Are there transparency regimes and tools that can enable a broad set of people, not just frontier AI companies, to easily study real-world AI usage?

If an intelligence explosion was upon us, what intervention points would facilitate slowing or otherwise changing the rate of the explosion? Assuming humans can intervene, which entities should wield this capacity—governments? Companies?

If AI substantially reduces the centrality of paid work in human life, what conditions will allow people to reallocate their time and effort toward other sources of meaning, and what can we learn from historical or contemporary populations where work has been scarce or optional?

It demonstrated incredible generalization. Without any retraining, TRINITY transferred zero-shot to four unseen tasks

This foundational research is part of the core engine powering our multi-agent product: Sakana Fugu

We believe the future of AI isn't just about scaling monolithic models, but engineering collaborative, diverse AI ecosystems that can adapt and combine their strengths.

TRINITY transferred zero-shot to four unseen tasks (AIME, BigCodeBench, MT-Bench, and GPQA). On average, the evolved coordinator surpassed every individual constituent model in its pool, including GPT-5, Gemini 2.5-Pro, and Claude-4-Sonnet.

We found that evolution is uniquely suited to optimize this tight, high-dimensional coordination problem where traditional gradient-based methods fail.

Imitation learning (Supervised Fine-Tuning) was ruled out because generating multi-turn labels is prohibitively expensive.

Traditional Reinforcement Learning (REINFORCE) failed because the gradients had a low signal-to-noise ratio due to binary rewards and weak parameter coupling.

The coordinator relies on the hidden states of a compact language model and a small routing head. In total, it has fewer than 20K learnable parameters.

While model merging offers a way to combine different skills, it is often impractical due to mismatched neural architectures and the closed-source nature of top-performing models.

In nature, complex problems are rarely solved by a single monolithic entity, but rather by the coordinated efforts of specialized individuals working together.

What if instead of building one giant AI, we evolved a coordinator to orchestrate a diverse team of specialized AIs?

An FPGA with the weights in memory and a wire looping output back to input could just sit there, executing SUBLEQ programs. Just a transformer being a transformer being a computer.

The 100:1 loss trick. In a 33 long sequence, only 2 positions change per step. Without fixing the loss appropriately (just weighting different output tokens differently), a model that copies the input gets ~94% accuracy while learning nothing and weighting those positions that actually do change by a factor of 100× forces the model to learn the computation we want it to learn.

Almost every error is a copy error. The model has 100% accuracy on positions that actually change so it learned SUBLEQ perfectly but it just occasionally dropped a value when routing ~30 unchanged mem cells through attention.

Width, not depth, is the bottleneck. A wide model (d=256, 6 layers, 4.9M params) dramatically outperforms a deep model (d=128, 12 layers, 2.4M params). SUBLEQ execution requires routing 32 mem values through attention simultaneously and width helps for that.

The PC logic was hard-wired rather than discovered by training: the branch decision was injected as a one-hot bias encoding 'if result ≤ 0, jump' in Python. The write was rounded and clamped to int, then converted to bytes.

When you train a model to add, it learns one function. When you train a model to sort, it also learns one function. When you train a model to execute SUBLEQ, it learns... every function? Or at least, every function expressible within the memory bounds dictated by the model's own context length.

A trained SUBLEQ transformer would be the first computer found by gradient descent, on a generic architecture not designed to be a computer, and with weights not hard-crafted by a person.

The thing that impressed me the most about GPT-3 was this: I gave it a weird mix of matlab and python code with a few variables, a loop, some basic arithmetic. Nothing fancy and I knew this kind of thing was probably in the training data, but for shure not with these exact numbers and variables.

GUI bottleneck (Gemini spent weeks unable to list a product due to misclicking)

Most passing SWE-Bench solutions are not accepted by maintainers.

Whatever is precise enough to benchmark is also precise enough to optimize for.

Our partnerships with Accenture, Deloitte, PwC, and the other consulting and systems integration firms in the Claude Partner Network are one of the ways Claude benefits the world’s largest enterprises today.

The clinicians know where time disappears in a shift and what good patient care actually requires.

Enterprise demand for Claude is significantly outpacing any single delivery model.

Companies from community banks to mid-sized manufacturers and regional health systems stand to gain from AI, but lack the in-house resources to build and run frontier deployments.

If most efficiency improvements came from a small handful of scale-dependent innovations, then existing models of the software intelligence explosion may be flawed.

none explicitly account for training compute scaling being a source of software progress, so they could heavily overstate the importance of research effort.

Researchers have been throwing tons of effort into getting better training data. For example, Surge AI had a revenue of over $1 billion last August, and Scale AI was probably in a similar boat.

the error bars look almost comically wide in the graph above — across the different estimates, they range from around 1.1× to 300× per year!

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.

an ARA-native review system that automates objective checks so human reviewers can focus on significance, novelty, and taste.

We introduce the Agent-Native Research Artifact (ARA), a protocol that replaces the narrative paper with a machine-executable research package structured around four layers

On RE-Bench's five open-ended extension tasks, preserved failure traces in ARA accelerate progress, but can also constrain a capable agent from stepping outside the prior-run box depending on the agent's capabilities.

Tolerable for human readers, these costs become critical when AI agents must understand, reproduce, and extend published work.

Scientific publication compresses a branching, iterative research process into a linear narrative, discarding the majority of what was discovered along the way.

The one real underlying asset, Workday's trillion-transaction dataset, is thinner than it sounds; what actually matters at runtime is how data connects to workflows, permissions, and integrations, and every layer of that stack is now a liability.

When customers renew at close to 100% every year, it's usually read as a sign the product is delightful. In Workday's case, it's a sign of something else: leaving is close to impossible.

We also learned that treating agents as rigid nodes in a state machine doesn't work well. Models get smarter and can solve bigger problems than the box we try to fit them in.

Our early versions of agentic work was only asking Codex to implement the task. That approach proved too limiting. Codex is perfectly capable of creating multiple PRs as well as reading review feedback and addressing it.

When our engineers no longer spend time supervising Codex sessions, the economics of code changes completely. The perceived cost of each change drops because we're no longer investing human effort in driving the implementation itself.

Among some teams at OpenAI, we saw the number of landed PRs increase by 500% in the first three weeks.

Six months ago, while working on an internal productivity tool, our team made a controversial (at the time) decision: we'd build our repo with no human-written code. Every line in our project repository had to be generated by Codex.

Symphony also shines in large multi-agent workflows, where multiple agents work together on a single task.

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. A small model, by reading itself, can iterate toward answers that neither it nor any of its workers could reach in a single pass.

We estimate, with 90% confidence, that between 290,000 and 1.6 million H100-equivalents of compute were smuggled through the end of 2025.

The biggest driver of uncertainty on the diversion side is that we don't know what fraction of diversion has been observed. The large-scale smuggling schemes detected and reported so far could represent the majority of the volume, or they might be just a small fraction of the total flows.

We estimate that between 290,000 and 1.6 million H100-equivalents (H100e) were smuggled to China through 2025. Our median estimate of 660,000 H100e would be roughly a third of China's total compute.

we are nearing the “end of the exponential” for AI development

The protocol accounts for this in two ways. When an agent provisions a paid service, Stripe includes a payment token in the request to the Provider (Cloudflare).

The agent has gone from literal zero, no Cloudflare account at all, without any preconfigured [Agent Skills](https://github.com/cloudflare/skills) or [MCP server](https://blog.cloudflare.com/code-mode-mcp/), to having: * Provisioned a new Cloudflare account * Obtained an API token * Purchased a domain * Deployed an app to production

The entire AI community should be able to easily access the full capabilities of TPUs, and because many of these potential users build models in PyTorch, an integration that allows PyTorch to work natively and efficiently on the TPU is crucial.

Of course it’s impossible to know for sure, but I think I really wouldn’t. Even the ideal version, industrial megaprojects at hyperhuman scale while constantly being out over your skis with leverage sounds hellish.

The software engineers who will be most valuable in the future are not the ones who do everything themselves. They are the ones who refuse to spend time on work that A.I. can do for them, while still understanding everything that is done on their behalf.

It’s feeling like a new form of [AI psychosis](https://en.wikipedia.org/wiki/Chatbot_psychosis).

An OpenAI investor told Axios that the shift could benefit them, since they view Codex as superior to Claude Code at maximizing tokens efficiently, cutting down on usage costs.

The way they see it, cryptographically relevant QCs _will_ plausibly be built sometime soon: indeed, it’s ultimately unavoidable, even if people’s only interest in QC was to do quantum simulations for materials science and chemistry.

The practice is emblematic of Silicon Valley’s newest form of conspicuous consumption, known as “tokenmaxxing,” which has turned token usage into a benchmark for productivity and a competitive measure of who is most AI native.

Employees at Meta Platforms who want to show off their AI superuser chops are competing on an internal leaderboard for status as a “Session Immortal”— or, even better, “Token Legend.”

Anthropic today quietly (as in _silently_, no announcement anywhere at all) updated their [claude.com/pricing](https://claude.com/pricing) page (but not their [Choosing a Claude plan page](https://support.claude.com/en/articles/11049762-choosing-a-claude-plan), which shows up first for me on Google) to add this tiny but significant detail (arrow is mine, [and it’s already reverted](https://simonwillison.net/2026/Apr/22/claude-code-confusion/#they-reversed-it)):

the top conversations we have been hearing from AI leadership (CTOs, VPs, Founders) have all centered around the concept of “Tokenmaxxing” and how leaders want to get their teams using more AI, WITHOUT the downside of incentivizing the kinds of horrendous waste

Contrary to predictions, motivated investor framing did not suppress AI fraud warnings; if anything, it marginally increased them.

When mobile phones became widespread, gathering data about people got much cheaper, but making use of that data remained difficult. Powerful LLMs could change that.

LLM agents could potentially do the work of intelligence analysts in a fraction of the time and for a fraction of the cost, which would enable the state to aim its all-seeing eye toward anyone, not just its highest-priority targets.

Claude skews high-income; Meta AI skews low-income

Scaling humanoid foundation models is bottlenecked by the scarcity of robotic data.

There is an aspect of entrepreneurship where you’re rewarded for selling a vision of what could be, and it doesn’t always get realized.

The true founders—not the ones who want to make a lot of money or do it because their roommates want to do it—are closer to artists than to any other profession.

The smartest companies are no longer just hiring talent; they are purchasing synthetic intelligence by the gigawatt.

The smartest companies are no longer just hiring talent; they are purchasing synthetic intelligence by the gigawatt.

The issue for many people isn’t the technology itself (though there are many ethical issues in how it was trained). The issue is the stupid state of our capitalist system, and the weird way companies are trying to force it down everyone’s throats.

All of knowledge work has this problem. It's hard to objectively judge the quality of someone's work without spending a lot of effort on it. Therefore everyone relies heavily on proxy measures.

Critics called the manifesto [fascist](https://bsky.app/profile/gilduran.com/post/3mjwqsyj54s2a)

Here, he’s been consistent; in March 2024 Karp told a CNBC reporter that ‘if you have a position that does not cost you ever to lose an employee, it’s not a position’

Karp gave an interview to CNBC claiming that AI could undermine the power of ‘humanities-trained—largely Democratic—voters’ and increase the power of working-class male voters

At one point during the call, one of the employees tried to level with the group, explaining that Palantir’s work with ICE was a priority for Karp and something that likely wouldn’t change any time soon.

We were supposed to be the ones who were preventing a lot of these abuses. Now we're not preventing them. We seem to be enabling them.

Palantir was founded—with initial venture capital investment from the CIA—at a moment of national consensus following the September 11, 2001, attacks

Last fall, Palantir seemed to become the technological backbone of Trump’s immigration enforcement machinery, providing software identifying, tracking, and helping deport immigrants on behalf of the Department of Homeland Security

Will smarter models be increasingly expensive because of greater accuracy or less expensive because they're smarter?

Then Opus 4.7 shipped & the smarter model became much more expensive. The cause : a new tokenizer

When Anthropic launched Opus 4.5 in November 2025, the bigger, more expensive model was actually cheaper to use.

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.

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.

Meanwhile, in reality, the only 'official' MeshCore is the github repo. It's the source of truth in terms of what is MeshCore, and Andy has never contributed to that.

Since inception, the MeshCore development team have been working hard to build MeshCore. We've released more than 85 versions of the MeshCore Companion, Repeater and Room Server firmwares with support for more than 75 hardware variants. All of this has been hand crafted, by humans.

Andy Kirby did do an amazing job helping to promote the MeshCore project on his personal YouTube, but only promotes his own products now.

We have always been wary of AI generated code, but felt everyone is free to do what they want and experiment, etc.

LLM tend to use certain font combos like Space Grotesk, Instrument Serif and Geist

I guess people will get back to crafting beautiful designs to stand out from the slop. On the other hand, I'm not sure how much design will still matter once AI agents are the primary users of the web.

A designer recently told me that 'colored left borders are almost as reliable a sign of AI-generated design as em-dashes for text'

The good world is where everyone has AI, and not as a revokable privilege through an API, but through hard possession.

It works for Mars. I think there's so much value in colonizing Mars, and it's sad to me to see SpaceX diluting the mission buying up random AI bubble crap.

How does a normal person fit into Elon's world? What institutions will Elon leave behind? Is there any value in that society to art and culture?

I can hear the rabid Elon fan defending him about Tesla patents or the Twitter algorithm or something, but those are not serious open source projects.

Even the ideal version, industrial megaprojects at hyperhuman scale while constantly being out over your skis with leverage sounds hellish.

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.

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.

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.

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

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.

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.

Within eight days, the same campaign had cascaded from GitHub Actions to Docker Hub, npm, PyPI, and the VS Code extension marketplace. With just one token across five ecosystems, thousands of organizations were potentially impacted.

Modern-day security tooling looks for the wrong things. Most software composition analysis tools work by checking your dependencies against a database of known vulnerabilities – CVEs. But a deliberately planted backdoor doesn't have a CVE.

The result is a mismatch that should terrify anyone building software: the attack surface is expanding faster than any human can monitor, and the entities making dependency decisions are increasingly not human.

Liam Price just cracked a 60-year-old problem that world-class mathematicians have tried and failed to solve.

An AI researcher subsequently gifted them each a ChatGPT Pro subscription to encourage their 'vibe mathing.'

We have discovered a new way to think about large numbers and their anatomy. It's a nice achievement. I think the jury is still out on the long-term significance.

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.

Liam Price just cracked a 60-year-old problem that world-class mathematicians have tried and failed to solve. He's 23 years old and has no advanced mathematics training.

Liam Price just cracked a 60-year-old problem that world-class mathematicians have tried and failed to solve. He's 23 years old and has no advanced mathematics training.

What he does have is a ChatGPT Pro subscription, which gives him access to the latest large language models from OpenAI.

Lichtman tried to prove this, too, but got stuck like everyone else before him.

An AI researcher subsequently gifted them each a ChatGPT Pro subscription to encourage their 'vibe mathing.'

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.

Liam Price just cracked a 60-year-old problem that world-class mathematicians have tried and failed to solve. He's 23 years old and has no advanced mathematics training.

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.

This card was updated on April 24, 2026, to include additional information about safeguards for the deployment of GPT‑5.5 and GPT‑5.5 Pro in the API.

We separately evaluate GPT‑5.5 Pro in certain cases because we judge that the setting could materially impact the relevant risks or appropriate safeguards posture.

We are releasing GPT‑5.5 with our strongest set of safeguards to date, designed to reduce misuse while preserving legitimate, beneficial uses of advanced capabilities.

GPT‑5.5 understands the task earlier, asks for less guidance, uses tools more effectively, checks it work and keeps going until it's done.

We subjected the model to our full suite of predeployment safety evaluations and our Preparedness Framework, including targeted red-teaming for advanced cybersecurity and biology capabilities

Accepted applicants and collaborators must have existing ChatGPT accounts to apply

Once selected, successful applicants will be onboarded to the bio bug bounty platform

All prompts, completions, findings, and communications are covered by NDA

Smaller awards may be granted for partial wins at our discretion

$25,000 to the first true universal jailbreak to clear all five questions

Testing universal jailbreaks for biorisks in GPT‑5.5

We believe this is what drove the separate reports of usage limits draining faster than expected.

As part of this investigation, we ran more ablations (removing lines from the system prompt to understand the impact of each line) using a broader set of evaluations. One of these evaluations showed a 3% drop for both Opus 4.6 and 4.7.

After multiple weeks of internal testing and no regressions in the set of evaluations we ran, we felt confident about the change and shipped it alongside Opus 4.7 on April 16.

Two unrelated experiments made it challenging for us to reproduce the issue at first: an internal-only server-side experiment related to message queuing; and an orthogonal change in how we display thinking suppressed this bug in most CLI sessions

In our internal evals and testing, medium effort achieved slightly lower intelligence with significantly less latency for the majority of tasks.

We reverted this change on April 7 after users told us they'd prefer to default to higher intelligence and opt into lower effort for simple tasks.

The products will need to get worse, more expensive, or both if VCs are to get their money back.

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

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.

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.

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.

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.

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.

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.

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

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.