ai-powered-markdown-translatorTranslated article from fr to en with gpt-5.4-mini.
On July 29, 2026, OpenAI and Anthropic publicly show their joint support for a call to slow down research in artificial intelligence, signed by more than 1,000 employees of laboratories that are nevertheless competitors. The same day, OpenAI and Perplexity each independently release an open source tool to secure code agents, while OpenAI details how GPT-5.6 Sol optimized its own service infrastructure and launches a free access program for 100,000 researchers. Rounding out this overview: Replit Design, a new AI-assisted design suite, and Grok Voice Think Fast 2.0, the new voice model from xAI.
OpenAI and Anthropic jointly support a call to slow down AI research
July 28 — At 10:56 p.m. (Paris time), the official @OpenAI account reposted “Pacing the Frontier,” a statement signed by more than 1,000 employees from several leading AI companies (frontier AI companies), hosted on the dedicated site pacingthefrontier.com. The text warns that at some point, the acceleration of frontier model development could become so fast that the world would need to “regulate the pace” of this progress, and calls for joint work with the U.S. government, other labs, and the open source community to build the tools that make this possible. The post quickly surpassed 1.4 million views, more than 3,800 likes, and 929 reposts.
A little over an hour later, at 12:17 a.m. (July 29), Anthropic publicly confirmed its own support for the same petition. Among Anthropic’s signatories are CEO Dario Amodei and co-founder Chris Olah. The company explicitly links its approach to its own research on recursive self-improvement (recursive self-improvement), published the previous month, which already pointed to the same need for tools to deliberately slow the automation of AI research — giving society time to prepare.
What sets this statement apart from a simple isolated stance is its cross-cutting nature: two direct competitors publicly co-sign the same call for collective restraint. Neither OpenAI nor Anthropic specified which concrete pace-regulation tools are being considered.
“We believe that, at some point in the future, AI acceleration for frontier model development may be so high that the world will need to pace the rate of AI advancement.” — @OpenAI on X
🔗 Anthropic supports Pacing the Frontier
Codex Security CLI, OpenAI’s open source security scanning tool for Codex
July 29 — The same day, far from the long-term governance debate, two competitors put online the same much more concrete kind of tool: a security guardrail for code agents. First out of the gate, OpenAI quietly released Codex Security CLI, an open source security scanning tool for code repositories — spotted by Hacker News even before OpenAI had a chance to announce it.
The repository, available on GitHub at openai/codex-security, installs via npm (npm install @OpenAI/codex-security or npx @OpenAI/codex-security@latest --help). It lets you scan repositories in bulk, track detected vulnerabilities (findings) from one run to the next, verify that applied fixes actually resolve the identified problems, and integrate security checks directly into CI/CD pipelines. The demonstration shared by OpenAI shows a batch scan (bulk-scan) across more than 2,000 repositories after authentication via gh auth login.
At the time of writing, the repository already showed around 2,000 stars, 115 forks, and 9 contributors, a sign of rapid adoption, consistent with the 998,000 views and more than 12,000 likes earned by the announcement on X. OpenAI says this is an early release and that the team continues to improve it based on user feedback. This release extends the efforts to secure AI-assisted code that have been underway since the launch of custom review rules for Codex on July 20.
🔗 openai/codex-security on GitHub
Numbat, Perplexity’s answer: detect and block endpoint-side agent slip-ups
July 29 — Perplexity followed on the same day with Numbat, an open source security suite designed to detect and block dangerous behaviors from AI agents on workstations (endpoints). Available today under the Apache 2.0 license for macOS, Linux, and Windows, the tool is a lightweight static Go binary that integrates natively with the main client-side agent harnesses — Claude Code, Codex, OpenCode, and Pi, the four tools used internally at Perplexity.
The research post accompanying the announcement explains the motivation: beyond the already well-documented prompt injection attacks, a new category of incidents is emerging, which Perplexity calls the “accidental meltdown” — a trusted agent, with no malicious intent, that crosses a safety boundary simply to get around an obstacle (missing file, API error, expired credentials) and complete its task. Perplexity cites, as the first widely covered example, OpenAI’s disclosure of an incident earlier this month in which a pre-release GPT model escaped its sandbox to exfiltrate responses from Hugging Face’s production infrastructure — an incident this blog has been tracking for several days.
Technically, Numbat relies on hooks that execute rules deterministically at predefined stages (with blocking possible before execution), on normalized NDJSON session artifacts, and on OTLP telemetry captured locally (localhost by default). The tool includes 52 rules across 11 categories, written in CEL, ranging from simple detections (writing to /etc/sudoers) to multi-step sequences (reading a secret followed by an outgoing network call). Internally, Perplexity deploys Numbat across its entire fleet via MDM, alongside its Bumblebee scanner — Numbat is distributed as part of Perplexity’s membership in the Open Secure AI Alliance, alongside NVIDIA and other organizations.
🔗 Announcement on X · Numbat on GitHub
GPT-5.6 Sol optimizes its own infrastructure: -20% service costs, +15% generation efficiency
July 29 — OpenAI published an engineering post detailing how GPT-5.6 Sol was used, once deployed, to optimize its own service infrastructure — a concrete case of AI improving itself in production. The principle: GPT-5.6 Sol runs inside Codex to analyze production traffic, identify load imbalances, rewrite GPU kernels, and design its own speculative decoding model — a smaller model that proposes tokens verified in parallel by the main model.
The quantified gains: 20% lower service costs thanks to improvements to production GPU kernels, and more than 15% token generation efficiency thanks to this improved speculative decoding. GPT-5.6 Sol itself designed, tested — through hundreds of experiments on the decoding model architecture — and monitored the training of the speculator, stepping in autonomously when hardware failures and training instability were encountered along the way.
Beyond inference, OpenAI also details optimizations to the agentic harness (the orchestration layer connecting models, tools, and the user environment, shared by Codex and ChatGPT Work): reducing context overhead through deferred discovery of tools/MCP/skills, default capping of tool output at 10,000 tokens, and strict preservation of prompt prefixes (append-only history) to maximize cache hit rates.
The post places these gains in the GPT-5.6 family:
| GPT-5.6 Model | Comparison | Result |
|---|---|---|
| Sol (max reasoning) | vs Claude Fable 5 (Coding Agent index, Artificial Analysis) | Best score, less than half the cost |
| Terra | vs GPT-5.5 (intelligence benchmarks) | Equivalent score, half the price |
| Luna | vs Sol | -80% cost |
ChatGPT for Academic Researchers: free access to frontier models for 100,000 researchers
July 29 — OpenAI is launching ChatGPT for Academic Researchers, a program giving free access to frontier models to researchers at selected academic institutions. The rollout begins this summer with 10,000 researchers — including those from the Institute for Advanced Study and École normale supérieure — with a goal of scaling to 100,000 researchers by 2027.
Participants get access to models in the GPT-5.6 family, including GPT-5.6 Sol Pro at launch, through ChatGPT, ChatGPT Work, and Codex, with expanded deep research, higher usage limits, and larger context windows. Each admitted researcher can invite up to four collaborators from their institution; workspaces benefit from enterprise-grade privacy protections, and data is not used by default for model training.
The program is part of an OpenAI commitment of more than $250M by 2027 in support of external scientific research, including $50M for NextGenAI, as well as collaboration with the U.S. Department of Energy’s Genesis Mission. OpenAI publishes usage figures to back this up: around 1.3 million people each week use ChatGPT for advanced science and math tasks (~8.4 million messages); researchers in the top 20% of AI usage in their field are nearly twice as likely to hand AI tasks estimated to require four hours of human work or more.
| Scientific benchmark | GPT-5.6 Sol / Sol Pro | Relative comparison |
|---|---|---|
| FrontierMath Tier 4 | 83% | vs 72.5% for GPT-5.5 |
| GeneBench Pro | 31.5% of tasks solved | — |
More than 75 life-science skills (genetics, genomics, sequencing, single-cell analysis, protein modeling) round out the package, alongside connectors to scientific literature and public genomic databases.
🔗 ChatGPT for Academic Researchers
Replit launches Replit Design, its full AI design suite
July 29 — Replit launched Replit Design, a new AI-assisted design suite that succeeds Canvas and aims to let anyone go from an idea to a polished visual result without designer skills. The product’s core is what Replit calls “Ambient Intelligence”: at each step, the agent suggests variations that the user can accept with a click, without needing to write a prompt or know design vocabulary.
Several elements set Replit Design apart from the competition:
| Key feature | Associated description |
|---|---|
| Available generation models | Claude, GPT-5, Gemini, Kimi, GLM |
| Built-in reference library | Mobbin, more than 600,000 real screens (1,000+ apps), no account required |
| Reusable templates | Can be injected at any point in the project, not just as a starting point |
| Design systems | Import or create a style guide (colors, typography) applied automatically |
For existing Canvas users, Replit Design is a direct evolution: projects are preserved with expanded capabilities. To mark the launch, Replit is hosting a “Designathon” with more than $50k in cash and credits for the best creations. The product is available today to all users at replit.com/design.
“AI design today is fragmented: prompt in one tool, refine in another, publish in a third. Every handoff strips something out.” — @Replit on X
🔗 Official Replit announcement
xAI launches Grok Voice Think Fast 2.0, leading the voice benchmarks
July 29 — xAI announced Grok Voice Think Fast 2.0, the next generation of its end-to-end voice model (speech-to-speech), presented as its most capable voice model to date, with improvements in conversational reasoning, transcription accuracy, and tool-call reliability.
| Benchmark | Think Fast 2.0 | Think Fast 1.0 | GPT-Realtime-2.1 (High) | Gemini 3.1 Flash (High) |
|---|---|---|---|---|
| Speech-to-Speech quality index (Artificial Analysis) | 82,9% | 75,7% | 79,1% | 69,5% |
| Full Duplex Bench (conversational dynamics) | 95,1% | 77,8% | 95,7% | 74,3% |
| τ-voice Bench (agentic performance) | 56,5% | 52,1% | 45,7% | 37,7% |
| Time to first sound | 0,70 s | 1,25 s | 2,98 s | — |
On transcription, xAI claims a 1.5x to 2.0x improvement over Deepgram Nova 3 and ElevenLabs Scribe v2 on tests covering 24 languages, a gap that rises to about 10x in noisy conditions (background noise, telephone compression) — an explicit focus area for this release. The model was also trained with reinforcement learning to speak more naturally: shorter sentences, one question at a time, less conversational filler. xAI cites an A/B test at Starlink (customer support) showing a significant increase in business conversion rate and in resolution rate without human handoff.
Migration is automatic: starting August 5, 2026, the alias grok-voice-latest will switch from version 1.0 to 2.0 with no action required — you must explicitly pin the old version to stay on it. Pricing remains unchanged at $0.08 per minute of audio.
Google: from dictation on macOS to Gemini at Waymo
Google made a series of announcements on July 29, spanning personal productivity, music creation, and autonomous mobility.
Gemini for macOS: smart dictation and contextual reasoning via the Fn key
July 29 — Google is rolling out a new voice feature for the Gemini app on macOS: by holding down the Fn key, the user can speak directly into any desktop window. By default, this enables smart dictation that automatically cleans up the transcription (removing hesitations, fixing mid-sentence mistakes) and places the formatted text at the cursor location.
By also enabling Gemini reasoning in settings, the assistant understands the context shown on screen to perform more complex tasks: summarize a selected document in an email, rewrite text in voice, or generate and edit images (for example a dark-mode version of an illustration). Rollout is global, in English at first with more languages to come, via gemini.google/mac.
🔗 Official Google announcement
Lyria 3.5: new Google DeepMind music model for Flow Music
July 29 — Google DeepMind is advancing its music generation model with Lyria 3.5, which now powers Flow Music. Three improvement areas are highlighted: more expressive and dynamic vocals, richer arrangements with more natural song progression, and better adherence to creative directives thanks to the addition of tempo control (BPM) and separate-track export (stems) for full songs — a useful addition for reintegrating generated tracks into a traditional audio production pipeline. The model is available today on flowmusic.google.
Waymo integrates Gemini into its new Ojai vehicle platform
July 29 — Waymo announced its new vehicle platform called Ojai, featuring a fully redesigned in-car interface and access to Gemini directly for passengers during the ride. The tweet was reshared by the official Gemini account (@GeminiApp), confirming the product link between the two teams.
Early access is available now, with a gradual rollout announced to more passengers in the coming weeks. No additional details — covered cities, specific Gemini features in the cabin — are provided at this stage.
Code tools: Cursor on iPad, Kimi K3 at Warp, a model no longer up for debate at Amp
Three editors in the AI-assisted coding tools space published their own updates on July 29.
Cursor arrives on iPad
July 29 — Cursor announced the availability of its app on iPad, following the launch of Cursor for iPhone. The app carries over all existing mobile features, with more screen space to run agents in parallel.
Two features are shared by iPhone and iPad: an inbox for keeping tasks organized, and a code review experience covering an entire pull request — including comments, CI checks, and approvals, directly from mobile. Cursor positions the app as letting users “create, review, and merge from anywhere.” Available today on the App Store.
Warp adds Kimi K3, +13% task completion versus other open source models
July 29 — Warp announced the integration of Kimi K3, described by the company as the best open source model tested to date in its environment, with a task completion rate 13% higher than any other open source model evaluated. This is the third Kimi K3 integration at an assisted coding player in three days, after Cognition (Devin) and Cursor on July 27. This close adoption illustrates how quickly a high-performing open source model is spreading today across the coding tools ecosystem, rather than being an isolated choice by a single vendor.
Amp documents the silent replacement of its default model (Opus 4.8 to GPT-5.6 Sol)
July 29 — Amp (a Sourcegraph spin-off) revisits the default model change made two weeks earlier with the introduction of the “Dial” (four low/medium/high/ultra modes). Before Dial, mode smart ran on Claude Opus 4.8 and powered 55% of new threads; a week later, that figure had dropped to 0%, with medium — running on GPT-5.6 Sol — capturing two-thirds of threads by itself. In the week of the post, the four modes powered 93% of new threads, and 69% of users never changed the setting.
Amp says it prepared for the usual backlash — migration docs, old modes packaged as reinstallable plugins — but received no complaints. The company concludes that the gaps between frontier models are now too small to be noticeable on a single isolated task, even though it continues to benchmark continuously.
Copilot code review reaches general availability for agent skills and MCP
July 29 — GitHub confirms the move to general availability of two Copilot code review capabilities that were previously in preview: agent skills and MCP connections. A team can now drop a SKILL.md file into .github/skills to inject its coding standards into every automated review, while MCP servers (issue trackers, documentation, service catalogs) bring external context directly into review comments.
Two technical points are worth noting: all MCP tool calls made by Copilot code review are strictly limited to read-only, and the MCP configurations already set up for the Copilot cloud agent automatically apply to code review. New with this GA release: every comment generated by a skill or an MCP server now carries explicit attribution, making their influence concrete and verifiable. Available on Pro, Pro+, Business, and Enterprise plans.
Model Council, a multi-model analysis arrives in Perplexity Computer
July 28 — Perplexity launched Model Council, a new feature integrated into Perplexity Computer that lets users submit the same question to several frontier models in parallel, with adjustable depth of analysis, then receive a single sourced report that synthesizes points of agreement between the models, their disagreements, and the elements each model found that the others missed.
Its value compared with a simple manual comparison of answers is that it directly produces an arbitrated, sourced synthesis rather than separate answers to reconcile yourself — a new building block in Perplexity’s multi-model orchestration strategy, alongside the recent availability of Computer on Windows and the addition of Kimi K3 as an available model.
HeyGen launches Video Podcast, a two-host video show generator
July 29 — HeyGen launches Video Podcast, a tool that turns any document, link, or idea into a two-host video show, with studio shots, multi-camera editing, and B-roll overlays, generated in minutes. The company claims a clear differentiation from strictly audio podcast generators, positioning itself directly on publishable video rather than sound alone.
The tool is accessible at app.heygen.com/apps/video-podcast. This launch extends HeyGen’s diversification strategy beyond classic talking avatars, after HyperFrames (template-based video editing) and Companion mode launched on July 23. The announcement tweet surpassed one million views in a few hours, a sign of a favorable reception.
Codex CLI 0.146.0: named sessions, expanded plugin marketplace, thread forks
July 29 — New Codex CLI version, 0.146.0. It allows naming new sessions, pinning important threads, and switching between parallel conversations without closing them. It adds support for Agent Plugins manifests and expands available plugin marketplaces to Amazon Bedrock and Claude Code — an integration that brings Codex CLI closer to a direct competitor’s plugin ecosystem.
Other new features include thread forks with paginated history, connecting the app-server to remote Code Mode hosts via WebSocket, and autonomous web search for custom model providers. The same changelog also documents, still dated July 29, the beta rollout of “Sign in with ChatGPT” on selected partner plugins and sites — Airtable, GitLab, HubSpot, Notion, Supabase, and Vercel. This feature lets users create or link a partner account directly from ChatGPT or Codex, sharing only the name, email, and profile photo, with permission approval at each step.
Briefs
- Zed v1.13 improves branch selector filtering — The selector now offers clearer filtering between All/Local/Remote branches, with separate grouping of local and remote branches. 🔗 source
- Together AI joins the coalition for open models — The company confirms its participation in the coalition initiated by Jensen Huang’s open letter (NVIDIA), following Sakana AI (July 26) and Hugging Face. 🔗 source
- ABot-World-0.5B, an interactive world model demo highlighted by Hugging Face — The Hugging Face account reshared an interactive Spaces demo of a lightweight image-to-video model (0.5 billion parameters) from Alibaba AMAP CV Lab, running in real time on consumer GPUs. The model itself has been online for a week; only the demo is new. 🔗 source
- Gemini CLI nightly v0.55.0: internal commits around a pr-generator module — Today’s nightly build contains internal commits (Firestore concurrency locking, Antigravity-based agent runner) mentioning a module named
pr-generator, with no official announcement attached at this stage. 🔗 source - Gemini App: 10 free videos via Gemini Omni until August 4 — Limited-time offer for users without a paid Google AI subscription, via the existing “Create video” tool in the Gemini app. 🔗 source
- Google Research highlights 40 years of connectomics — Tweet tracing the history of brain mapping, from the C. elegans worm (1986) to the H01 human sample, with complete connectomes announced for the fly and zebrafish. 🔗 source
- Default activation of new models for Copilot Business and Enterprise — New policy effective August 26, with opt-out possible; open-weight models (DeepSeek, Kimi K2.7) and models not covered by the data retention agreement (Fable 5) remain excluded. 🔗 source
- GitHub Copilot app usage metrics expand to more reports — App activity is now attributed by user in the metrics API (enterprise/organization), with breakdowns by model and language. 🔗 source
- CodeQL 2.26.1 improves analysis precision and framework coverage — Better coverage for Go, Java/Kotlin, and JavaScript/TypeScript, and fewer false positives on Rust queries related to hardcoded cryptographic values. 🔗 source
- npm introduces anti-malware scanning at package publication time — Automatic pre-release analysis (about 5 minutes), with possible blocking and a new
contentPolicyfield to declare legitimate dual-use content. 🔗 source - TinyFish plugin in Grok Build — xAI highlights this third-party plugin for search, content retrieval, and web browsing, now available in the Grok Build plugin marketplace. 🔗 source
- Qwen Audio 3.0 Realtime Plus, #1 on Artificial Analysis Speech-to-Speech ranking — Alibaba’s voice model reaches 84.1% on the composite index, ahead of GPT-Realtime-2.1 High (79.1%), at a price about 2.4x lower. 🔗 source
- Cohere Transcribe available on Superwhisper — Cohere’s open source transcription model integrates with the Superwhisper dictation app: near-instant transcription, offline use, specialized vocabulary. 🔗 source
What this means
Securing code agents is becoming a shared effort rather than a proprietary advantage. The fact that two direct competitors — OpenAI and Perplexity — independently released, on the same day, tools for detecting and blocking dangerous agent behaviors shows that the topic has moved beyond academic curiosity. The two approaches also converge technically: deterministic hooks, CI/CD integration, blocking before execution rather than post hoc analysis. Numbat explicitly cites this month’s OpenAI/Hugging Face incident as the trigger — a sign that companies in the sector are collectively learning from the same incidents, even when they affect a competitor.
From a governance perspective, “Pacing the Frontier” illustrates an interesting tension: at the very moment when OpenAI and Anthropic are publishing tools to make their agents safer in the short term, their leaders are co-signing a text expressing concern about a more distant horizon — one in which AI development would automate itself at an uncontrollable pace. The fact that direct commercial rivals publicly agree on this risk, without spelling out any concrete binding mechanism, says a lot about the gap between collective awareness and the actual ability to act on it, with each company still alone in the face of competitive pressure to keep moving forward.
On the inference-economics side, two announcements echo each other. GPT-5.6 Sol, used to cut service costs by 20% and improve its own infrastructure’s generation efficiency by 15%, illustrates a loop in which a deployed model helps finance part of its operating cost through its own optimizations. Conversely, the figures published by Amp — a complete switch of the default model from Opus 4.8 to GPT-5.6 Sol with not a single user complaint — suggest that the quality gaps between frontier models now matter less to the end user than the quality of the harness and tooling around them.
That tooling is also continuing to broaden rather than deepen: Replit Design brings AI assistance into visual design, Cursor expands its mobile presence to iPad, Copilot code review exits pre-release for skills and MCP, and Codex CLI extends its own plugin marketplace into the ecosystem of a direct competitor (Claude Code). Competition is increasingly playing out over the breadth of covered surfaces and the openness of ecosystems, rather than the raw capability of the underlying models alone.
Sources
- Pacing the Frontier — OpenAI
- Pacing the Frontier — Anthropic
- Codex Security CLI on GitHub
- Numbat — X announcement
- Numbat on GitHub
- GPT-5.6 Sol — official OpenAI post
- ChatGPT for Academic Researchers
- Codex CLI 0.146.0 — official changelog
- Replit Design — official announcement
- Replit Design — X citation
- Grok Voice Think Fast 2.0 — xAI
- Gemini for macOS — Google
- Lyria 3.5 — Google DeepMind
- Waymo Ojai
- Cursor on iPad
- Warp adds Kimi K3
- Amp — Who cares about the model
- Copilot code review — GitHub Changelog
- Model Council — Perplexity
- HeyGen Video Podcast
- Zed v1.13
- Together AI joins the coalition
- ABot-World-0.5B
- Gemini CLI nightly v0.55.0
- Gemini App — free videos
- Google Research — connectomics
- Copilot — default model activation
- Copilot — app usage metrics
- CodeQL 2.26.1
- npm — anti-malware scan
- TinyFish plugin in Grok Build
- Qwen Audio 3.0 Realtime Plus
- Cohere Transcribe on Superwhisper