Firefly IPO and GPT-5 keeps AI cheap

Welcome to Cautious Optimism, a newsletter on tech, business, and power.

Friday! The politicization of the American central bank continues, with a POTUS ally joining the voting ranks of the Fed. You can imagine my view on the matter, but for a more in-depth look at how adding Stephen Miran to the board of governors will impact interest rate decisions, the Times has you covered.

In micro macro news, a datacenter project in Tucson was voted down by its city council — to literal cheers. This is a two-sided situation: On one hand, we need more data centers lest the domestic economy fall behind its global rivals; on the other, planning to use drinking water to cool 290 acres worth of data center in a dry state during a drought was an error.

Here’s an idea: Datacenters where there’s surplus water. For example, along the Columbia River, where there are already a good number of compute hubs. Oddly, it’s in the very dry South West that we’re seeing the largest projected growth in new datacenter projects. So long as that continues, expect communities to raise a stink. Especially if they can’t shower. To work! — Alex

• 📈 Trending Up: Response to incentives … self-sabotage … Russian aggression in Ukraine … open-weight bragging … synthetic data … response to incentives … burn … tariffs on Russia? … Intel’s spine after Trump’s dig …

  • One to watch: German AI startup n8n could raise a round from Accel that pushes its value up to $2.3 billion. Add it to our list of European AI startups of note.

• 📉 Trending Down: Age verification online? … data science gigs? … India-US relations … Dojo … goods affordability … two states … my sanity … Pinterest, after earnings … Twilio, after earnings …

Is GPT-5 good?

After releasing a generally well-received set of open-weight models, OpenAI’s main event for the quarter came yesterday with the drop of GPT-5. The model family comes in three sizes, and developers can select varying levels of thinking when they call the new AI system.

But is it good? The short answer is yes. The slightly longer answer is yes, but.

GPT-5 is a good model. LMArena users voted it the top model on its platform, taking pole position in coding, math, writing, longer queries, and instruction following. In blind tests, it does well, which is an excellent start to GPT-5’s life.

Artificial Analysis ranks the model as the leader in its Artificial Analysis Intelligence Index, but only by a point over Grok-4, and a mere two points ahead of the previously released o3 model. OpenAI’s own benchmarking of the model shows improvements in most categories.

GPT-5, however, does not appear built purely to crush benchmarks. The idea of building an AI model that is good in a way that isn’t precisely measured by leading public benchmarking tools is not new. Anthropic’s Claude 4 Sonnet Thinking ranks eighth on the same Artificial Analysis scoring chart, for example, while the model is the literal market-leader over on OpenRouter, handling twice the tokens of its nearest rival, Google’s Gemini 2.5 Flash model.

So, the proof will come in the usage pudding.

Not that the benchmark Olympics don’t still have adherents. xAI’s leader made several public comments yesterday about how his Grok 4 models will best OpenAI on the statistical leaderboards.

• Ironically enough, it was market laggard Apple who made the most public noise this year about benchmark saturation, or data contamination regarding AI testing.

Apart from coding tests and other similar hurdles, OpenAI stressed that GPT-5 is better at writing and health-related queries, hallucinates less, and has an improved voice mode. It’s also a unified model, so you can use it without having to tell it to think or not to think or use search or not — it handles that for you. Which means that the infamous model-soup that OpenAI previously offered now gives you a single option.

Startups like Cursor had nice things to say about GPT-5, Simon Willison was impressed, and Raindrop AI CTO Ben Hylak gave it a big thumbs up. On the other hand, folks who expected OpenAI to crush more benchmarks were less impressed.

But what I think matters the most when it comes to GPT-5 is that it’s cheap and fast.

GPT-5 costs $1.25 per million input tokens, and $10 per million output tokens. Meanwhile, Grok 4 via the xAI API costs $3 per million input tokens and $15 per million output tokens. Claude Opus 4.1 via the Anthropic API will cost you $15 per million input tokens, and $75 per million output tokens — hell, even Claude Sonnot 4 is more expensive! Gemini 2.5 Pro from Google’s API costs $1.25 to $2.50 per million input tokens (lower prices for smaller loads), and $10 to $15 per million output tokens (again, hgiehr prices for longer results), for reference.

And GPT-5 is blisteringly fast in my testing. Better benchmarks, work on the model in ways that won’t precisely show up in tuned tests, and an attractive price, and speed? That mix seems like a win to me.

What’s nice is that we don’t have to actually make a final call between OpenAI made a good new model but maybe not one good enough to cement its market leadership for a year and OpenAI made a model good enough to cement its market leadership for a year. The market will for us. Either GPT-5 sees wide developer adoption, or not. The data will bear out.

• In consumer, of course, GPT-5 is about to become the most used model in the world. ChatGPT is huge, with 700 million weekly active users.

In other words, go play with GPT-5 and come to your own conclusion for your personal use. Then we’ll check the aggregate data in a few days and weeks to see what everyone else thinks.

Three more things

• Lower AI API prices are a big deal. TechCrunch reports that the Windsurfs and Cursors of the world struggle from very low, or even negative gross margins. Paying more for your inputs than you can charge for your outputs is a bad business. Especially when Anthropic’s expensive models are very popular amongst developers. GPT-5 being cheaper than many of its peers could make a lot of startups more profitable.

• Some folks aren’t worried: Reading the data engineering subreddit this morning, it appears that the folks there aren’t too concerned about GPT-5 taking their gigs.

• Startups are not out of the blender: For TWiST yesterday I prepared a list of startups that GPT-5’s improved capabilities may run up against. Think education AI startups, vibe-coding tools like Lovable, website builders like Squarespace, health-focused AI services like Hippocratic AI and Ada health, writing tools like Copy AI, and voice chat apps like Chatacter AI and Chai AI.

Firefly’s IPO blasts off

Despite concerns that the domestic economy is slowing, the tech IPO market is staying hot. After Figma’s electric debut earlier this month, Firefly Aerospace’s own debut went off like a bag of firecrackers.

After raising its IPO range, Firefly wound up pricing its launch and space vehicle business at $45 per share, ahead of its heightened price expectations. And it sold more shares, to boot.

And then it opened at $70 and closed at $60.35 per share, giving the company a heft first-day pop. Shares of Firefly are coming back to Earth a little today, but nothing catastrophic. We’re not witnessing a rapid, unscheduled disassembly of Firefly’s IPO win, in other words.

Worth just under $8 billion this morning, Firefly had revenues of ~$71 million in the first half of 2025, giving it an effective run-rate multiple of 56x, the sort of figure we see attached to the most valuable cloud-AI companies, like Palantir and Cloudflare. Not hardware companies whose products can, at times, explode.

My view that the Firefly IPO says a smidgen more about investor interest in the spacetech economy than it does in Firefly per se may seem rude, but since the result is the same for the newly-public company, we can afford to be slightly unkind.

No matter, the folks who can’t get SpaceX shares have another place to park a bet. And they seem stoked at the opportunity.

Grab-Bag

• Stablecoins get tamped down in China: Regulators in China are working to slow stablecoin hype, with Bloomberg reporting that “China told local brokers and other bodies to stop publishing research or hold seminars to promote stablecoins, seeking to rein in the asset class to avoid instability.” The idea of stablecoins — third-party tokens backed by hard assets pegged to a single currency’s value — is predicated on central govenrments having less control over their currency. Creating alternatives has that effect. China, however, is not a place where economic power apart from state control is smiled upon by the government.

• Open-Closed: OpenDoor became a meme stock recently, though its second-quarter earnings appear to have partially quenched that trade. Elsewhere, the other well-known ‘open’ company from Keith Rabois — an investor now back at Khosla — OpenStore, is also struggling to stay afloat. Not that I am one to write entire bullet points from pique. Of course not.