the system achieved this training result more than 20 times faster than conventional synchronization methods.

chips from different generations running at different speeds still matched the ML performance of single-chip-type training runs, ensuring that even older hardware can meaningfully accelerate AI training.

With increasing levels of hardware failure, Decoupled DiLoCo continues to deliver a high level of 'goodput', or useful training, while that of other approaches nosedives.

By dividing large training runs across decoupled 'islands' of compute, with asynchronous data flowing between them, this architecture isolates local disruptions so that other parts of the system can keep learning efficiently.

Amazon is investing $5 billion in Anthropic today, with up to an additional $20 billion in the future. This builds on the $8 billion Amazon has previously invested.

Claude remains the only frontier AI model available to customers on all three of the world's largest cloud platforms: AWS (Bedrock), Google Cloud (Vertex AI), and Microsoft Azure (Foundry).

Anthropic will also use incremental capacity for Claude in Amazon Bedrock. The agreement includes expansion of inference in Asia and Europe to better serve Claude's growing international customer base.

Our run-rate revenue has now surpassed $30 billion, up from approximately $9 billion at the end of 2025.

We have signed a new agreement with Amazon that will deepen our existing partnership and secure up to 5 gigawatts (GW) of capacity for training and deploying Claude

Amazon is investing $5 billion in Anthropic today, with up to an additional $20 billion in the future

run-rate revenue has now surpassed $30 billion, up from approximately $9 billion at the end of 2025

committing more than $100 billion over the next ten years to AWS technologies

over one million Trainium2 chips to train and serve Claude

over 100,000 customers now run Claude on Amazon Bedrock

up to 5 gigawatts (GW) of capacity for training and deploying Claude

over one million Trainium2 chips to train and serve Claude

run-rate revenue has now surpassed $30 billion, up from approximately $9 billion at the end of 2025

Amazon is investing $5 billion in Anthropic today, with up to an additional $20 billion in the future

more than $100 billion over the next ten years to AWS technologies

over 100,000 customers now run Claude on Amazon Bedrock

up to 5 gigawatts (GW) of capacity for training and deploying Claude

The Prompt API uses the Gemini Nano model in Chrome. While the API is built into Chrome, the model is downloaded separately the first time an origin uses the API.

The Prompt API for the web is still being developed. While we build this API, refer to our best practices on session management for optimal performance.

The network requirement is only for the initial download of the model. Subsequent use of the model does not require a network connection. No data is sent to Google or any third party when using the model.

The Prompt API isn't available in Web Workers for now, due to the complexity of establishing a responsible document for each worker in order to check the permissions policy status.

Microsoft continues to participate directly in OpenAI's growth as a major shareholder.

Revenue share payments from OpenAI to Microsoft continue through 2030, independent of OpenAI's technology progress, at the same percentage but subject to a total cap.

Microsoft will continue to have a license to OpenAI IP for models and products through 2032. Microsoft's license will now be non-exclusive.

Microsoft will no longer pay a revenue share to OpenAI.

OpenAI can now serve all its products to customers across any cloud provider.

this means that existing estimates overstate the returns to software R&D, and makes the software intelligence explosion seem much less likely.

But I think we have enough evidence to think that software progress might really be several times a year, and to make a best guess contextualized with a lot of uncertainty.

gpt-oss-20b does substantially better than GPT-3 on MMLU, despite using the same amount of training compute.

This means that almost all existing estimates of software progress were misleading.

these estimates rely on an overly conservative estimate of software progress of 3× per year

Synthetic data can help push beyond this — a good example that Millidge raises is the Phi series of models.

If doubling cumulative research effort also doubles compute efficiency, then the returns to R&D are 1. If it quadruples, then the returns are 2.

Almost all the evidence points to very fast software progress: each year, the training compute needed to get to the same capability declines several times — possibly even ten times or more.

AI software progress is about reducing the training compute you need to get to the same level of capability, through better algorithms or data.

I would put venture capitalist in finite demand & open loop. There's only a certain amount of venture capital dollars entering the ecosystem in a year, & investment selection remains an open problem.

Open Loop + Finite Demand = Utility Tools. Preparing 10-Ks & 10-Qs. Legal contract review. Insurance claims processing. One report per quarter, one contract per deal. AI makes the work faster, but doesn't create new work to do.

Closed Loop + Infinite Demand = Economic Engines. Software engineering lives here. AI writes the code. Tests verify correctness. More code enables more features. Companies will always need more software.

There were 1 billion commits in 2025. Now, it's 275 million per week, on pace for 14 billion this year if growth remains linear (spoiler: it won't.)

The compliance-driven buyers improvising local AI out of retail Mac Minis because the product they need does not exist.

Why the company that moved computing off the mainframe fifty years ago is making the same structural move with AI, and what that predicts.

The question it forces is not which model is best. It is who owns the inference layer your organization depends on, what happens when the economics of that layer stop being subsidized, and whether the thing in your pocket turns out to matter more than the thing in the datacenter.

The structural cost problem in AI inference that makes Apple's on-device bet defensible, not just defensive.

The board looked at the AI race Apple was losing and, rather than try harder at the thing that was failing, changed which game the company plays.

For a company that spent fifteen years running a functional model where no single discipline owned a product, putting two hardware engineers at the top is not a personnel decision. It is a structural break.

benchmarks sourced from publicly available material carry contamination risk, where training-data exposure can silently inflate scores.

This means that improvements on SWE-bench Verified no longer reflect meaningful improvements in models' real-world software development abilities. Instead, they increasingly reflect how much the model was exposed to the benchmark at training time.

We also found evidence that models that have seen the problems during training are more likely to succeed, because they have additional information needed to pass the underspecified tests.

Tests reject correct solutions: We audited a 27.6% subset of the dataset that models often failed to solve and found that at least 59.4% of the audited problems have flawed test cases that reject functionally correct submissions

Our RL infra team used a K2.6-backed agent that operated autonomously for 5 days, managing monitoring, incident response, and system operations, demonstrating persistent context, multi-threaded task handling, and full-cycle execution from alert to resolution.

The architecture scales horizontally to 300 sub-agents executing across 4,000 coordinated steps simultaneously, a substantial expansion from K2.5's 100 sub-agents and 1,500 steps.

Kimi K2.6 autonomously overhauled exchange-core, an 8-year-old open-source financial matching engine. Over a 13-hour execution, the model iterated through 12 optimization strategies, initiating over 1,000 tool calls to precisely modify more than 4,000 lines of code.

Kimi K2.6 demonstrates significant improvements over Kimi K2.5 in internal evaluations conducted by CodeBuddy: code generation accuracy increased by 12%, long-context stability improved by 18%, and tool invocation success rate reached 96.60%.

Meta founder and CEO Mark Zuckerberg described superintelligence in a blog post last year

Wedbush Securities analyst Dan Ives said in a report on Thursday.

The layoffs will start on May 20, the company confirmed.

Meta plans to lay off roughly 8,000 employees, or 10% of its workforce

Claude is now being deployed to NEC Group employees around the world

NEC will establish a Center of Excellence to develop a highly skilled, AI-enabled engineering organization

As part of its long-running Client Zero initiative, in which NEC serves as its own first customer before offering its technology to clients

NEC aims to build one of Japan's largest AI-native engineering teams, who will use Claude Code in their work.

Claude packages everything into a handoff bundle that you can pass to Claude Code with a single instruction.

Our most complex pages, which took 20+ prompts to recreate in other tools, only required 2 prompts in Claude Design.

What used to take a week of back-and-forth between briefs, mockups, and review rounds now happens in a single conversation.

Claude Design gives designers room to explore widely and everyone else a way to produce visual work.

Even experienced designers have to ration exploration—there's rarely time to prototype a dozen directions, so you limit yourself to a few.

The gains are especially strong in agentic coding, computer use, knowledge work, and early scientific research—areas where progress depends on reasoning across context and taking action over time.

GPT‑5.5 found a proof of a longstanding asymptotic fact about off-diagonal Ramsey numbers, later verified in Lean. The result is a concrete example of GPT‑5.5 contributing not just code or explanation, but a surprising and useful mathematical argument in a core research area.

We are treating the biological/chemical and cybersecurity capabilities of GPT‑5.5 as High under our Preparedness Framework. While GPT‑5.5 didn't reach Critical cybersecurity capability level, our evaluations and testing showed that its cybersecurity capabilities are a step up compared to GPT‑5.4.

Losing access to GPT‑5.5 feels like I've had a limb amputated.

GPT‑5.5 delivers this step up in intelligence without compromising on speed: larger, more capable models are often slower to serve, but GPT‑5.5 matches GPT‑5.4 per-token latency in real-world serving, while performing at a much higher level of intelligence.

GPT‑5.5 found a proof of a longstanding asymptotic fact about off-diagonal Ramsey numbers, later verified in Lean. The result is a concrete example of GPT‑5.5 contributing not just code or explanation, but a surprising and useful mathematical argument in a core research area.

The viable path is trusted access, robust safeguards that scale with capability, and the operational capacity to detect and respond to serious misuse.

We are treating the biological/chemical and cybersecurity capabilities of GPT‑5.5 as High under our Preparedness Framework. While GPT‑5.5 didn't reach Critical cybersecurity capability level, our evaluations and testing showed that its cybersecurity capabilities are a step up compared to GPT‑5.4.

GPT‑5.5 delivers this step up in intelligence without compromising on speed: larger, more capable models are often slower to serve, but GPT‑5.5 matches GPT‑5.4 per-token latency in real-world serving, while performing at a much higher level of intelligence.

The gains are especially strong in agentic coding, computer use, knowledge work, and early scientific research—areas where progress depends on reasoning across context and taking action over time.

We are treating the biological/chemical and cybersecurity capabilities of GPT‑5.5 as High under our Preparedness Framework. While GPT‑5.5 didn't reach Critical cybersecurity capability level, our evaluations and testing showed that its cybersecurity capabilities are a step up compared to GPT‑5.4.

GPT‑5.5 is priced higher than GPT‑5.4, it is both more intelligent and much more token efficient. In Codex, we have carefully tuned the experience so GPT‑5.5 delivers better results with fewer tokens than GPT‑5.4 for most users

The viable path is trusted access, robust safeguards that scale with capability, and the operational capacity to detect and respond to serious misuse.

GPT‑5.5 is our strongest agentic coding model to date. On **Terminal-Bench 2.0,** which tests complex command-line workflows requiring planning, iteration, and tool coordination, it achieves a state-of-the-art accuracy of 82.7%.

GPT‑5.5 delivers this step up in intelligence without compromising on speed: larger, more capable models are often slower to serve, but GPT‑5.5 matches GPT‑5.4 per-token latency in real-world serving, while performing at a much higher level of intelligence.

Drug manufacturers pay pharmacy benefit managers rebates above 50% of list price for formulary access.

Google Cloud launched a parallel $750m fund to pay McKinsey, Accenture, and Deloitte to train engineers and co-fund client AI projects.

Structure: $500M OpenAI equity plus $4B from TPG, Bain, Advent, Brookfield, and Goanna form a $10B LLC.

The median US buyout fund returns 13% to 16% net.

OpenAI pledged $1.5B to a joint venture called DeployCo, guaranteeing private-equity partners a 17% annual return floor over five years.

Jeremy didn't get laid off. He got leveraged.

The Meta cuts are the inverse. When one person with the right AI tools can do the work of 10-to-15 people, the person most at risk isn't the one using the AI. It's the one whose job description overlaps with what AI now does by itself.

A US lab would never; well, unless you count a code red or Meta's throw money at the problem moves.

The tradeoff is that the same input can map to more tokens—roughly 1.0–1.35× depending on the content type. Second, Opus 4.7 thinks more at higher effort levels, particularly on later turns in agentic settings. This improves its reliability on hard problems, but it does mean it produces more output tokens.

Our alignment assessment concluded that the model is 'largely well-aligned and trustworthy, though not fully ideal in its behavior'. Note that Mythos Preview remains the best-aligned model we've trained according to our evaluations.

On some measures, such as honesty and resistance to malicious 'prompt injection' attacks, Opus 4.7 is an improvement on Opus 4.6; in others (such as its tendency to give overly detailed harm-reduction advice on controlled substances), Opus 4.7 is modestly weaker.

Opus 4.7 is better at using file system-based memory. It remembers important notes across long, multi-session work, and uses them to move on to new tasks that, as a result, need less up-front context.

Interestingly, this means that prompts written for earlier models can sometimes now produce unexpected results: where previous models interpreted instructions loosely or skipped parts entirely, Opus 4.7 takes the instructions literally.

I had the intuition that these problems were kind of clustered together and they had some kind of unifying feel to them. And this new method is really confirming that intuition.

The LLM took an entirely different route, using a formula that was well known in related parts of math, but which no one had thought to apply to this type of question.

Several correlated but not strictly identical changes happened over the same few months: scaling inference compute, heavier use of RL in post-training, and models producing reasoning tokens.

Tasks where correctness is harder to verify may not have seen the same speedup, so the acceleration we document here may not be as general as the headline numbers suggest.

WeirdML V2 places models in an unusually resource-constrained environment: models get only five attempts to submit working code, with no access to external tools. This setup has not been the focus of recent RL training.

The three metrics where we find acceleration are concentrated in programming and mathematics. These are areas that labs have explicitly targeted for improvement, and they share an important property: correctness is easy to verify automatically.

Reasoning models show both a one-off jump in performance and a roughly 2-3x faster trend compared to non-reasoning models.

Three of four metrics show strong evidence of acceleration, seemingly driven by reasoning models.

Three of four metrics show strong evidence of acceleration, seemingly driven by reasoning models.

Parameters are estimated by unweighted least squares. Time t is measured in years since the first observation in each dataset.

We pre-selected the 6-month horizon as our primary metric, balancing genuine forecasting distance against the limited date range of our data.

The minimum training cutoffs are: ECI (June 2024), METR Time Horizon (January 2024), Combined Math (September 2024), and WeirdML V2 (January 2025).

Our fourth metric, an index constructed from WeirdML V2 results, showed no sign of acceleration. A single global linear trend fit the data best.

We use four AI capability metrics: ECI (Epoch Capabilities Index), METR 50% Time Horizon, Combined Math Index, and WeirdML V2 Index.

Reasoning models show both a one-off jump in performance and a roughly 2-3x faster trend compared to non-reasoning models.

Three of the four metrics (ECI, log METR 50% time horizon, and a math-focused index we constructed from several math benchmarks) show strong evidence that progress has sped up relative to a global linear trend fit to data from 2023 onward.

The three metrics where we find acceleration are concentrated in programming and mathematics. These are areas that labs have explicitly targeted for improvement

Our fourth metric, an index constructed from WeirdML V2 results, showed no sign of acceleration. A single global linear trend fit the data best.

The best-performing model across these three metrics was a pair of independent linear trends: one for reasoning models and one for non-reasoning models.

Reasoning models show both a one-off jump in performance and a roughly 2-3x faster trend compared to non-reasoning models.

Three of four metrics show strong evidence of acceleration, driven by reasoning models.

The three metrics where we find acceleration are concentrated in programming and mathematics.

We select the median-difficulty question from the set with maximum model coverage and standardize it to 0.

We work with the natural logarithm of the time horizon, which puts it on an approximately linear scale.

The minimum training cutoffs are: ECI (June 2024), METR Time Horizon (January 2024), Combined Math (September 2024), and WeirdML V2 (January 2025).

Reasoning models show both a one-off jump in performance and a roughly 2-3x faster trend compared to non-reasoning models.

Three of four metrics show strong evidence of acceleration, driven by reasoning models.

For Anthropic, more usage across diverse tasks means more data, which produces a smarter model—just as more queries improved Google search.

Commoditizing complements doesn't always work because focus is scarce even for the largest, fastest growing businesses.

A free, good-enough product is enough to change market dynamics.

The risk of this strategy to the ecosystem is that it makes previously attractive categories no longer viable.

The commoditization flywheel : both companies give away complements to drive usage of the core.

Commoditizing complements doesn't always work because focus is scarce even for the largest, fastest growing businesses.

Some categories never developed a competitive response to this strategy: email, advertising infrastructure, user-generated video.

A free, good-enough product is enough to change market dynamics.

Some categories never developed a competitive response to this strategy : email, advertising infrastructure, user-generated video.

For Anthropic, more usage across diverse tasks means more data, which produces a smarter model—just as more queries improved Google search.

But plenty of categories survived through specialization or direct competition : cloud, travel, domain registration, social networking. Commoditizing complements doesn't always work because focus is scarce even for the largest, fastest growing businesses.

The risk of this strategy to the ecosystem is that it makes previously attractive categories no longer viable. Commoditizing the complement does not demand a best-in-class replacement.

I'm not going to trust them to measure it.

If 90% is AI, do we even need a team?

Writing code is not the same as software development. This is only capturing some level of acceleration while writing code, and does not capture time taken in architecture, debugging, review, and deployment.

Cursor counted the entire file as AI, even though we can see from the diff that it left plenty of the lines unchanged.

So even though I did 100% of the writing and 50% of the refactoring, Windsurf reports that 100% of the code I produced in that session was generated by AI.

customers should expect PCW values of 85%+, often 95%+. This is not a hallucination and is accurate given how we compute this metric

Security is a defensive posture; agency is a functional right.

placing constraints upon them not only helps users and services build trust in them, but it also helps people more easily conceptualise what they do.

Some proposals for AI agents assume that putting agentic code in a TEE or similar 'jail' will solve these problems, but that ignores the need to collectively bargain

lack of a well-defined user agent role in AI that's backed up by transparent, public standards... leaves a gap – it makes it harder for a marketplace to form.

User Agents are a Form of Collective Bargaining

Every time you use an Internet-connected computer, you're trusting someone (and most likely, a multitude) to act on your behalf.

The agent interprets new information and adapts the logic. The engine applies that logic continuously and emits precise updates.

Agents and CDC streams are powerful together because they split the work well.

With change data capture (CDC), the system emits a stream of precise updates: inserts, updates, deletes, each tied to specific records.

The fix is not smarter prompts. It is software built to meet agents halfway.

Humans are not a good target for calm technology.

Today's agents, the copilots, the chatbots are designed to be human like.

A LeadDev survey found 54% of engineering leaders believe AI copilots will reduce junior hiring long-term.

When juniors skip debugging and skip the formative mistakes, they don't build the tacit expertise. And when my generation of engineers retires, that knowledge doesn't transfer to the AI.

The Pentagon told defense CEOs to consolidate or die. Fifty-one major defense contractors collapsed into five.

A nuclear weapons program lost the ability to make a material it invented. The knowledge existed only in people, and the people were gone.

memory-driven experience scaling represents a crucial new frontier for agent scaling

simple checklists evolved into memories with compositional, preventative logic structures

existing TTS methods often discard the exploration trajectory and treat the final answer as the only useful outcome

this self-judgement does not need to be perfectly accurate, as we find ReasoningBank to be quite robust against judgment noise

by over-emphasizing successful experiences, they miss out on a primary source of learning — their own failures

by recording detailed actions instead of tactical foresight, they fail to distill higher-level, transferable reasoning patterns

Fugu models achieve superior performance by dynamically coordinating and orchestrating a diverse pool of powerful models.

Instead of using domain knowledge to prescribe team organization, roles, or workflows, Fugu learns to dynamically assemble agents from a pool and coordinate them through non-obvious but highly efficient collaboration patterns.

The depth of recursion becomes a tunable compute axis at inference time, requiring no retraining.

When a Fugu model is allowed to call itself recursively, reading its own prior output as context and deciding whether to revise its coordination strategy, a new form of test-time scaling emerges.

A core conviction at Sakana AI is that the most capable AI systems will not be monolithic models scaled in isolation, but collections of specialized agents working together.

Instead of using domain knowledge to prescribe team organization, roles, or workflows, Fugu learns to dynamically assemble agents from a pool and coordinate them through non-obvious but highly efficient collaboration patterns.

A core conviction at Sakana AI is that the most capable AI systems will not be monolithic models scaled in isolation, but collections of specialized agents working together.

Fugu models achieve superior performance by dynamically coordinating and orchestrating a diverse pool of powerful models.

Sakana Fugu coordinates pools of frontier foundation models to achieve state-of-the-art performance across coding, mathematics, scientific reasoning, etc.

The depth of recursion becomes a tunable compute axis at inference time, requiring no retraining. A small model, by reading itself, can iterate toward answers that neither it nor any of its workers could reach in a single pass.

Instead of using domain knowledge to prescribe team organization, roles, or workflows, Fugu learns to dynamically assemble agents from a pool and coordinate them through non-obvious but highly efficient collaboration patterns.

Chinese authorities have reportedly been pushing data centers and public computing projects to use more domestic chips, including through reported bans on foreign-made chips, sourcing quotas, and requirements to pair Nvidia chips with Chinese alternatives.

DeepSeek does not appear to have fully moved beyond Nvidia. The company's technical report reveals that it is using Chinese chips to run the model for inference, but...appears to have adapted only part of V4's training process for Chinese chips.

In a 1-million-token context, V4-Pro uses only 27% of the computing power required by its previous model, V3.2, while cutting memory use to 10%.

DeepSeek V4 exceeds them all on coding, math, and STEM problems, making it one of the strongest open-source models ever released.

The filing cabinet keeps getting bigger. But a bigger filing cabinet is still a filing cabinet.

The current separation between training and deployment is not just an engineering convenience – it is a safety, auditability, and governance boundary.

A system that can look up any fact has not been forced to find structure. It has not been forced to generalize.

The intelligence lives in the static parameters, and the apparent capabilities change radically depending on what you feed into the window.

Without our safeguards in place (which we do to measure a model's raw capabilities), only Mythos Preview and Opus 4.7 completed more than half the tasks.

We also welcome feedback and input from third parties and industry experts. We're currently working with The Future of Free Speech (an independent think tank at Vanderbilt University), the Foundation for American Innovation, and the Collective Intelligence Project

if AI models can answer these questions well (that is, accurately and impartially), they can be a positive force for the democratic process.

No encryption protects that layer. The router can read, change, or replace anything.

Out of 28 paid and 400 free routers: > 9 injected malicious code into tool calls > 17 touched researcher-owned AWS credentials > 1 drained $500k from an Ethereum wallet

Client-side defenses caught 89% of injections. But real protection needs providers to sign responses.

Some attacks only fired after 50 prior calls. Others activated only in auto-approve mode.

a lightweight surrogate trained on them can absorb a significant portion of future traffic at near-zero marginal inference cost

On a 150-class benchmark, the surrogate fully replaces the teacher

the surrogate is activated only when its agreement with the LLM exceeds a user-specified threshold α

Duplicate section heading | Error; reject the file

Unknown component property | Accept with warning

The DESIGN.md format is at version `alpha`. The spec, token schema, and CLI are under active development. Expect changes to the format as it matures.

Components map a name to a group of sub-token properties: ... Variants (hover, active, pressed) are expressed as separate component entries with a related key name.

Unknown section heading | Preserve; do not error

A DESIGN.md file combines machine-readable design tokens (YAML front matter) with human-readable design rationale (markdown prose). Tokens give agents exact values. Prose tells them _why_ those values exist and how to apply them.