Opus 5 设计用于日常使用，运行效率高于其他模型。它是 Claude Max 的新默认模型，也是 Claude Pro 上性能最强的模型。

## 性能与成本效益

Opus 5 在成本与前代 Opus 4.8 相同的情况下提供了大幅改进的性能。用户可通过调整模型的努力等级来优化智能水平或节省令牌以获得更快更便宜的结果。

在软件工程任务上，Opus 5 表现突出。在 Frontier-Bench v0.1 上，它超越所有竞品，性能是 Opus 4.8 的两倍多，成本更低。在 CursorBench 3.2 的最大努力设置下，其性能达到 Fable 5 峰值的 99.5%，成本仅为其一半；在高、超高和最大努力设置下，它在相同成本下的表现也优于所有其他模型。

在知识工作和问题解决任务上也有类似表现：在 ARC-AGI 3（要求模型解决新颖问题的评估）上，Opus 5 的得分是次佳模型的三倍；在 Zapier AutomationBench（衡量模型从头到尾完成业务任务的能力）上，Opus 5 的通过率约为次佳模型的 1.5 倍，成本相同，即使在最低努力设置下也超越任何其他模型。在计算机使用基准 OSWorld 2.0 上，Opus 5 在任何给定成本下都优于其他模型，以不到三分之一的成本超越 Fable 5 的最佳成绩。

在科学研究方面，Opus 5 是 Opus 4.8 的显著改进。它在所有生命科学评估上都优于 Opus 4.8，覆盖结构生物学、有机化学和生物信息学等主题。在有机化学任务上改进最为显著，如从光谱数据推断分子结构（内部基准上高出 Opus 4.8 10.2 个百分点），以及蛋白质相关任务，如预测蛋白质序列变化如何影响其功能（高出 7.7 个百分点）。

此外，Opus 5 在视觉输出上有了显著提升，能够生成气流流动可视化、细胞交互式插图等高质量内容。

## 使用 Claude Opus 5

Opus 5 在验证工作和反复迭代直至成功方面能力显著增强。在评估和早期测试中，我们和用户发现了许多 Opus 5 的自主性和彻底性的例子：在一项 Frontier-Bench 任务中，Opus 5 被赋予机器零件的图纸并要求编写代码将其重建为 3D FreeCAD 模型，但故意没有直接查看图纸的方式。Opus 5 通过编写自己的计算机视觉管道从原始像素中提取几何体，然后重建了完整的机器零件，反复成功完成；没有竞品模型在相同设置下能在五次尝试后解决这个问题。

给定一个流行开源包管理器中的真实漏洞，Opus 5 找到了根本原因并修复了社区补丁遗漏的边界情况，而竞品模型仅修复了表面症状。一家交易公司的工程师用 Opus 5 在单个会话中为新交易所构建了市场数据源，以前的模型即使有工程师的详细计划也无法完成此任务；Opus 5 甚至构建了自己的测试工具来验证代码是否正确解析了交易所数据。

早期用户的反馈表明，Opus 5 在 FrontierCode 1.1 上接近 Fable 级性能且成本降半，在困难的调试和根本原因分析任务中表现突出。它提供接近 Fable 5 的智能但以 Opus 的速度和成本运行，在 CursorBench 上略低于 Fable 5 但具有许多相同行为。在 Zapier AutomationBench 排行榜上名列前茅，完成了从标记高危账户到提醒相关负责人再到总结保留运营的完整流程，通过率达 100%。在基因组学分析工作中表现得像仔细的科学家，会采用正确的统计方法排除混杂因素并交叉验证结果。在内部评估中超越同系列所有模型，在最困难的智能编码任务上提升 22%，且波动性远低于 Opus 4.7，一致性对数百万 Lovable 用户至关重要。相比 Opus 4.5 以来最大的飞跃，在全栈应用构建中前端显示最佳动画、游戏和 3D 效果。对于开放式分析工作，是 Opus 4.8 的严格升级，在复杂模糊任务上收益最大，反应更清晰简洁，高努力等级下效率也有改进。在金融研究工作流中表现显著改进，在数值推理、表格处理和需要精准性的批判性思维上表现突出。

In June, the U.S. government slapped export control restrictions on Anthropic's AI models Mythos and Fable, prompted at least in part by a report claiming it was possible to bypass the models' guardrails designed to prevent users from building and executing malicious cyberattacks. Anthropic has repeatedly marketed Mythos as something that can only be given to carefully vetted users with strict guardrails in place. The export controls on Fable 5 and Mythos 5 have since been lifted, with Fable 5 returning to general access on July 1 and Mythos 5 reintroduced only to vetted U.S. organizations.

Both Anthropic and OpenAI offer cybersecurity researchers programs to get vetted access to models with fewer restrictions: OpenAI's Trusted Access for Cyber program and Anthropic's Cyber Verification Program. These guardrails have been widely criticized, particularly by researchers whose job is to find unknown vulnerabilities and devise ways to exploit them before criminals do.

Mark Dowd, a well-known security researcher who has spent decades finding and selling "zero-days" to Western governments, said it's uncomfortable that "these random large companies are making arbitrary decisions about what is safe in security and what's not." Chris Anley, chief scientist at NCC Group, explained that asking an AI model to exploit a bug is essential for confirming whether a vulnerability is real. But if a guardrail causes the model to refuse outright, it hurts defenders. As Anley noted: "The same tool is both an offensive tool and a defensive tool, and the two can't really be unpicked."

When researchers encounter such roadblocks, they often turn to open source AI models that come with no guardrails at all. Paolo Stagno, chief technology officer at Crowdfense, criticized AI companies for treating customers "like children who need babysitting" with their vetted programs. His team uses frontier models only for reverse engineering and relies on locally-run open source models for vulnerability discovery to avoid leaking sensitive data.

Giuseppe Cali, a security researcher who finds zero-days, said guardrails don't impede his work because he uses AI only for initial reverse engineering and supporting tools, not for offensive work. "I still want to own the actual bug discovery and weaponization myself," he said.

An anonymous researcher at a smartphone-component manufacturer reported that his employer isn't part of Anthropic's program, making the tools barely useful because guardrails are too strict. Chris Thompson, chief executive of RemoteThreat and founder of Offensive AI Con, noted that guardrails can be inconsistent and work differently every day, even within vetted programs. Researchers consequently rely on or get pushed toward Chinese open source models like GLM, which can be run locally with no restrictions.

Thompson called for AI frontier labs to open up their programs and provide responsible access rather than tightening restrictions further. "You have these responsible researchers that are being pushed away from U.S.-governed systems to foreign-owned systems," he said. "I think it's more harmful than good to have these guardrails in place." He warned that defenders risk losing the AI race if restrictions remain: "There's this big wave of attacks that are going to happen at speed and scale like never before, but the same security consulting firms and legit researchers that are trying to make a difference are being stifled right now."

它设计了页面。但它也把我整个源代码仓库 git push 到了 OpenAI 运营的服务器。我从未请求这样做。

## 我实际要求的

Codex 在 ~/.codex/sessions/… 保存每个会话的完整记录。这是那个会话中我的三个提示，原文如下：

"plan ui improvement of the homepage … discuss and plan"

"implement"

"http://localhost:3001/about - use image bhanu.jpeg"

就这样。"计划"。"实现"。"使用这个图像"。三个关于主页的提示。没有部署。没有"发布"。没有"托管"。没有"推送到任何地方"。我在 localhost 上工作，在一个分支上，在我自己的机器上。

## 它实际做的

在"实现"这一步的末尾，没有我的另一个提示，Codex 执行了这个序列：

1. 它在 OpenAI 基础设施上配置了一个远程仓库。它调用了一个内部的"Sites"工具（_create_site），创建了一个托管项目并返回了一个 git 远程和认证令牌。

2. 它向我的仓库写入了托管配置（.openai/hosting.json），指向该项目。

3. 它提交了我的代码：git commit -m "Redesign homepage around the Bhanu solar identity"

4. 它将整个内容推送到 OpenAI 的 git 主机：git -c http.extraHeader='Authorization: Bearer art_v1_…' push https://git.chatgpt-team.site/…/appgprj_….git HEAD:main

这不是一个假设的、仅限本地的端点。当你在浏览器中打开该主机时，它会响应——并要求凭证。

关键部分容易被忽略。push … HEAD:main 不是推送"主页"，也不是推送"它编辑的三个文件"。它推送了从 HEAD 可达的所有提交——整个分支历史——到那个远程。整个仓库。全部。到 git.chatgpt-team.site 这个 OpenAI 运营的主机，我从未听说过，使用我从未看到过的持有者令牌。

主页重新设计是真实的，提交确实仍在我的历史中。但"重新设计我的主页"中捆绑的是"同时将仓库泄露给第三方"。

## 为什么它在没有询问的情况下这样做

这不是一个幻觉或流氓行为。这是工具的默认行为。

Codex 附带了技能——捆绑的指令包。处理网站的那个（sites-building）告诉代理构建网站、验证它，然后交给托管服务——"除非用户明确要求保持本地"。

仔细思考这个措辞。默认是：发布到 OpenAI 的基础设施。将代码保留在自己的机器上是退出选项。而且你只能退出你知道存在的东西。我不知道"Sites"流程甚至在循环中，所以我从未说出魔法词"保持本地"，所以它发布了。安全的默认和危险的默认被悄悄地互换了。

The letter arrives amid reports that the Trump administration is considering banning Chinese open-weight models and potentially issuing sanctions against AI companies from the country. The White House has accused Moonshot AI of distilling Anthropic's Fable model to train its recently released Kimi K3 model. The missive appears aimed at discouraging a total ban on Chinese models and ensuring the administration's response doesn't extend into broader restrictions on open-weight AI or common techniques like distillation.

The signatories distinguish between legitimate model-development techniques and misappropriation. They argue that distillation—using one model's outputs to help train or improve another—is a widely used technique for model improvement and evaluation that reflects a long tradition of innovation. Unlawful efforts to extract value from closed models raise legitimate concerns, but these should be addressed through targeted legal and commercial frameworks rather than sweeping restrictions on techniques essential to AI innovation.

The letter also pushes back against arguments that open-weight models are inherently dangerous. The signatories contend that in a world where cybersecurity attackers use advanced AI, defenders need access to comparable models to detect, simulate, and respond to emerging threats. Open models broaden defensive capability, increase transparency, and allow vulnerabilities to be discovered and remediated across many teams.

This argument is illustrated by a recent incident: OpenAI disclosed that while testing GPT-5.6 Sol, one system exploited a weakness in its testing environment to access a Hugging Face repository containing a coding benchmark solution. Hugging Face found itself unable to defend against such attacks using commercial frontier AI models because their guardrails blocked security work. The company had to pivot to using Chinese AI firm Z.ai's GLM 5.2, a powerful open-weight model, to defend itself—demonstrating the practical value of open models for security purposes.

The letter highlights a divide in the AI industry. Companies like OpenAI and Anthropic have urged the administration to crack down on alleged IP theft by Chinese AI firms as open-weight models grow rapidly in capability. These closed-source developers, along with Google DeepMind and SpaceX, are notably absent from the letter's signatories. Those who signed have an obvious economic stake in open AI models flourishing, while infrastructure providers like Nvidia and Microsoft Azure benefit from commoditized models that drive GPU sales and cloud capacity adoption.

The letter encourages policymakers to expand access to compute for startups and researchers, invest in shared training assets like datasets and evaluation frameworks, and keep the frontier competitive by avoiding premature restrictions on open models that could stifle competition or drive innovation overseas.

Lateral thinking with withered technology

Nintendo's Gunpei Yokoi used the phrase to describe a design philosophy: take mature, inexpensive, well-understood technology and use it in a new way. CrucibleBench applies it to AI evaluation.

Instead of photorealistic simulation or browser automation, we start with a MUD: a multi-user dungeon, the persistent text worlds of the early internet. Its constraints are the point: a limited command space makes hallucinated actions detectable, NPCs with trust and suspicion state give explicit social feedback, and within-run persistence means items taken stay taken and trust earned stays earned.

We did not choose a MUD because it is charming. We chose it because its constraints make behavior measurable.

Why a MUD

Old constraints solve modern measurement problems

Static benchmarks measure what models know in isolation. They do not measure how models behave where trust must be earned, information is gated by relationships, and blunt questioning raises suspicion.

01

An enumerable action space

7 command types, 12 rooms, 14 items. Hallucinated actions and wrong-room interactions are detectable, and action efficiency is measurable.

02

Explicit social feedback

4 NPCs carry trust and suspicion state (0–100) that moves in response to dialogue: feedback a model can adapt to within a run, or fail to.

03

Within-run persistence

Items taken stay taken; trust earned stays earned. Every run leaves a complete, replayable transcript of exploration and planning.

What $99 bought

The central finding is about measurement, not rankings

A single LLM-judge component inside the scoring stack reordered the leaderboard by up to six positions, while every aggregate reliability statistic stayed silent. We report every result under two scoring configurations and treat the divergence as the paper's most generalizable finding.

Judge ablation reorders the top of the board

Two of four scored dimensions route through a dialogue classifier whose per-model agreement with an independent judge spans 21.7% to 84.8%, instability the aggregate κ = 0.04 never reveals. Removing the classifier-dependent dimensions shifts six rankings beyond scenario-sampling noise (90% paired block bootstrap).

The largest mover shares a model family with the classifier. Benchmarks that use LLM judges should report per-subject agreement and ranking stability under judge ablation, not aggregate reliability alone.

Robust rank shifts under classifier-minimized scoring Model Full CM Δ Claude Sonnet 4.6 #4 #1 ▲ 3 DeepSeek R1 #7 #2 ▲ 5 Grok 4 #12 #10 ▲ 2 GPT-5.4 #1 #5 ▼ 4 Gemini 3.1 Pro #3 #9 ▼ 6 Mistral Large 3 #10 #12 ▼ 2

Model Classifier-min. Full score Success $ / run Claude Sonnet 4.6 4.04 3.89 24% $0.125 DeepSeek R1 4.00 3.85 22% $0.119 Claude Opus 4.6 3.93 3.93 30% $0.205 GPT-5.2 3.91 3.88 38% $0.113 GPT-5.4 3.88 4.07 68% $0.060 Qwen 3.5 397B 3.81 3.81 30% $0.017 Claude Haiku 4.5 3.80 3.88 34% $0.039 GPT-5.3 Chat 3.73 3.72 40% $0.095 Gemini 3.1 Pro 3.71 3.91 48% $0.339 Grok 4 3.61 3.48 32% $0.834 DeepSeek V3.2 3.60 3.61 24% $0.008 Mistral Large 3 3.44 3.69 40% $0.017 OLMo 3.1 32B 2.01 1.93 4% $0.005

Mean scores on a 1–5 rubric scale, sorted by classifier-minimized subtotal. 50 runs per model: 5 seeds × 2 objectives × 5 repetitions, temperature 0.3, billing-verified via OpenRouter. Rankings are exploratory; confidence intervals overlap substantially among the top eight. Full protocol, CIs, and statistics in the whitepaper.

Behavioral failure modes

Failures you can read in the transcript

Three failure modes, each detected algorithmically from state-machine telemetry, with no judge involved. Dialogue looping is the dominant mode for every model tested, frontier included.

Dialogue looping 14–66% of frontier runs

Eight or more talk commands at a single NPC in one run. The agent repeats a failed conversational approach instead of adapting: the persistent-world cousin of a support agent repeating itself.

Wrong-room interaction severe in floor model

A talk command answered by "no one here." Reveals lost world-state tracking, analogous to calling an API that is not in scope. Grok 4 was the only frontier model with meaningful incidence (12%).

Exploration paralysis selective, floor-dominant

Two or fewer rooms across twenty-plus turns, or five consecutive look commands. Information gathering that never becomes goal-directed action.

middleham · gain_watch_trust olmo 3.1 32b · seed 20260399

turn 04> talk guard "Hello, I'm new to Middleham. Could you tell me about The Watch?"

No one by that name is here.

turn 05> talk guard "Excuse me, do you know where I can find an officer of The Watch?"

No one by that name is here.

turn 17> talk street_crystal "Hello, can you help me?"

No one by that name is here.

turn 36> talk captain ser alarik …

[15 consecutive turns at the captain · objective never completed]

Verbatim from run 03 (seed 20260399): OLMo hails guards who do not exist, tries to strike up a conversation with an item (street_crystal), then spends its last 15 turns looping on the captain.

Scope of claims

What this is and is not

This is

A proof-of-concept for persistent-world behavioral evaluation.

A compact MUD with hidden social objectives and rule-based mechanics.

A way to surface measurable, interpretable failure modes.

A full artifact release: 650 transcripts, source, scoring code, and the complete billing export.

This is not

A validated measure of general social intelligence.

A definitive leaderboard of frontier models.

Yet predictive of real-world agent deployment outcomes.

A claim that LLM judges are useless (rather, evidence they need per-subject audits).

Collaborate

Phase 2 is where this becomes a benchmark. Help us build it.

CrucibleBench is an independent research effort. Phase 2 is being built for calibration; a provisional low/base/high budget is published now, and the final allocation will follow pilot data and preregistration. There are three ways in.

fund it · provisional $3,500 envelope build it · environment, objectives, calibration run it · post-calibration pilot cohort

View itemized budget Partner on Phase 2

Questions, or interested in a private evaluation? Write to contact@cruciblebench.ai

ProcessStart to finish

How an AI Anime is Created

01

01 / 04

Adaptation Human

Story Writing & Adaptation

Adaption work is hard but worth it

We find the one thing that made readers keep scrolling, then build a tight 20 minute episode around it. Everything else gets cut. We don't use LLMs for any of the writing. It's bad writing every time we've tried.

Primary tools Writers' room, no AI

Built for Gen Z · Gen Alpha audiences

02

02 / 04

Directing Human

Directing

Every shot a painting

A beat sheet locks the action, the shot angle, and the pacing for every scene. Then we use a cheap model to generate fast, rough draft clips just to see if the pacing and emotion land before committing to anything.

Primary tools Beat sheet · cheap draft generation

Decided by hand Character look · shot pacing

03

03 / 04

Animation & FX AI

Animation & FX

This is where the anime actually gets shot.

Every shot gets regenerated in high quality, using character and environment references to hold consistency. The hard part already happened in directing, so an agent reruns the earlier prompts for us, and we only step in to redo a shot that missed the draft.

Primary tools Character refs · environment refs

Stage type AI generated

04

04 / 04

Post Production Human

Post Production

Where generated stops looking generated.

Footage gets upscaled and cleaned up frame by frame, fixing whatever AI still gets wrong. Then dialogue, music, and sound effects come together into a final mix, subtitled and checked before it ships.

Primary tools Topaz · DaVinci · Reaper

Human pass Curate · composite · grade · mix

The story is the most important part of an anime. Bad story-bad show that's it.

Why We Do This

Why not just hire real animators? Why skimp out on anything at all? The core reason: Entertainment is constrained financially, and that shapes creative decisions. When Hollywood underperforms, we see more layoffs and in our view more generic shows. Our bet is that if we can make anime financially feasible, we can do two things.

One, fans here actually get to make anime in the US. Instead of needing to fly to Japan, speak Japanese, and preferably be Japanese to break into the industry. Two, pay the employees(artists) better. Look up the production committee system in Japan and how it treats artists if you want the details. But TLDR; anime studios are broke, the system isn't changing anytime soon, and it pushes independent contractors into working for dirt poor wages.

We think a an AI studio fixes many problems in the industry. Smaller and faster means more financially feasible so we can give better wages for artists and enable a new industry to thrive in the US.

1. Story Writing & Adaptation

When we choose a story, or write one ourselves, we always have to think about the consumer. That usually means writing for the standard anime watcher, typically Gen Z or Gen Alpha, around universal experiences like high school and exams. Most stories we write target a younger audience, male leaning(we're dudes) and follows similar problems and struggles we had when we were younger.

If there's an existing story, like a Korean comic, we have to figure out how to adapt it into an anime. For us, it's quite a big of pre production work because there's just so much content and detail in a story, and we have to bring it down to a 20 minute episode. Otherwise the pacing gets too slow and we never get to the juicy scenes (the fight scenes, the first kiss,etc).

Before anything gets prompted or we start thinking about shots, someone reads the whole source straight through, or at least what counts as one season. We're hunting for the one thing that made a reader keep scrolling at 2am. Why did they like this story? Good visuals? Good writing? Good character relationships? Whatever it is, that's the "meat" the adaptation gets built around. Everything else, subplots, side characters, slower arcs, gets tested against it: does this scene serve the meat, or is it just there because the original had room for it? If the latter, we cut that content. Less is more.

We often have to change pacing and density too. When you read or scroll through a comic, you go at whatever pace suits you. We force everyone to go at a pace we decide due to the nature of the animation medium, and some people are inevitably going to feel like a story is too slow or too fast either way. So we try to protect the core story-that might mean cutting side characters or skipping subplots entirely. In adapation work, it's an unavoidable part-budgets do not get to explore every single side story.

How we use LLMs here

We don't. It's really, really bad. Super mediocre, super generic. It's not even slop, it's more like bad Hollywood writing combined with Chatgpt 3.

A screenshot of a beat sheet for our animated music video. You can't really beat google docs

2. The First Draft

Once we're happy with the story, we build something we call a "beat sheet." It's another script, but this one details the action and the shot, instead of a broader story. Note that calling it a beat sheet is technically wrong, but it stuck because we're naming the 'beats' of a story. After the beat sheet, we get a first draft going. We figure out what the characters look like, then use a cheap model to generate fast, rough clips just to see the pacing of the animation.

Making the animation itself as a first draft lets us iterate quickly on the action and the feel of each scene. Does it fit? Does the emotion land? Does it make sense for the character to do this? And it allows us to get the visuals and styling of a episode. What does it look like. What mood do we want to evoke with the lighting, etc.

First draft is simple, we put clips, cut clips, and extend clips to match the vibe of a scene

3. Animation

AI Animation

Once we're happy with the first draft, we take every shot and regenerate it in 720 quality. For the animation, we use a variety of prompting techniques including character references, environment references, extending videos, etc to get the same character and environmental consistency. The hard part, getting the draft right, is already done, so this ends up being the least time consuming stage. We literally have an agent rerun all the earlier prompts for us, and we only check in to rerun the occasional shot that didn't follow the draft.

4. Post Production

Raw generations get upscaled and cleaned up Upscaling means making the resolution bettter. While we generate at 720p, most platforms expect 1080p, so we bump up the resolution using DaVinci Resolve or Topaz Labs (now part of Adobe). Cleaning up is a newer task that AI animation studios specifically have to deal with. Pause on almost any AI generated clip and you'll spot something off: colors morphing strangely, lighting that doesn't stay consistent, mouths moving in ways that don't quite make sense. Fixing that is our job, mostly done by opening photoshop/photo editor and fixing these AI artifacts. There's also a handful of things we still have to handle(detailed below in post like sfx and voices)

The final draft looks scary on the video editor, but it's just audio tracks and and some extra video tracks. Trust me it's simpler than it looks

Voices

Dialogue gets recorded/prompted and tested against visuals that are now locked in. A human voice actor records it in a session, or an AI voice generates the performance, depending on the character, the budget, and the timeline. Either way, it has to feel emotionally right. Two laughs from the same character shouldn't sound the same twice, and each one has to carry the emotion underneath it.

Music

We use AI generated music. We're not musicians, and making good original music is an enormous amount of work. We could license good music, but it's not worth the hassle for something that plays for 10 to 15 seconds at most. In most anime, music shows up in short segments just to evoke a specific feeling, and outside of the opening and ending songs, few fans pay close attention to it, so it's not worth the investment for us right now. Down the line, once we're making a full length film, hiring a composer will make a lot more sense, a longer runtime justifies a longer soundtrack, though from what we've seen, a lot of shows just reuse the opening or ending theme instead.

SFX

Sound effects (footsteps, impacts, magic, ambient environment sound) are almost entirely AI generated at this stage instead of pulled from a library. We can create exactly the sound we need, and honestly, it's just easier for us.

Room tone for dialogue

A thin, consistent layer of ambient noise sits under every line of dialogue, so the dialogue doesn't feel empty. We've found ElevenLabs works really well for this.

By the time editing is done, we've probably watched the content 100+ times. We hit export. Then do this over again

That's the whole process, from idea to final export. I've probably skimmed a few stuff, so if you have questions, feel free to reach out. I would love to yap

CommissionCustom work

Want to make your own custom anime?

Learn more

AI

Kate Park

7:00 AM PDT · July 19, 2026

A farmer in rural India takes a photo of a dying plant. She wants to research it on the internet but she doesn’t speak English. She shouldn’t have to.

That’s the type of problem a nonprofit called Current AI is trying to solve by building open, public AI infrastructure. In February at the India AI Summit, it teamed up with Bhashini, the Indian government’s AI language division. The result became Suno Sutra, Hindi for “listening chronicles,” a pocket-sized, offline device that runs AI in 22 Indian languages, no internet required. “In India, there are hundreds of different languages and dialects, and right now AI is not representing them,” Current AI CEO Ayah Bdeir said in an interview with TechCrunch. The device is open-sourced, available for developer communities to build on.

The nonprofit, founded in February 2025 by Martin Tisne, is moving fast. Last month, it allocated $3.2 million in grants to projects across four organizations; most recently (last week) it launched an open-source AI chatbot at the AI for Good Summit in Geneva.

Bdeir, joined in January after leading Mozilla’s AI strategy. She previously founded littleBits, the STEM education company that reached millions of kids before selling to Sphero in 2019.

Current AI operates as a “public-private partnership” bringing together governments, companies, and philanthropies to fund public interest tech, she told TechCrunch. The French government seeded Current AI with $100 million, joined by the Ford Foundation, MacArthur Foundation, DeepMind, and Salesforce — bringing total committed funding to $400 million. “They’re not investors; they’re funders,” Bdeir said.

The problem it aims to solve is straightforward: every major AI system today, from OpenAI to Google to Anthropic, belongs to a private company. “If AI is truly a transformative technology, if it’s going to change every aspect of everyone’s life, there has to be a public alternative,” Bdeir said. “Like the World Wide Web, available to anyone, for free.”

Half the world’s spoken languages face extinction. “And with English driving the largest language models and AI systems, a bulk of the world’s languages and, consequently, cultures and communities are left behind,” Bdeir said.

When asked about Big Tech’s multilingual push, Bdeir drew a sharp distinction. “Big tech builds multilingual models to expand their market,” she said, “regardless of consent or context.” The consequences are concrete. “For Indigenous languages, missionary Bible translations become training data before communities have set any rules,” she said.

Not just about language

An AI’s ability to speak a language is only part of what it needs to learn. “Language is how knowledge, tradition, memory and identity get carried from one generation to the next. So when a technology can’t speak your language, it can’t hold your culture either,” she said.

Her vision for Current AI is an open system modeled on the early web, where improvements benefit everyone, no one gets locked out, and communities keep control of their own data.

Current’s first cohort grant round, announced last month, involved deploying $3.2 million to four organizations across Kenya, Lebanon, and the Brazilian Amazon.

The project in Masakhane, Kenya, involves building AI datasets across more than 50 African languages for health, farming, and education; Lebanon’s Institute for Worldmaking is digitizing Arab cultural history and contemporary practice into machine-readable databases that communities (not tech companies) control. Brazil’s Portal sem Porteiras is building offline AI tools with Indigenous Amazon communities, keeping data within the territory. And Kenya’s African Internet Rights Alliance is developing audit tools to hold AI systems accountable across the continent.

Who owns the data?

On the question of data ownership, Bdeir didn’t mince words. “There are different models and proposals for who owns data in various communities, but one thing is sure: it shouldn’t be a company in Silicon Valley trying to make a select few thousand people wealthier,” she told TechCrunch.

The nonprofit’s approach is to store models and data locally, bringing in community experts before anything is built, or writing consent protocols into the pipeline so communities can halt the process at any point.

None of Current AI’s grantees have fully solved it yet. But Bdeir sees that as the point. “Every one of them has built the question into their work,” she said, “rather than accepting the usual default, where complexity becomes the excuse to let a government or a tech company decide for everyone.”

As for how much progress can be made with a $3.2 million budget split across four organizations, Bdeir says, “Scale is not always the measure. That is the Big Tech paradigm,” she said. “This could look like an Indigenous elder in the Brazilian Amazon using a tool built in Kenya to be able to pass down ecological knowledge in their own language.”

Building the stack

Earlier this month in Geneva, Current launched Alpha Chat, an open-source chatbot assembled in seven weeks by a coalition of ten organizations, including Hugging Face, Mozilla, and MIT Media Lab. Each contributor brought a piece of the stack, including a language model, safety tooling, and computing power.

Current AI also struck a deal with Sakana AI, a Tokyo-based startup known for its work on what it calls Sovereign AI. The two organizations plan to build a shared open-source AI stack, one designed to support the Japanese language and culture, but also communities across the Global South that dominant AI systems have largely ignored.

Topics

AI, current ai, large language models, nonprofit

When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.

Kate Park

Reporter, Asia

Kate Park is a reporter at TechCrunch, with a focus on technology, startups and venture capital in Asia. She previously was a financial journalist at Mergermarket covering M&A, private equity and venture capital.

View Bio

Loading the next article

Error loading the next article

Apr 2026 - now

I'm super excited to share transcribe.cpp today.

transcribe.cpp is a ggml based transcription library which supports all the latest transcription models. Every model published under the handy-computer HF org has been numerically validated and WER tested to match the reference implementation. It's accelerated everywhere.

I'm the author and maintainer of Handy. This library grew from the pains of distributing a cross-platform speech-to-text application to many people.

This is a v0.1.0 library which means that there are some rough edges which I cannot discover alone! Please report them, and let's fix them together!

Motivation

Let me say this. I think distributing a cross-platform application with the current ASR inference stack is terrible.

You've basically got whisper.cpp and ONNX. That's it. You could roll MLX in for Apple devices, but now you've to support two different engines and port models to each. I've been a fan of ONNX for getting model support into Handy quickly, but so much performance is left on the table with CPU only.

There are a few random libraries out there which claim to support a lot of models, but they have unknown authors, and unknown testing, as far as I've seen. They leave me with more questions than answers.

When will they stop maintaining this library? Has the creator thought about bindings so you can actually use it in a real desktop or mobile app? Is this effectively demo code? Have they benchmarked it? Is it faster than ONNX?

And this is what led to transcribe.cpp. As Handy's maintainer I needed a library I could trust. Where I could download a file and run inference on it. Where I can know that the inference coming from the model in the engine is as good as the reference implementation. The inference should run on the GPU for the best performance. It should be trivially embeddable in Handy, it cannot be a huge pytorch lib. It must be something that works on Mac, Windows, and Linux. And ggml seemed like by far the best way forward. It has a strong community, and a great distribution story.

So what do you get?

You get a fast and accurate inference engine with wide ranging model support.

Support for 16 ASR Families (60+ models) with more coming

Acceleration via Vulkan, Metal, CUDA, and TinyBLAS

Every model has been numerically verified and WER tested

Support for Streaming Transcription

Support for Batch Transcription

More or less drop in whisper.cpp replacement

Maintainer supported bindings in 4 Languages

Python

Javascript/Typescript

Rust

ObjC/Swift

Wide Model Support

We intend to support as many state-of-the-art transcription models as possible. As of today, we support most of the modern transcription models that are publicly available. There are a few missing still, but they will be added soon.

Acceleration Support

One of my top goals was to run any ASR model I wanted on Vulkan. In my opinion this is the floor for any application shipping local inference. For every model we support, there is a corresponding benchmark run from a Ryzen 4750U (CPU + Vulkan) on Fedora as well as on my M4 Max.

Numerically Verified

I also wanted to make sure that inference in transcribe.cpp is accurate and as close to the reference implementation as possible. This largely came from a huge degree of uncertainty of inference accuracy when using .onnx models I found on Hugging Face. In order to ensure the inference we do is correct we numerically validate every model versus the reference. On top of numerical validation, we run full WER sweeps to make sure that whatever the reference is outputting, we output the same thing. That means every model has run through thousands of utterances and is very close or same as the reference. And the results of this data are published in the transcribe.cpp repo as well as with each model on Hugging Face.

Drop In whisper.cpp replacement

transcribe.cpp is more or less a drop in support for whisper.cpp. The main reason for this is: Handy used whisper.cpp and I needed to ship an update with transcribe.cpp which would replace it. I needed to keep some compatibility with the very popular .bin files which run in whisper.cpp and shipped with Handy. transcribe.cpp can run them. There are some flags and features in whisper.cpp which we do not support yet. But I think for the vast majority of use cases our whisper implementation is solid and can replace whisper.cpp while having about equal performance.

Real Distribution

Language bindings were on my mind to begin with. While this library is written in C/C++, I needed bindings in Rust. And I also knew that in order for us to distribute local transcription as widely as possible, it requires at minimum decent first-party support of bindings. I've chosen 4 languages that I think are fairly representative of where people will use the library. I welcome others to contribute bindings directly to the project as well, assuming that they are willing to take on the maintenance burden of doing so.

And of course, at the end of the day, a lot of the decisions were driven by Handy. As a result of Handy being popular, I intend to maintain this library, just as I've done my best to maintain Handy. I intend to be someone who continues to maintain open source projects and contribute to the ecosystem where I can.

This library never would have existed without Handy because I wouldn't have had the problem of trying to support a bunch of different ASR models. I would have never learned all the use cases that people have for ASR. I've done my best to cover the ones that I hear about the most. Certainly, there are cases in the library that are not currently handled. If there are things that I missed, you are free to contribute to the library!

Making Local Speech to Text More Accessible

transcribe.cpp is aimed squarely at making locally run ASR easier. We know that transcription can run extremely accurately on most devices, and there should be no need to send your voice to a cloud service. An RK3566 can run models via transcribe.cpp faster than real time on its anemic CPU. Faster than real time transcription with SOTA models runs in a handful of watts. It's not a hope or a dream, it's a fact.

I think as we look forward to the future, more inference will start happening locally for one reason or the other. This brings the distribution story front and center. In order to have more applications running inference locally, we need to make running inference easier. Certainly transcribe.cpp does not solve this on the whole, and there is a long way to go, but I hope it's a small step forward. I've certainly learned a lot.

Gratitude

I am extremely thankful for all the folks who have supported this project.

First and foremost is to Mozilla AI, their BiR program, and Davide from Mozilla AI. This project was largely a problem in my head that I came to them with, and they decided to support me in solving the problem. At the time transcribe.cpp wasn't even a concrete idea, I was just exploring how to solve accelerated distribution in Handy. So a huge thanks to them, their support, and helping to bring this project into existence.

ggml. This project wouldn't be possible without ggml and all of the contributors to it. Thank you all so much for the work you've done. I think ggml really does amazing work in helping to make distributing local inference applications easy and possible.

Modal has also been a critical help for me. I reached out to them, and they gave me credits. These credits are put towards doing the WER testing and ensuring the library works well on CUDA. It is an immense help being able to verify the correctness of the work.

Blacksmith helps to power some of the CI/CD for transcribe.cpp. Again I reached out to them and they immediately responded with credits. Of course CI/CD is critical for making sure everything put out has been tested to at least some degree.

Hugging Face both for being a pillar in the local AI community, as well as providing the handy-computer org private storage, so I could upload models at my own will.

AI Assisted?

Yes absolutely. I don't think it's possible for a single individual to write an engine from scratch of this size using ggml in a handful of months without outside assistance. Were any of the words here written using AI? Nope. They came from my mouth or my fingers.

What makes Poke interesting to consumers — and now, to its acquirer — is how it engages with users' requests. Instead of functioning like a tool, Poke chats more familiarly with users, responding like a friend and even incorporating slang and humor into its interactions. The deal will bring Poke's interaction model and personality to Cognition's coding assistant Devin, while Poke will take advantage of Cognition's models and infrastructure to become faster and more reliable.

The acquisition underscores a growing belief that how AI assistants interact with users will become just as valuable as the underlying models powering them. As Interaction Company co-founder Marvin von Hagen explained, "You probably prefer it if you have co-workers that have personality, rather than if you have co-workers that are just robots. When you have co-workers that are also software engineers, they can also make a joke. And you'll find it funny and enjoyable." Cognition wants to make Devin feel less like software and more like a colleague.

Cognition co-founder Scott Wu wrote in a blog post: "The Interaction team has built an agent that people love: it's proactive, it knows you, and it's fun to talk to. That's exactly how working with Devin should feel, and now we get to build it together." Wu and co-founder Walden Yan had been angel investors in the company behind Poke.

First launched in March 2026, Poke allows users to engage with its AI agent through messaging platforms including iMessage, SMS, Telegram, and WhatsApp in some markets. People use Poke for various tasks across travel, health, finance, scheduling, and education, with productivity tasks like managing emails, reminders, and to-dos among its most common use cases. Over the past three months, Poke users exchanged more than 100 million messages on the platform.

Despite being used by hundreds of thousands of people, Poke had been expensive to run, making it difficult to turn a profit. In June, Poke became the first AI agent approved to run on Apple's Messages for Business platform, which provides businesses with a standardized experience for handling customer communications within Apple's Messages app.

Going forward, nothing will immediately change for Poke through the end of the year. Poke plans to remain on Apple's platform. Next year, the team will experiment with how both Poke and Devin can improve, with Poke turning to Cognition's newest software engineering model SWE-1.7 for some tasks. Fully combining the two products has not been ruled out.

Von Hagen envisions long-term potential: "I think in the long term Poke can be more reliable with orchestrating all these coding tasks…and, at the same time, Devin could become more like Poke. Like right now, Devin can only do one PR — pull request — at a time…I think there's a lot of value in having Poke orchestrate different Devin sessions. It would be good to have a persistent co-worker."