AI is getting cheaper more quickly than any other transformative tech in history. At a given level of performance, cost has fallen ~47%/quarter since 2023. That’s 4× faster than DNA sequencing, 6× faster than compute, 18× faster than lithium batteries, and (up to 1973) 54× faster than electricity.

The share of math preprints on arXiv that acknowledge AI has risen rapidly, from 4% in April to 25% in August. This increase holds even when filtering to papers with at least one author who published regularly before 2023.

Another problem from FrontierMath: Open Problems has been solved! The solution was elicited by Becker, Greger, and @DominikPeters in an interactive session with GPT-6 Astra. Peters originally suggested the problem for the benchmark. He had this to say.

Introducing Benchmark Reviews: our new initiative to audit AI benchmarks. We are launching with 15 benchmarks: 4 Verified, 9 Flawed, and 2 with not enough information for a review.

Trade data is consistent with more than $3B of chips smuggled into China via Malaysia. Between April 2024 and June 2025, China recorded $3.8 billion in server imports from Malaysia, averaging $106,000 each. Malaysia declared the same shipments at $17,000 each.

Earlier this week, we showed that GPT-5.6 and Claude 5 exhibit very different long-context TTFT scaling, which corresponds to their different pricing structures at long context lengths. OpenAI recently released GPT-6 Astra, which shows the same pattern of increased API pricing beyond 272k input tokens. We collected additional latency measurements, which show a similar curvature to that of GPT-5.6 models.

Every FrontierMath Tier 4 problem has now been solved by AI, with GPT-6 Astra solving the last problem standing. Mathematicians often commented that AI found unintended shortcuts when solving their Tier 4 problems. Not so for this last one, which was created by Jay Pantone.

OpenAI has grown its compute nearly 20-fold since 2023, the sharpest example of an industry-wide surge. Our new AI Chip Users explorer tracks the growth of compute use across five of the world's top frontier AI developers: OpenAI, Google DeepMind, Anthropic, Meta Superintelligence Labs, and SpaceXAI. Estimates include all compute used for AI research, training, and inference, which we model using power capacity disclosures from these developers where available, along with financial filings, third-party analyst estimates, and our own analysis of major AI data centers.

OpenAI GPT-5.6 models and Anthropic Claude 5 models have different pricing structures at long context lengths. GPT model costs increase in price past 272k input tokens, while Claude model costs remain fixed. Does this reflect an underlying difference in the architecture of these models? Our measurements of serving latency suggest so. We studied time to first token (TTFT) on these models and how it scales with increasing context length. We found a significant difference in how they scale, with GPT showing a noticeable quadratic component, while Claude models remain closer to linear.

Will Huawei catch up to Nvidia by 2030? Almost certainly not. We analyzed Huawei’s chip roadmap and supply chain, from the Ascend 950 to its 3D-stacking bet and domestic HBM supply. Our estimate: Huawei will produce less than 4% of Nvidia's AI compute in 2026. Relying on domestic HBM alone, that share could fall to just 1% by 2028. 🧵

GPT-6 Astra has set a new ECI record, with a score of 169. This is a substantial jump from the prior best (163), but is within our uncertainty range for the reasoning-era ECI trend. Astra also set new records on our math, continual learning, and game-puzzles benchmarks. On our long-horizon coding benchmark, MirrorCode, Astra ranks between Opus 4.7 and Fable 5. OpenAI gave us pre-release access to test Astra. Charts and more details for Astra’s individual benchmark results in the thread.

In 2025, Anthropic and OpenAI were already among the fastest-growing companies of their size in history. And yet, their revenue growth has accelerated in 2026. @justjoshinyou13 and @lynette_bye discuss what this means for the future of AI in the latest edition of our newsletter.

We discovered that US GDP statistics miss most of the value Nvidia adds to the US economy. As a result, GDP growth has been understated by ~0.3 percentage points over the last year.

Continuing to scale AI compute at recent rates will require exponentially more capital. Is financing a bottleneck? Probably not yet. That’s the takeaway from a new Gradient Update by @chutchesonai, which analyzes nearly $50B of debt associated with Anthropic’s buildout. Anthropic is a useful case because in November 2025 it announced plans to invest $50B in US compute infrastructure. It had less than $9B in annualized revenue at the time. To raise the necessary funds, Anthropic turned to institutional capital and vendor credit support: Investors supply the money upfront, lending against long-term lease payments, while Broadcom and Google make the leases easier to finance by agreeing to cover some losses if payments stop. The buildout has two legally separate sides: - Compute: $34.5B for TPU systems leased to Anthropic, $30B of which has Broadcom support - Infrastructure: ~$15.2B for data center capacity leased to Fluidstack, which provides capacity and operations to Anthropic, with Google supporting specified lease obligations On the compute side, a dedicated equipment company borrows as Google TPU systems arrive, buys the racks and leases them to Anthropic for five years. Broadcom conditionally supports the lease obligations backing $30B of debt, with a reported maximum exposure of $29B. Investors also committed $4.5B without Broadcom’s support. That junior tranche pays 8.5%, versus 5.75% for a supported senior tranche. Because the tranches also differ in seniority, the 2.75-point gap represents an upper bound on the backstop’s value. On the data center side, project companies use investor debt to build 1.43 GW of critical IT capacity. Fluidstack’s rent repays that debt. Google plays a role similar to Broadcom’s: If Fluidstack stops paying rent, it conditionally covers part of investors’ losses. Broadcom, Apollo, and Blackstone call the $35B compute deal the first transaction in a platform designed to support >20 GW through 2028, including deployments for Anthropic and OpenAI. This first buildout is a test case for investors. If it performs as expected, it will provide confidence for financing even larger projects at lower cost. This week’s Gradient Update was written by @chutchesonai. All Gradient Updates are informal, opinionated analyses that represent the views of individual authors, not Epoch AI as a whole. Read the full essay here:

AI appears to be finding software vulnerabilities at scale. In June 2026, 21 notable organizations disclosed ~1,500 high- and critical-severity CVEs, over 3.5× the previous monthly record set before Claude Mythos Preview's release.