Speed, Context, and the Cost Illusion

00:00:04 Two things shipped this week that are quieter than benchmark headlines but probably more consequential. Cursor and SpaceX AI released a teach-a-task feature for Grok Bot—you hit a button, record yourself doing something in the browser, and the model learns your workflow.

00:00:23 OpenAI rolled out Computer History for ChatGPT on the same day, which ingests everything you do on your computer so it can pick up exactly where you left off. On paper these are incremental UX additions. But together they mark a shift that's been building: the bottleneck in AI has moved from capability to access to context.

00:00:45 Just because a model can do something doesn't mean it knows what you know. Microsoft tried a version of this with Windows Recall back in 2024, but the backlash forced a retraction. What makes the difference now is the actual workflow: an agent doing your work instead of just summarizing it.

00:01:06 Two different companies arriving at the same constraint independently suggests we've hit the wall on what prompting alone can solve.

00:01:15 While context wins are interesting, the daily operational math is where the immediate pressure sits. Google released Gemini 3.7 Flash on Thursday at three hundred and forty tokens per second according to Artificial Analysis. That's more than twice the pace of GPT-5.6 Luna and even a hair faster than Nvidia's Nemotron 3.5 Lightning.

00:01:40 Google also slashed prices by half compared to 3.6 Flash. But here's where the model sits in an uncomfortable spot: it costs about forty cents per task on Artificial Analysis's benchmark run. That places it at the same price as MuSpark 1.2 and slightly above Nemotron 3 Ultra or GLM 5.2—roughly eight times more expensive than GPT-5.6 Luna.

00:02:05 Brandon Galang of Vercel pushed back on the FUD around this. He noted that while 5.6 Luna technically edges out 3.7 Flash on benchmarks, it's a much smaller model and his team wouldn't turn to it for coding tasks. For years, Gemini has held its own on vision and multimodal tasks.

00:02:26 He's giving 3.7 Flash a shot as a daily driver. What Galang's take points to is partially synchronous coding—the kind of work where you sit at the terminal watching what the agent produces and correct course in real time. Speed matters there more than peak benchmark quality.

00:02:47 If the model can iterate with you instead of making you wait for it, that changes the tradeoff entirely.

00:02:55 There's a study that came out of Alpha Sense worth looking at because it directly challenges an assumption a lot of buyers are making: that switching to cheaper models, especially Chinese ones like Kimi K3 and GLM 5.2, automatically saves money. AlphaSense tested seven models on real financial analysis tasks—sifting through earnings call transcripts, SEC filings, and news articles, scoring outputs for factual accuracy and perspective diversity.

00:03:28 GPT-5.6 Soul beat Kimi K3 by delivering twenty percent higher quality at thirteen percent lower cost. Gemma 4 and Inkling matched GLM 5.2's quality for under one-fifth the price. And Opus 5—there's always an outlier—delivered worse results than Opus 4.8 while costing more than five times as much.

00:03:49 AlphaSense CEO Jack Kokko framed it simply: some models that look expensive based on per-token pricing actually end up cheaper because they use tokens more efficiently. For enterprise buyers, the faster you shift your conventional wisdom around this and figure out the models that are most efficient for your particular tasks, the better off your budget is going to be.

00:04:17 Token efficiency isn't about parameter count or architecture labels. It's about whether a given model wastes fewer tokens repeating itself, hedging too much, or going off-track and needing re-reading. The task-level cost is what actually shows up on the budget sheet.

00:04:36 Cost and speed only cover half the equation. Ethan Mollick framed it slightly differently yesterday, flagging that Google AI Overview alone will reshape the web and what we consume on it over time. That shift's already starting. For teams building with AI right now, measure what actually matters at your layer: How many retries does a given model need?

00:04:58 Does the speed bump justify the quality trade-off for synchronous workflows? Can you teach it your context once and stop paying per-token? Those are the questions that will separate the pilots from the production deployments in the next six months. — Seln.