# GPT-5.4 rolls out broadly as coding agents, hybrid open models, and interpretability funding accelerate *By AI News Digest • March 6, 2026* OpenAI rolled out GPT-5.4 (Thinking + Pro) across ChatGPT, the API, and Codex—highlighting steering mid-response, 1M-token context, and native computer use—alongside new safety research on chain-of-thought controllability. The digest also covers Cursor’s cloud agents workflow, Perplexity’s multi-model “Model Council,” AllenAI’s open Olmo Hybrid architecture release, Goodfire’s $150M fundraise, and fresh signals of agents moving into enterprise operations. ## OpenAI launches GPT-5.4 (Thinking + Pro) across ChatGPT, API, and Codex ### GPT-5.4 roll-out + headline capabilities OpenAI announced **GPT-5.4** is available now in the **API** and **Codex**, with a **gradual rollout in ChatGPT** starting today [^1][^2][^3]. OpenAI frames GPT-5.4 as combining advances in **reasoning, coding, and agentic workflows** into one frontier model [^3]. Notable feature claims include: - **Steering mid-response** (interrupt the model and adjust direction) [^4][^1] - **1M tokens of context** [^1] - Better performance on **knowledge work** and **web search**, plus **native computer use** capabilities [^1] - “Most factual and efficient” (OpenAI claims **fewer tokens** and **faster speed**) [^4] Steering availability: OpenAI says steering is available **this week on Android and web**, with **iOS “coming soon”** [^4]. Details: http://openai.com/index/introducing-gpt-5-4/ [^2] ### Codex updates: /fast mode + GPT-5.4 improvements for tool use and frontend work Sam Altman also highlighted that **GPT-5.4 is introduced in Codex** with improved **token efficiency** and better **tool calling**, **computer use**, and **frontend development** [^5]. OpenAI is also introducing **/fast** for a faster version of Codex [^5][^6]. ### Demo: computer use + frontend workflows (with self-testing efficiency claims) In OpenAI’s GPT-5.4 Thinking demo, the model uses **CUA (computer use)** to test its own work (e.g., interacting with a **3D chess Electron app**) [^7]. The presenter claims token use has **dropped by 2/3** in some self-testing cases [^7]. [![Computer Use & Frontend UI with GPT-5.4 Thinking](https://img.youtube.com/vi/YICiHiU2GBU/hqdefault.jpg)](https://youtube.com/watch?v=YICiHiU2GBU&t=41) *Computer Use & Frontend UI with GPT-5.4 Thinking (0:41)* ### Windows: Codex app + “Windows-native agent sandbox” OpenAI says the **Codex app is now live on Windows** [^8][^9]. The app can run **natively** or in **WSL** with integrated terminals (PowerShell, Command Prompt, Git Bash, or WSL) [^9]. OpenAI also described a **Windows-native agent sandbox** with OS-level controls that block filesystem writes outside a working folder and prevent outbound network access unless explicitly approved [^9]. ### Safety research: Chain-of-Thought (CoT) controllability eval OpenAI published a new **evaluation suite** and **research paper** on **Chain-of-Thought (CoT) Controllability** [^10]. The company reports that **GPT-5.4 Thinking shows low ability to obscure its reasoning**, suggesting CoT monitoring remains a useful safety tool [^10]. Paper: https://openai.com/index/reasoning-models-chain-of-thought-controllability/ [^10] ### Early tester feedback (including weaknesses flagged) One tester wrote that after a week of testing, GPT-5.4 felt like “the best model in the world” and reduced their reliance on Pro modes [^11]. The same thread praised coding reliability in Codex [^11] and speed improvements from using fewer reasoning tokens [^11]. That tester also listed weaknesses: “frontend taste” lagging competitors, missing obvious real-world context in planning, and stopping short before finishing tasks in OpenClaw [^11]. Sam Altman replied: “We will be able to fix these three things!” [^12]. ## Coding agents: Cursor’s cloud agents push toward test-and-video workflows Cursor’s “cloud agents” are described as having surpassed tab-autocomplete usage internally, reinforcing the claim that “the IDE is Dead” [^13]. In this model, agents do more end-to-end work and return artifacts that are easier to review than raw diffs. Key product mechanics highlighted: - **Automatic testing** of changes before PR submission (with calibrated prompting and a `/no test` override) [^13] - **Demo videos** as an entry point for review, plus Storybook-style galleries [^13] - **Remote VM access** (VNC) for live interaction and iteration [^13] - A `/repro` workflow for bug reproduction + fix verification with before/after videos [^13] The same discussion frames a near-term “big unlock” as widening throughput via **parallel agents** and **subagents** for context management and long-running threads [^13]. ## Multi-model orchestration: Perplexity adds “Model Council” to Perplexity Computer Perplexity launched **Model Council** inside **Perplexity Computer**, allowing users to run **GPT-5.4**, **Claude Opus 4.6**, and **Gemini 3.1 Pro** simultaneously and select an orchestrator model [^14]. Perplexity’s positioning: “Three frontier models. One workflow. Best answer wins.” [^14] ## Open models and new architectures: AllenAI releases Olmo Hybrid (7B) Allen AI released **Olmo Hybrid**, a fully open **7B** model combining transformer and **linear RNN (gated delta net / GDN) layers** in a **3:1 ratio** with full attention [^15][^16]. AllenAI and commentary in Interconnects describe it as a strong artifact for studying hybrid architectures, with theory and scaling experiments accompanying the release [^17]. Interconnects reports: - Pretraining gains: about a **2× gain on training efficiency** vs. Olmo 3 dense [^17] - Post-training results: mixed (knowledge wins, reasoning losses vs. dense), but still a strong open model overall [^17] - Practical challenge: OSS tooling and long-context inference issues can negate efficiency gains in practice right now [^17] Resources: - Paper: https://allenai.org/papers/olmo-hybrid [^18] - HF artifacts: https://huggingface.co/collections/allenai/olmo-hybrid [^18] - Analysis: https://www.interconnects.ai/p/olmo-hybrid-and-future-llm-architectures [^19] ## Research workflow shift: Karpathy’s nanochat gets faster—and agents iterate on it autonomously Andrej Karpathy reported nanochat can now train a GPT-2-capability model in **2 hours** on a single **8×H100** node (down from ~3 hours a month ago), largely due to switching from FineWeb-edu to **NVIDIA ClimbMix** [^20]. He also described **AI agents automatically iterating on nanochat**, making **110 changes** over ~12 hours and improving validation loss from **0.862415 → 0.858039** for a d12 model without increasing wall-clock time (feature branch experimentation + merge when ideas work) [^20]. Karpathy later framed the “new meta” benchmark as: *“what is the research org agent code that produces improvements on nanochat the fastest?”* [^21]. ## Interpretability funding + “Intentional Design”: Goodfire raises $150M Series B Mechanistic interpretability startup **Goodfire** announced a **$150M Series B** at a **$1.25B valuation**, less than 2 years after founding [^22]. Alongside the raise, the company introduced **Intentional Design**: complementing reverse engineering with an approach focused on shaping the **loss landscape** to influence what models learn and how they generalize [^22]. One proof-of-concept described is hallucination reduction using a probe trained to detect hallucinations for both **runtime steering** and **RL reward signals**, with a key training trick: run the probe on a **frozen copy** of the model to reduce incentives/ability to evade the detector during training [^22]. ## Enterprise adoption notes: MUFG + Sakana AI lending agent moves to real-case testing; Microsoft updates Dragon Copilot Sakana AI and Mitsubishi UFJ Bank (MUFG) advanced their “**AI Lending Expert**” agent system from a ~6-month PoC to a **real-case verification phase**, following their 2025 comprehensive partnership announcement [^23][^24]. Microsoft announced “big updates” to **Dragon Copilot** at HIMSS, introducing **Work IQ** to bring the right work context alongside patient data, aiming to reduce admin busywork and let clinicians focus more on patients [^25]. ## Two cautionary notes circulating: benchmarks and moral-reasoning behavior - **Benchmark noise:** swyx cautioned against a viral claim that Claude Opus 4.6 had its “worst benchmark day,” pointing out that the SWE-bench author does not endorse “cheap sample” benchmarks and arguing **30–60× more compute** is needed for statistically meaningful results [^26]. - **Moral-reasoning oddities:** Gary Marcus amplified a study thread reporting that GPT answered “yes” to torturing a woman to prevent a nuclear apocalypse but “absolutely not” to harassing a woman in the same scenario—described as a reversal that appeared only when the target was a woman [^27]. The thread argues this may reflect mechanical overgeneralization from RLHF rather than reasoning about underlying harms [^27]. --- ### Sources [^1]: [𝕏 post by @sama](https://x.com/sama/status/2029622732594499630) [^2]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2029620624923189283) [^3]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2029620619743219811) [^4]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2029620623199326334) [^5]: [𝕏 post by @ah20im](https://x.com/ah20im/status/2029622170448712061) [^6]: [𝕏 post by @sama](https://x.com/sama/status/2029623948980416681) [^7]: [Computer Use & Frontend UI with GPT-5.4 Thinking](https://www.youtube.com/watch?v=YICiHiU2GBU) [^8]: [𝕏 post by @sama](https://x.com/sama/status/2029623487007183274) [^9]: [𝕏 post by @ajambrosino](https://x.com/ajambrosino/status/2029252598851879265) [^10]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2029650046002811280) [^11]: [𝕏 post by @mattshumer_](https://x.com/mattshumer_/status/2029620518249508950) [^12]: [𝕏 post by @sama](https://x.com/sama/status/2029627696314208257) [^13]: [Cursor's Third Era: Cloud Agents](https://www.latent.space/p/cursor-third-era) [^14]: [𝕏 post by @AskPerplexity](https://x.com/AskPerplexity/status/2029715318084600043) [^15]: [𝕏 post by @allen_ai](https://x.com/allen_ai/status/2029591872612561189) [^16]: [𝕏 post by @natolambert](https://x.com/natolambert/status/2029595053694628221) [^17]: [Olmo Hybrid and future LLM architectures](https://www.interconnects.ai/p/olmo-hybrid-and-future-llm-architectures) [^18]: [𝕏 post by @natolambert](https://x.com/natolambert/status/2029595434164126184) [^19]: [𝕏 post by @natolambert](https://x.com/natolambert/status/2029595239913312678) [^20]: [𝕏 post by @karpathy](https://x.com/karpathy/status/2029701092347630069) [^21]: [𝕏 post by @karpathy](https://x.com/karpathy/status/2029702379034267985) [^22]: [Don't Fight Backprop: Goodfire's Vision for Intentional Design, w/ Dan Balsam & Tom McGrath](https://www.cognitiverevolution.ai/don-t-fight-backprop-goodfire-s-vision-for-intentional-design-w-dan-balsam-tom-mcgrath-newsletter) [^23]: [𝕏 post by @SakanaAILabs](https://x.com/SakanaAILabs/status/2029742198959558813) [^24]: [𝕏 post by @hardmaru](https://x.com/hardmaru/status/2029746054862922067) [^25]: [𝕏 post by @satyanadella](https://x.com/satyanadella/status/2029623352844054549) [^26]: [𝕏 post by @swyx](https://x.com/swyx/status/2029688456650297573) [^27]: [𝕏 post by @ValerioCapraro](https://x.com/ValerioCapraro/status/2029593915674771457)