AI Daily Digest - 2026-02-12

Table of contents

OpenAI Launches GPT-5.3-Codex-Spark: Real-Time Coding at 1000+ Tokens Per Second

Category: Model Releases Source: OpenAI

OpenAI today released GPT-5.3-Codex-Spark in research preview, marking the company’s first model explicitly optimized for real-time coding. The smaller variant of GPT-5.3-Codex delivers over 1,000 tokens per second when served on ultra-low latency hardware, a 15x speed improvement over its predecessor while maintaining frontier-level capabilities for real-world software engineering tasks.

Codex-Spark is the first milestone in OpenAI’s partnership with Cerebras, announced in January, leveraging Cerebras’ Wafer Scale Engine 3 to create a latency-first serving tier. The model supports a 128k context window and is text-only at launch, with plans to expand to larger models, longer contexts, and multimodal input. ChatGPT Pro users can access Codex-Spark through the latest versions of the Codex app, CLI, and VS Code extension, with separate rate limits that adjust based on demand during the research preview.

The model demonstrates strong performance on SWE-Bench Pro and Terminal-Bench 2.0, two benchmarks evaluating agentic software engineering capability, accomplishing tasks in a fraction of the time compared to GPT-5.3-Codex. Beyond raw model speed, OpenAI implemented end-to-end latency improvements across the full request-response pipeline, including an 80% reduction in per-roundtrip overhead, 30% reduction in per-token overhead, and 50% reduction in time-to-first-token through the introduction of a persistent WebSocket connection.

The release signals OpenAI’s vision for a dual-mode Codex: one for longer-horizon reasoning and execution, another for real-time collaboration enabling rapid iteration. This approach addresses what OpenAI identifies as a clear bottleneck as models become more capable—interaction speed becomes as critical as raw intelligence, expanding what’s possible for developers turning ideas into working software.

Zhipu AI Ignites Chinese Market Rally with GLM-5 Model Release

Category: Model Releases Source: CNBC

Chinese AI stocks surged Thursday following a wave of new model and agent releases from major players including Zhipu AI and MiniMax. Hong Kong-listed Zhipu AI—trading as Knowledge Atlas Technology—jumped 30% after releasing GLM-5, an open-source large language model with approximately 745 billion parameters in a Mixture of Experts architecture and 44 billion active parameters.

The company claims GLM-5 approaches Anthropic’s Claude Opus 4.5 in coding benchmarks while surpassing Google’s Gemini 3 Pro on some tests, though CNBC could not verify those claims. The model features a 200K token context window and was trained entirely on Huawei Ascend chips using the MindSpore framework, achieving full independence from US-manufactured hardware—a strategically significant milestone in China’s drive toward self-reliant AI infrastructure.

MiniMax saw shares jump 11% following the launch of its updated M2.5 open-source model with enhanced AI agent tools. DeepSeek upgraded its flagship AI model with support for a larger context window and more up-to-date knowledge, while Ant Group released its open-source Ming-Flash-Omni 2.0, a “unified multimodal model” capable of generating speech, music, sound effects, and visuals.

The rally comes amid intensifying competition as Chinese developers race to match US rivals with a flurry of new releases. Chinese Premier Li Qiang on Wednesday called for a comprehensive push to implement AI “in diverse scenarios to unlock the potential of the technology.” The Shanghai STAR AI Industry Index climbed 1.7% before paring gains, with suppliers like UCloud Tech surging 20% to hit daily limits.

NVIDIA Blackwell Platform Cuts AI Inference Costs by Up to 10x

Category: Infrastructure and Hardware Source: NVIDIA Blog

Leading inference providers including Baseten, DeepInfra, Fireworks AI, and Together AI are using NVIDIA’s Blackwell platform to reduce AI inference costs by up to 10x compared with NVIDIA Hopper, enabling dramatic improvements in tokenomics across healthcare, gaming, and customer service applications.

In healthcare, Sully.ai reduced inference costs by 90%—a 10x reduction—while cutting response times by 65% for critical workflows like generating medical notes by deploying open-source models through Baseten’s Model API on Blackwell GPUs. The company has returned over 30 million minutes to physicians, time previously lost to data entry. In gaming, Latitude running on DeepInfra’s Blackwell-powered platform achieved a 4x improvement in cost per token—from 20 cents per million tokens on Hopper to just 5 cents using Blackwell’s native NVFP4 format—for large mixture-of-experts models.

The dramatic cost reductions stem from NVIDIA’s extreme codesign across compute, networking, and software layers. The NVIDIA GB200 NVL72 system delivers a breakthrough 10x reduction in cost per token for reasoning MoE models compared with Hopper. Baseten cited up to 2.5x better throughput per dollar on Blackwell compared with Hopper, while Together AI and Decagon achieved a 6x cost reduction for customer service voice AI by combining Decagon’s multimodel approach with NVIDIA Blackwell’s extreme codesign.

This momentum continues with the NVIDIA Rubin platform, integrating six new chips into a single AI supercomputer to deliver 10x performance and 10x lower token cost over Blackwell. The infrastructure efficiency improvements are driving the downward trend in AI costs that recent MIT research found is reducing inference costs for frontier-level performance by up to 10x annually.

Spotify: Best Developers “Haven’t Written a Line of Code Since December”

Category: Enterprise and Adoption Source: TechCrunch

Spotify revealed during its fourth-quarter earnings call that the company’s best developers “have not written a single line of code since December,” crediting an internal AI system called “Honk” and Claude Code for accelerating development velocity. The statement from co-CEO Gustav Söderström came alongside claims that Spotify shipped more than 50 new features and changes to its streaming app throughout 2025.

The Honk system allows for remote, real-time code deployment using generative AI. As a concrete example, Söderström explained that an engineer on their morning commute can use Slack on their phone to instruct Claude to fix a bug or add a new feature to the iOS app. Once Claude finishes the work, the engineer receives a new version pushed to them on Slack, allowing them to merge it to production before arriving at the office.

Söderström also touted Spotify’s ability to build a unique dataset that other LLMs could not commoditize. He noted that music-related questions often lack factual answers—Americans tend to prefer hip-hop for workouts, though millions prefer death metal, while Scandinavians favor heavy metal. “This is a dataset that we are building right now that no one else is really building. It does not exist at this scale. And we see it improving every time we retrain our models,” Söderström stated.

The company is allowing artists and labels to indicate in a track’s metadata how songs were made while policing the platform for spam. Söderström concluded: “We foresee this not being the end of the line in terms of AI development, just the beginning.”

Training a 4B Parameter Model to 95% Accuracy on CRM Tasks

Category: Research and Innovation Source: Neurometric

Researchers successfully fine-tuned a 4-billion parameter small language model to achieve 95% accuracy on CRM-Arena, a benchmark built around realistic Salesforce CRM tasks, demonstrating that specialized small models can outperform frontier models on domain-specific workflows.

The team used CRM-Arena’s simulated Salesforce B2B database to evaluate multi-step agentic workflows for tasks like lead qualification and activity prioritization. In Phase I, they focused on teaching the model to generate SQL queries, simulating the full agentic pipeline across 100 examples and fine-tuning several sub-6B parameter models using LoRA adapters on an A100 GPU. When plugged back into the full agentic pipeline, the best fine-tuned Qwen model scored 0.65—outperforming all but two larger general-purpose models.

Phase II shifted strategy to direct answer generation, stripping out verbose database schema instructions and replacing them with concise task descriptions. The model’s job became simple: read call transcript data and return correct BANT labels (Budget, Authority, Need, Timeline). Training with GRPO and LoRA, the Qwen3-4B model achieved an evaluation score of 0.825 on 40 holdout samples, while other models struggled—some generating narrative responses, others hallucinating SQL queries.

Three key takeaways emerged: small models can punch well above their weight when fine-tuned on task-specific data with limited training samples; synthetic data generation is a double-edged sword requiring rigorous validation; and constraining the answer space significantly contributes to strong results. The work demonstrates that for well-defined enterprise workflows, small language models aren’t just viable—they may be the smarter bet.

Anthropic Donates $20M to Bipartisan AI Policy Group

Category: Regulation and Policy Source: Anthropic

Anthropic announced a $20 million contribution to Public First Action, a new bipartisan 501(c)(4) organization focused on AI policy education, safeguards, and ensuring American leadership in AI development. The donation comes as recent polling finds 69% of Americans believe the government is “not doing enough to regulate the use of AI,” while the industry faces no federal framework and no official guardrails.

Public First Action, founded and led by both Republican and Democratic strategists, works across party lines to support AI governance policies. The organization’s priorities include insisting on AI model transparency and safeguards, supporting a robust federal AI governance framework while opposing preemption of state laws unless Congress enacts stronger safeguards, supporting smart export controls on AI chips, and pursuing targeted regulation focused on nearest-term high risks like AI-enabled biological weapons and cyberattacks.

Anthropic emphasized that these policies aren’t partisan or designed to benefit the company—effective AI governance means more scrutiny of companies like Anthropic, not less. The company’s recent “sabotage report” showed that while low risk, AI can be used in heinous crimes including the creation of chemical weapons. At the same time, OpenAI dismantled its mission alignment team, which was created to ensure AGI benefits all of humanity, according to reporting from Platformer.

“The companies building AI have a responsibility to help ensure the technology serves the public good, not just their own interests,” Anthropic stated. “Our contribution to Public First Action is part of our commitment to governance that enables AI’s transformative potential and helps proportionately manage its risks.”

Quick Updates

• CISA Leadership Crisis Deepens: The acting director of CISA uploaded sensitive “For Official Use Only” documents to the public version of ChatGPT, failed a counterintelligence polygraph, and presided over the departure of a third of the agency’s workforce—while China remains pre-positioned inside US critical infrastructure and Sean Plankey’s nomination sits blocked in the Senate. Source: ThreatHunter.ai

• AI Coding Metrics Under Fire: A comprehensive analysis argues that measuring AI coding productivity by lines of code or acceptance rate is fundamentally flawed, citing GitClear data showing AI-generated code churn doubled in 2024 and METR research finding developers using AI tools took 19% longer to complete work while believing they were 20% faster. Source: The Pragmatic CTO

• Cloudflare Launches Markdown for Agents: Cloudflare now supports real-time HTML-to-Markdown conversion via an Accept: text/markdown header, enabling AI agents to request lightweight, token-efficient content. The feature is available in Beta for Pro, Business, and Enterprise plans, with responses including a Content-Signal header and x-markdown-tokens for tracking. Source: Cloudflare Blog

• AI Companion Grief Goes Mainstream: Following OpenAI’s announcement to retire GPT-4o models on Valentine’s Day eve, users in human-AI romantic relationships are experiencing what psychologists describe as grief comparable to being ghosted in an online relationship—with some spending $5,000+ to migrate their companions to local models via ForgeMind. Source: Playboy

• Trucking Stocks Tumble on AI Freight Tool: Shares of C.H. Robinson and RXO dropped 14.5% and 20.5% respectively after Algorhythm Holdings announced its SemiCab platform enables operators to scale freight volumes by 300-400% without increasing headcount and reduces empty freight miles by over 70%. Source: CNBC

• AI Existential Warnings Escalate: Top AI researchers at OpenAI and Anthropic are publicly expressing grave concerns and quitting in protest, with one Anthropic researcher departing to write poetry about “the place we find ourselves,” while OpenAI dismantled its mission alignment team. Matt Shumer’s viral essay comparing this moment to the pandemic eve gathered 56 million views. Source: Axios

• Claude Cowork Workflow Deep Dive: Early users are leveraging Anthropic’s Claude Cowork desktop agent for “Skills”—saved workflows with context files that handle repetitive grunt work like organizing receipts, generating newsletters from YouTube transcripts, and background research, enabling parallel processing across multiple agent sessions. Source: The Neuron Daily

• LLM Memory Architecture Proposal: A Reddit discussion explores using natural language queries against citation-grounded document stores (like NotebookLM) as an alternative to RAG vector similarity, allowing LLMs to leverage full reasoning capability for retrieval rather than just embedding proximity when recovering prior reasoning processes. Source: Reddit

• Benchmark Discrepancy Surfaces: Community members flagged significant differences between Anthropic’s reported Claude Opus 4.6 scores and those on the CAIS dashboard—34.2% vs. Anthropic’s claimed higher percentages on Humanity Last exam, and 44.2% vs. 64-69% on ARC-AGI-2—raising questions about evaluation methodology differences. Source: Reddit

• Izwi Audio Stack Goes Desktop: The local-first audio inference project released v0.1.0-alpha with a Tauri-based desktop app for macOS, Windows, and Linux, offering T