Token Costs Threaten Frontier Lab Dominance
Research Rundown #191, plus new memos on Sakana and Flow Engineering
Rising Cost Competition to Frontier Labs
In AI, the high token cost of using frontier models is becoming an increasingly loud part of the conversation. This week, OpenAI launched GPT-5.6 Sol at $5 input / $30 output per million tokens, half the input cost of Fable 5, which costs $10 input / $50 output. The price gap with open source is even wider. DeepSeek V4 Flash is available at $0.14 input / $0.28 output, about 36x cheaper than GPT-5.5 on input and over 100x cheaper on output. Meanwhile, China’s Z.ai released GLM-5.2, which costs about an eighth as much as Claude Opus 4.8 for certain tasks and is now the third most widely used AI system in the world.
Because China produces most of the top-performing open-source models, six of the ten models on the leading AI leaderboard were developed in China, meaning the cost pressure on US frontier labs is increasingly coming from state-backed competitors with no profit motive. Enterprises are starting to respond to the realities of token economics. It was reported this month that Uber burned through its entire annual AI budget in four months and has since implemented spending tiers on AI tools. Nevertheless, roughly 95% of enterprise AI usage still runs on frontier models, suggesting the shift to cheaper alternatives has barely begun.
Meta Eyes Prediction Markets
This Tuesday, it was reported that Mark Zuckerberg had tasked a small team with building a standalone prediction-markets app, internally called “Arena,” to rival Polymarket and Kalshi. Independent from Facebook, Instagram, WhatsApp, and Messenger, the app would initially use video-game-style points rather than cash, but hasn’t ruled out using real money in the future.
Arena would be Meta’s second attempt at a prediction market, after launching Forecast in 2020, a points-based COVID-era prediction app, which sunset in 2021 after failing to gain traction with a no-money, conversation-based strategy. Other tech companies have tried and failed with similar projects: Google ran the largest corporate markets ever, Prophit (2005–2011), which died after Dodd-Frank killed its regulatory path, and the internal Gleangen in 2020. The biggest differences between these experiments and rapidly-growing platforms Polymarket and Kalshi are (1) the ability to bet with real money and (2) the inclusion of sports contracts, which make up the vast majority of prediction market trading today.
US EVs Score Much-Needed Regulatory Win
This week, the Department of Transportation proposed removing the brake-pedal requirement for vehicles “designed to be driven exclusively by automated driving systems,” removing an important barrier for Tesla’s wheel-and-pedal-free Cybercab and Amazon’s Zoox. The same day, the Commerce Department barred Polestar, the Geely-owned EV maker, from selling new cars in the US under the “Connected Vehicle Rule,” restricting Chinese software or hardware only months after granting the same authorization to Polestar’s sibling company, Volvo, also owned by Geely.
This support to the EV industry comes as US auto sales just fell for an eighth straight month, with EV sales collapsing 35.5% year-to-date and EV market share sliding to 5.1%. The EV slump has been framed as a competitiveness crisis with China, which accounts for about 60% of global EV sales, and with BYD surpassing Tesla as the world’s largest EV maker in 2025.
Latest Research
Flow Engineering is an AI-native physical engineering platform building the system of record for complex hardware programs. Read our new report here.
Sakana is a Japanese AI model developer that launched a multi-agent orchestration system purported to rival Mythos. Read our new report here.
What We’re Reading
OpenAI is considering delaying its blockbuster IPO into 2027 after SpaceX’s turbulent post-IPO trading spooked advisors and rekindled internal tension between CEO Sam Altman and CFO Sarah Friar, who has urged waiting to manage cash burn and compute commitments.
OpenAI and Broadcom unveiled Jalapeño, OpenAI’s first custom inference chip, completing the design-to-tapeout in nine months (claimed to be the fastest ASIC cycle ever), with OpenAI’s own models accelerating parts of the work; the first deployment is set for late 2026 at a gigawatt scale with Microsoft and other partners.
Anthropic accused Alibaba of conducting the largest known distillation attack on Claude, alleging operators tied to Alibaba’s Qwen lab ran 28.8 million exchanges through ~25K fake accounts to extract reasoning and coding capabilities; the June 10 letter to the Senate Banking Committee follows three earlier industrial-scale campaigns Anthropic attributed to DeepSeek, Moonshot, and MiniMax.
The NSA lost access to Anthropic’s Mythos 5 after the Trump administration’s June 12 export-control order forced Anthropic to disable the model globally, depriving analysts of a tool that had found vulnerabilities in classified networks within hours during an internal Project Glasswing red-team test (the source of Sen. Mark Warner’s widely misread “broke into almost all our classified systems” comment).
Meta paused its Model Capability Initiative, the mandatory employee-monitoring program that recorded keystrokes and screen activity to train internal AI models, after a misconfiguration left tens of thousands of internal data tables of private conversations, performance reviews, and transcriptions readable by any Meta employee; a 1.5K-signature staff petition predated the breach.
President Trump signed two executive orders pushing for a research-grade quantum computer by 2028, also ordering the federal migration to post-quantum cryptography by 2030-2031 and Pentagon deployment of quantum sensors by 2028; the orders follow the Commerce Department’s $2 billion in equity stakes across nine quantum-computing companies last month, including a new IBM venture.
NVIDIA introduced its Rubin reference data-center design, a closed-loop liquid-cooling system claiming up to a 100% reduction in on-site water use by running outdoor dry coolers instead of evaporative chillers. However, the closed loop addresses only ~25-33% of total AI data center water consumption, since electricity generation and chip fabrication also draw water.
SpaceX launched and deployed Starfall on its first demonstration flight from Cape Canaveral, a saucer-shaped reentry capsule designed to return up to 1,000 kg of cargo from orbit anywhere on Earth; the FAA assessment describes Starfall as a potential “proliferated successor” to the ISS for orbital pharma and bio-manufacturing, and the Pentagon has parallel point-to-point cargo agreements with Blue Origin, Rocket Lab, and Anduril.
Kalshi opened talks to raise fresh capital at a roughly $40 billion valuation, nearly doubling its $22 billion May valuation in seven weeks as the FIFA World Cup drove monthly volume above $17 billion; the talks land amid an escalating jurisdictional fight, with the CFTC suing nine states and CME suing the CFTC over Kalshi’s perpetual-futures approval.
China retook the world’s fastest supercomputer crown for the first time since 2017 when the Shenzhen-based LineShine hit 2.198 exaflops on the TOP500 list, dethroning Lawrence Livermore’s El Capitan using a CPU-only design built entirely on Chinese-designed processors; organizer Jack Dongarra said US export controls give China “a strong incentive to develop domestic alternatives.”
Meta and EssilorLuxottica rolled out Meta Glasses starting at $299 across three frame styles (Meta Adventurer, Meta Fury, and a Kylie Jenner collaboration), all running Meta AI powered by Muse Spark (the first model out of Meta Superintelligence Labs) from day one, with a new Rx Lens Swap for post-purchase prescription lenses and live translation now covering 20 languages.
Apple raised prices across its Mac and iPad lineups (MacBook Neo from $599 to $699, MacBook Air from $1,099 to $1,299, iPad Air from $599 to $749) while sparing the iPhone for now, citing an “unprecedented” memory and storage shortage driven by AI data-center demand; Counterpoint Research says memory prices have roughly quadrupled in three quarters, and CEO Tim Cook last week called it a “hundred-year flood.”
At Contrary Research, we’ve built the best starting place to understand private tech companies. If you're interested in researching and writing about tech, apply here for our Research Fellowship!