AI DAILY / DEV
←
FRIDAY
April 24, 2026
→

    OpenAI Ships GPT-5.5 in ChatGPT and Codex, 88.7% SWE-Bench and 60% Fewer Hallucinations
    • Standard, Thinking, and Pro variants rolled out April 23 to Plus/Pro/Business/Enterprise plans plus Codex.
    • 88.7% SWE-bench, 92.4% MMLU, and a reported 60% drop in hallucinations over GPT-5.4.
    • API pricing: $5/$30 per M tokens standard, $30/$180 Pro; Codex user base now at 4M.
    • TechCrunch frames the release as OpenAI's step toward a ChatGPT 'super app' — longer agentic runs with less user prompting.
    models openai.com

    DeepSeek Drops V4 Open Weights Same Day as GPT-5.5, 1.6T MoE Under Apache 2.0
    • V4 Pro: 1.6T total / 49B active MoE; V4 Flash: 284B / 13B active — both Apache 2.0 with 1M-token context.
    • Bloomberg: new Hybrid Attention Architecture (Compressed Sparse + Heavily Compressed) plus an 'Engram' memory layer hits 97% needle-in-a-haystack.
    • Release timing against GPT-5.5 is explicit — DeepSeek is pitching V4 as the open-source answer to every closed frontier model.
    • Lands as Tencent and Alibaba are reportedly in talks to lead DeepSeek's first outside round at a $20B+ valuation.
    models bloomberg.com

    xAI Launches Grok Voice Think Fast 1.0, Under 1-Second Latency via Starlink
    • Flagship voice model built with Starlink for reasoning-in-the-background on multi-step voice workflows.
    • Time-to-first-audio under 1 second — roughly 5× faster than the nearest competitor per xAI.
    • Targets customer support, sales, and high-volume tool-calling; positioned against OpenAI Realtime and ElevenLabs.
    • Companion Grok STT/TTS APIs launch alongside — $0.10/hr batch, $0.20/hr streaming, with speaker diarization and expressive speech tags.
    models x.ai

    Meta to Cut 8,000 Jobs While Guiding 2026 AI Capex to $115B–$135B
    • Internal memo April 23: ~10% of headcount (about 8,000 roles), cuts starting May 20, plus 6,000 open reqs scrapped.
    • Meta Superintelligence Labs is exempt — reductions concentrate in non-AI product, recruiting, and support orgs.
    • Funds 2026 capex guidance of $115B–$135B on AI infra and talent; Zuckerberg's 'personal superintelligence for everyone' push cited in the memo.
    • Second major workforce cut in roughly a year — the sharpest Big Tech reshape yet to fund AI spend.
    industry cnbc.com

    Hugging Face ml-intern Beats Claude Code on GPQA, Open-Source Autonomous ML Engineer
    • Agentic loop up to 300 iterations reads arXiv/HF Papers, finds datasets, launches training jobs, and iterates.
    • Demo: Qwen3-1.7B pushed from ~10% → 32% on GPQA in under 10 hours of autonomous fine-tuning.
    • Scores above Claude Code (~23%) on the same benchmark per Hugging Face; built on the smolagents framework.
    • On GitHub Trending daily and weekly; CLI plus web app, fully open source.
    open-source github.com