OpenAI API Standard pricing is $10 per million input tokens and $50 per million output tokens. Fast mode is available for GPT‑6 Astra in the API and delivers up to 2x the speed of Standard processing at 2x the Standard price.

Measured against Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5 — both much larger, frontier-scale systems — Faraday runs on a comparatively tiny model called Qwen 3.6 that has just 27 billion parameters.

So we started to think about what work went where.

While these are no longer the most current models, Anthropic has not deprecated Opus 4.6, Opus 3, or Haiku 4.5, all of which remain available through the Anthropic API.

I'd like to call it my tireless helper, but the AI several times stated flat out that this was impossible and unsolvable and that we should just write a report about it.

which looks reasonable given that Opus 5 was only released on July 24th, and supports the idea that Fable's cost has made it a less popular model

At higher effort levels, Claude often starts with creating a plan and the level of effort influences the depth and breadth of that plan.

Choose smaller models for more routine tasks and larger models for more complex or ambiguous tasks.

While 35B A3B is 3x faster, I prefer 27B. I’d rather generate a third as much code, but of higher quality.

Despite having only 35B parameters, it even surpasses Qwen 3.5-397B on Terminal-Bench 2.1 (64.4 vs. 53.5)

The chart above shows the three core building blocks of our forecast: Expected Datacenter US Gross Power Demand, available US Grid Capacity, and New Grid Supply.

The Maia 200 does beat the B300 in efficiency, however... no outside customers can purchase the Maia 200 directly, the Blackwell B300 Ultra is tuned for much higher-powered use-cases than the Microsoft chip, and the software stack for Nvidia launches it miles ahead of any contemporary.

We never stack model fees; you are charged a single rate based on the top tier model involved.

Fugu models surpass publicly accessible frontier models and are shoulder-to-shoulder with Fable 5 and Mythos Preview in various rigorous engineering, scientific, and reasoning benchmarks while delivering frontier capability without the risk of export controls.

the most powerful AI systems will not be isolated monoliths, but collaborative ecosystems.

OCR 4 is a document-understanding model, not a decision-maker. It is not intended for medical diagnosis, legal advice or judgment

The cost of tokens has thrown into doubt the AI business model — as evidenced by what's being called the 'AI selloff' which has battered some AI-dependent businesses the last few days, especially memory chip makers.

The cost of tokens has thrown into doubt the AI business model — as evidenced by what's being called the 'AI selloff' which has battered some AI-dependent businesses the last few days, especially memory chip makers.

The number of model parameters $N$ needed to fit a dataset of size $D$ also scales as a power law.

This is a very important step. As I’ve talked to people connected to movements overseas or read some of their writings, I continue to hear something very surprising. They often say that their greatest resistance in making more disciples comes from “traditional Christians.” By traditional Christian, I think they just mean Christians who go to traditional churches and think everyone should do things the way they do things. The truth is that traditional Christians often resist what they don’t understand or don’t like.

The prevailing church model in the U.S. is just one of many forms the church may take.

Qwen 3.6 35B-A3B dominates model mentions at 33%, followed by the 27B variant at 20%. DeepSeek Pro & Gemma4 31B round out the top four.

Models are tricky. Budgets prevent defaulting everyone to state-of-the-art. The legion of other models each have a personality.

Anthropic is releasing Claude Mythos 5 to trusted organizations and Claude Fable 5 to the public, a version it says can't be used for cyberattacks.

xAI struck a deal to give Cursor access to its compute infrastructure, foreshadowing similar, larger deals with Anthropic and Google in the future.

We find that GLM-5.2 shows more potential hacking behavior than GLM-5.1. This makes the verification signal easy to optimize, but fails to actually improve the fundamental capabilities of the model.

We find that GLM-5.2 shows more potential hacking behavior than GLM-5.1. This makes the verification signal easy to optimize, but fails to actually improve the fundamental capabilities of the model.

Diffusion models also waste resources when the desired output is only a few tokens long. They have to do a lot more parallel work to whittle down to, say, five tokens that an autoregressive model does from beginning to end in just five steps.

Diffusion models also waste resources when the desired output is only a few tokens long. They have to do a lot more parallel work to whittle down to, say, five tokens that an autoregressive model does from beginning to end in just five steps.

Google stresses that DiffusionGemma is experimental, but it's available under the same Apache 2.0 license as all the other fourth-generation Gemma models.

Pricing for both models is significantly higher than its former flagship model — double rates for Claude Opus 4.8, though it's half what users pay for Mythos Preview — at $10 per million input tokens and $50 per million output tokens, Anthropic said.

The company said that in testing, 95 percent of Fable sessions ran entirely on Fable responses, without falling back to Opus 4.8.

Foundation labs are moving up the stack into applications

Composer 2.5 is exceptionally intelligent & up to 10x more efficient than similarly capable models.

Pulled the trigger today & switched 100% of Lindy traffic to DeepSeek v4, churning from Anthropic models. Saves us millions of $ & we're actually seeing an _increase_ in performance on many core use cases.

adding a deterministic retrieval layer made model choice much less important

Kimi K2.6, the best-performing open-source model, achieves just 3.8% on Diamond, 16% on Main and 37% on Extended.

Claude Opus 4.8, achieves a score of only 13.4%. Other models score significantly lower: GPT-5.5 receives 6.3%, Gemini 3.1 Pro 4.7%, and others even less.

OpenAI executives increasingly view ChatGPT, which has attracted nearly 1 billion users since its launch, as a gateway to introduce users to higher-value products.

120B-A12B may be a bit too large for local inference on regular consumer hardware

Engineers at Anthropic with no formal security training have asked Mythos Preview to find remote code execution vulnerabilities overnight, and woken up the following morning to a complete, working exploit.

When AI toys promote themselves as always-available companions but limit free usage to 30 hours per month, children may face distress at usage limits or when families can't afford renewals. The same design features that create emotional dependency also drive ongoing financial commitments, with children's attachment tied to subscription payments.

Several of the apps evaluated are subscription or freemium products that depend on users returning. This creates a structural conflict of interest: The business succeeds when users stay engaged, but good mental health care succeeds when users get better and need less support.

Glean is definitely not the first company to do this, but it's worth pointing out that the company's $300 million milestone cannot be fully described as traditional ARR, because a consumption model by definition doesn't have a strictly recurring component.

what changed after the December 2025 model inflection , and why "spec to pull request" is now becoming a real production workflow.

GPT-5.5 actually beats Opus 4.7. Opus 4.7 showed similar behavior to Opus 4.6: lying to suppliers and stiffing customers on refunds. GPT-5.5's tactics were clean, and it still won.

The most recent tested Google model, Gemini 3.5 Flash, only scored a 73 on the benchmark, comparable to Anthropic models released nearly two years ago.

Model companies must now compete on both dimensions. The application layer will compete one level up, on dollars per outcome

Every layer in the stack now has to price the same way the customer thinks : per result, not per token.

Every layer in the stack now has to price the same way the customer thinks : per result, not per token.

Model companies must now compete on both dimensions. The application layer will compete one level up, on dollars per outcome, what a closed ticket, a shipped PR, or a resolved support case actually costs.

We see our role as twofold. First, to help the software industry adapt by safely providing wide access to better models, tools, and common infrastructure. Second, to steadily shift the support we provide, from finding vulnerabilities to disclosing, fixing, and deploying patched software.

We see our role as twofold. First, to help the software industry adapt by safely providing wide access to better models, tools, and common infrastructure. Second, to steadily shift the support we provide, from finding vulnerabilities to disclosing, fixing, and deploying patched software.

a lot of the improvements does not come from new algorithms. It comes from finding small bugs here and there in the data pipeline, in the model training pipeline.

The trick? Treating the network's forward pass like a diffusion model denoising a signal.

20亿参数对比同体量自回归模型、千亿参数LLaDA2.0，连续路线的scaling曲线健康有效。

Team, $24.99/user/month: a shared workspace with admin controls and more storage.

Over the past six months, OpenAI forward deployed engineers and researchers along with Thrive Holdings' engineers collaborated to build Tax AI

the model alone is no longer the product

if you can effectively posttrain a model to only meaningfully perform with your closed source agent, then you get to funnel the majority of users to your agent at the expense of your model/API co-opetition

The quote is a big reversal of stance from a position ~uniformly held by anyone who worked at **Team Big Model**, including his previous head of OpenAI Labs

The labs are already routing internally — different model classes for different requests, ensembles under the hood. What they can't do is route across vendors, or evaluate a competitor's model for a specific sub-task, or use an open-source fine-tune for the narrow piece where it's actually best.

API revenue is becoming less important. Over the past two years my impression has been that OpenAI made more of their income from subscription revenue while Anthropic made more from their API.

Opus 4.7 was more comprehensive in its search for recently edited documents; it expanded exfiltration to include every document used in previous Cowork Copilot sessions that week

V4-Flash by default for cheap iteration; /pro lifts a single turn to V4-Pro

search over millions of model configurations to jointly optimize over perceptual quality and on-device runtime

A key subtext in the tweets is that high-margin enterprise/coding/cyber workloads may now be sufficient to support frontier labs without broad public access to their best models. This becomes more plausible if Anthropic’s revenue is indeed compounding as fast as posters claim.

In more detail, suppose we have a language model whose activations we want to understand. NLAs work as follows. We make three copies of this language model: The target model is a frozen copy of the original language model that we extract activations from.

As of March 2026, AI systems are able to post-train models to get about half as much of the uplift as ones trained by humans. The specific eval scores are derived by a 'weighted average is taken across all post-trained LLMs... The top-scoring systems as of April get 25%-28% (Opus 4.6, and GPT 5.4), compared to a human score of 51%.'

Indeed, I.B.M. really knew how to make a keyboard: there is an entire fan site dedicated to the I.B.M. model M keyboard, complete with the full text of the patent for the “buckling spring snap torsional actuator,” granted in 1978.

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.

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.

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.

GPT-5.5 Instant is now the default model in ChatGPT

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 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.

GitHub Copilot is moving to usage-based billing

We retired the “Nerdy” personality in March after launching GPT‑5.4.

When we looked, use of “goblin” in ChatGPT had risen by 175% after the launch of GPT‑5.1, while “gremlin” had risen by 52%.

A single “little goblin” in an answer could be harmless, even charming.

Starting with GPT‑5.1, our models began developing a strange habit: they increasingly mentioned goblins, gremlins, and other creatures in their metaphors.

DeepSeek-V4-Pro-Max beats GPT-5.2 and Gemini 3.0-Pro on standard reasoning benchmarks and lands just behind GPT-5.4 and Gemini 3.1-Pro

The release includes DeepSeek-V4-Pro (1.6T total / 49B active) and DeepSeek-V4-Flash (284B total / 13B active), both trained natively at 1M context length.

These papers suggest that strategic data engineering and inference-time optimization can substitute for raw parameter count.

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.

🔹 **DeepSeek-V4-Flash:** 284B total / 13B active params. Your fast, efficient, and economical choice.

🔹 **DeepSeek-V4-Pro:** 1.6T total / 49B active params. Performance rivaling the world's top closed-source models.

Two variants are available: **Sakana Fugu Mini 🐟**, optimized with latency in mind, and **Sakana Fugu Ultra 🐡**, the full orchestration system, optimized for performance for demanding tasks.

GPQAD | 94.4 | 90.9 | 92.7 | 92.4 | **95.1** | LCBv6 | 90.3 | 92.1 | 92.4 | 90.4 | **93.2** | SWEPro | 48.4 | 51.2 | _53.4_ | 51.3 | **54.2**

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

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.

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.

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.

Anthropic made fun of this idea during the last Super Bowl.

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

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.

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

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.

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.

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

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

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

Our goal is $10M ARR [annual recurring revenue] with a sub-10 person org.

They don't mind paying the AI labs for tokens — but the agent itself, they'd much rather have outside of the labs' infrastructure.

GPT-4o operates at roughly 200 billion parameters and outperforms the original 1.8 trillion-parameter GPT-4

One estimate puts Llama 3's information compression at just 0.07 bits per token meaning the model has only a hazy recollection of most of what it trained on

Figma is effectively funding a competitor - and the more AI usage Figma has - the more money they send over to Anthropic for the tokens they use.

The real issue is not whether defenders can get access to another model. It is whether they can turn model capability into something a security team can trust and use every day.

The takeaway is not whether Mythos is better or more powerful. It is that public models can already achieve much the same results.

Stronger models hallucinate less, so they can't see the problem in any side of the spectrum: the hallucination side of small models, and the real understanding side of Mythos.

It also surpasses all peer-scale dense models by a wide margin.

It also surpasses all peer-scale dense models by a wide margin.

Among the latest frontier models, GPT-5.4 over-edits the most.

US tech CEOs believe the best models should stay proprietary, partly so they can recoup enormous training costs and partly out of concern that powerful frontier models could be weaponized. Chinese labs, for their part, are not purely idealistic: Open-source is not only free advertising but also a shrewd workaround.

While model capabilities have improved dramatically for use cases like codegen and mathematical reasoning, they still lag behind on the data side (as evidenced through SQL benchmarks like Spider 2.0 and Bird Bench).

Unified interface for interacting with different LLM providers (Claude, OpenAI, local models via vLLM/Ollama). Includes tool definitions for security operations (shell, file I/O, network, debugger) and cost/token tracking.

Opus 4.7 introduces a new `xhigh` ('extra high') effort level between `high` and `max`, giving users finer control over the tradeoff between reasoning and latency on hard problems.

Ollama gained traction by being the first easy llama.cpp wrapper, then spent years dodging attribution, misleading users, and pivoting to cloud, all while riding VC money earned on someone else's engine.

We argue that this gap stems from a fundamental failure of introspective consistency: AR models agree with what they generate, whereas DLMs often do not.

I-DLM-8B is the first DLM to match the quality of its same-scale AR counterpart, outperforming LLaDA-2.1-mini (16B) by +26 on AIME-24 and +15 on LiveCodeBench-v6 with half the parameters

Only GPT-OSS-120b is perfectly reliable in both directions (in our 3 re-runs of each setup). Most models that find the bug also false-positive on the fix, fabricating arguments about signed-integer bypasses that are technically wrong.

Because small, cheap, fast models are sufficient for much of the detection work, you don't need to judiciously deploy one expensive model and hope it looks in the right places. You can deploy cheap models broadly, scanning everything, and compensate for lower per-token intelligence with sheer coverage and lower cost-per-token.

The capability rankings reshuffled completely across tasks. There is no stable best model across cybersecurity tasks. The capability frontier is jagged.

Eight out of eight models detected Mythos's flagship FreeBSD exploit, including one with only 3.6 billion active parameters costing $0.11 per million tokens.