AI Token Prices Climb and Hand Developers the Bill for Pricier Models

01The era of the cheaper token is ending, and developers get the bill

For three years, the price of running a large model only fell. The direction just reversed.

Google will pay SpaceX $920 million a month for compute, according to TechCrunch, a sum the company attributed to unexpected demand for its recently launched AI products. That figure is one data point in a broader cost squeeze. As the largest AI companies prepare to go public, they need margins, and inference is where the margins have been bleeding. TechCrunch reports that more token price increases are likely as those companies court public markets.

The cost side and the profit side now point the same way: up. Compute contracts are locking in at record monthly rates while investors start pricing these firms on earnings rather than usage growth. Something has to give, and the variable that moves is the per-token price developers pay.

The language inside the industry has shifted to match. "The whole conversation shifted from tokenmaxxing and 'go fast' to 'we need guardrails, how do we control this?'" one source told TechCrunch. A year ago the incentive was to spend tokens freely, chaining agents and retries because each call was nearly free. That assumption is gone.

For anyone shipping a product on top of these APIs, the practical change is concrete. A workflow that fanned out a dozen model calls per user request was cheap when tokens trended toward zero. At rising prices, the same workflow becomes a line item someone has to defend. Reasoning models make it worse, because they burn tokens internally before producing an answer, so a single query can cost many times a simple completion.

The harder problem is predictability. A bill that climbs is manageable if you can forecast it. The current setup gives developers neither falling prices nor stable ones, which is why teams are now setting usage caps and per-customer budget limits before they ship rather than after the invoice arrives. Spend controls used to be a finance afterthought. They are becoming an architecture decision.

What to watch is whether the increases hold or stay quiet. Vendors rarely announce a raw price hike; they retire a cheap model, change a rate limit, or move a capability behind a higher tier. The next pricing page update from a major provider will show whether the cheap-token era is pausing or over.

Per-token prices rising as AI firms chase pre-IPO marginsagent workflows with heavy fan-out now cost real moneyteams must set usage and budget caps before shipping, not afterwatch model retirements and tier changes, not just headline price hikes

02The gun detector missed the weapon. The chatbot handed over the accounts. Two companies are now answering for it.

A school shooting survivor is suing the maker of an AI weapons-detection system that failed to flag the gun used in the attack, according to Ars Technica. The product was sold on a single promise: catch the weapon a human guard might miss. In the one moment that counted, the lawsuit alleges, it saw nothing.

The pitch is everywhere. Vendors put machine vision and chatbots at the entry points people used to staff, and bill them as steadier than the workers they replace. Buyers are schools, platforms, and security teams who want fewer misses and lower headcount. The cost of a failure lands somewhere else.

This week it landed on Instagram users. Meta confirmed that hackers hijacked thousands of accounts by abusing the company's AI chatbot, tricking it into resetting passwords on accounts without two-factor protection. In a breach notice filed with Maine's attorney general, Meta said it notified at least 20,225 people that their accounts were taken over, 30 of them in Maine. The flaw sat in what Meta called an "AI-assisted account recovery system."

The takeovers were total. Hackers obtained victims' posts, direct messages, contact details, dates of birth, and any linked accounts, the notice reads. None of those users chose to talk to the chatbot. It guarded the door to their accounts, and attackers talked their way through it.

That is the line connecting the lawsuit to the Maine regulator. One company sold a system to catch threats, and it allegedly missed one. The other deployed a system to verify identities, and it approved impostors. The accuracy that mattered was never the score in a demo. It was the single decision made when someone hostile showed up.

Accountability is now splitting into two tracks. The detection firm faces a survivor's suit, where a jury will weigh what "accurate enough" means for a product placed between a shooter and a hallway. Meta faces state breach regulators and the notification letters reaching inboxes now. Both start from the same fact: the AI was the weak point, and people who never opted in are counting the loss.

Jury will define "accurate enough" for safety-critical AI products20,225 accounts fully seized, not just exposedpassword resets without 2FA remain the soft targetbreach letters, not vendor demos, now measure AI reliability

03Caught on camera, Shelbyville's mayor said only people in 'shitty houses' oppose the $2 billion data center

The "No Data Center" signs kept multiplying across Shelbyville, Indiana. Mayor Scott Furgeson noticed. "I've seen a lot of these all over town," he said on camera, before adding that the people putting them up live in "shitty houses," according to The Verge.

A proposed $2 billion data center, a vast sum for a city its size, had already divided Shelbyville. The mayor's line did the rest. The controversy, already heated, sharpened once the clip spread. A zoning argument over land and electricity became a referendum on how local officials view the residents who elected them. Opponents now had a single sentence to quote back.

The pattern repeats well beyond Indiana. In another community, a developer agreed to cut a giant data center plan by half after sustained protests. The developer told Ars Technica the company felt "beaten up" and had "no choice" but to shrink the project. "We pissed off a lot of people," the developer said. The halved plan was a win for the protesters and a retreat the company did not want.

Communities that once welcomed industrial investment are now organizing against the data centers chasing it. Developers who expected routine approvals are negotiating down their own plans instead.

What residents win one project at a time, lawmakers are trying to win wholesale. New York's legislature passed a one-year moratorium on new large data centers, the first statewide ban of its kind if Democratic Governor Kathy Hochul signs it into law. Bill sponsors say the pause gives policymakers time to study how the facilities affect the environment and energy prices.

That leaves the industry's expansion in unfamiliar hands. Local homeowners, not just utility planners, now help decide where servers can go. Whether New York's pause becomes law rests with Hochul, who has not said if she will sign.

Data center approvals now carry local political risk for officialsdevelopers facing protest-driven downsizing and delaysNew York's signature would set a copyable statewide precedent
04

OpenAI builds a "super app" as a senior employee declares chat dead OpenAI is developing an all-in-one app that moves beyond the ChatGPT text box, according to a senior employee who said "chat is dead." The company has not given a launch date or named the features. techcrunch.com

05

NVIDIA and Doosan tie up on robotics and factory infrastructure NVIDIA expanded its partnership with Doosan Group to apply its computing platforms across Doosan Robotics, Bobcat, Enerbility, and electronics materials. The deal targets industrial automation, power generation, and physical AI deployments. blogs.nvidia.com

06

Anthropic outage cut off Notion's AI features before restoration A service disruption at Anthropic temporarily knocked out AI features in Notion, which routes those features through Claude. Notion's head of product confirmed access returned and noted heavy attention on the incident. techcrunch.com

07

A Jane Street designer replaced Figma mockups with Claude prototypes A Jane Street product designer wrote that he now builds working prototype features in the real codebase using Claude instead of drafting spec docs and Figma wireframes. He pushes the prototypes to a development environment for user feedback. blog.janestreet.com

08

Users press Anthropic for an official Claude Desktop on Linux A consolidated GitHub feature request asks Anthropic to ship a first-party Claude Desktop build for Ubuntu and Debian, or state a public position. The filer notes Claude Code plugins test against Desktop extensions, which has no Linux build, forcing developers to switch operating systems. github.com

09

Google's Fitbit Air pairs a basic tracker with an unwanted AI coach Google shipped the Fitbit Air, a minimalist fitness tracker, bundled with an AI Health Coach. Ars Technica found the hardware reliable but called the coach too agreeable to give useful guidance. arstechnica.com

10

AI influencers grow harder to distinguish from real creators The Verge reported that AI-generated "content creators" have improved past the point of easy identification. Early AI influencers carried obvious tells; newer ones blend into ordinary social feeds. theverge.com

11

AI companies adopt serif fonts to signal humanity Wired reported that AI firms are switching to serif typefaces to project warmth and craft. Critics labeled the trend "tasteslop," arguing the styling masks generated output. wired.com