, big ones diverge where they must. This comparison is confirmed from the ArtificialAnalysis source and exposes the pattern the raw list hides.

Speed, cost, and access limits

Up to 8x token generation is not 8x shorter task time; tool calls and tests eat most of the clock. A WebSocket stream is recommended for tool-heavy agents, and Ultrafast runs on US data residency only. The model reached the API with DevDay on September 29, 2026; default limits apply at 500K TPM for tiers 1-3, 1M for tier 4, and 5M for tier 5. This technical detail is confirmed from the OpenTools source, and the speed debate should be read inside these bounds.

The access gate is narrow: Ultrafast opens only to the $500 Pro tier plus eligible Enterprise and Edu plans. It arrives off by default on the Enterprise side, and buying credits on a lower plan unlocks nothing. Usage with an API key bills at API rates . Pro tiers stack at $100, $200, and $500; this plan detail is confirmed from the OpenAI listing and reads together with the Codex rule of 8x against included usage and 6x against purchased credits.

Then comes the reset game: nine manual resets collected, six from an event, one expiring within an hour while a 34% slice melts away. Tibo promises "an improvement or a reset every day for 28 days", yet resets always land late. Luckily serious work runs on the Opus side, so burning the Codex account to zero stings little. This usage picture is confirmed from the CodexUsage guide and shows plainly how quotas melt in daily practice.

Verdict: love it, but guard the wallet

The verdict is blunt: he loves the speed but finds the price unusable, and advises against upgrading to the $500 plan for this alone. Carefree usage belongs only to moments like incident response , where seconds print money. A GPT-6.1 Sol Ultrafast option may arrive later, he notes.

He sets the context too: GPT-5.5 retires from ChatGPT, Work, and Codex on October 14, 2026, with the API untouched. The $200 plan carries 6.1 Sol, Luna, fast mode, and resets; its detail waits for a separate video. Viewers thus keep the product calendar apart from the pricing decision.

The practical takeaway starts with measurement: time to first token, sustained rate, and total time deserve separate tracking . Latency-sensitive interactive work is this speed tier's true home. Input, cached input, writes ($75 per million), and output tokens each need their own meter.