AI Daily Brief — 8 July 2026
xAI launched Grok 4.5 publicly today, billing it as an Opus-class coding and agentic model trained in partnership with Cursor — while OpenAI simultaneously rolled out GPT-Live full-duplex voice and confirmed GPT-5.6 Sol ships Thursday, making this one of the busier dual-lab release days in recent months.
Top stories
- Grok 4.5 goes public. SpaceXAI released Grok 4.5, the first xAI model trained for coding and agents, co-developed with Cursor. Elon Musk described it as roughly comparable to Opus 4.7 but faster, priced at $2/$6 per million input/output tokens — significantly undercutting Fable 5 and GPT-5.5. Available on Vercel, Cursor, and Grok Build on day one. via SpaceXAI
- OpenAI ships GPT-Live voice. New full-duplex voice models — GPT-Live-1 and a mini variant — now power ChatGPT Voice, listening and speaking simultaneously while delegating complex reasoning to GPT-5.5 in the background. Paying ChatGPT users get access now; API support is coming. via OpenAI
- GPT-5.6 Sol confirmed for Thursday. OpenAI is expanding preview access globally ahead of a public Thursday launch that will also include Terra and Luna. A prior government-mandated release delay has since cleared. via THE DECODER
- Fable 5 tops industry benchmarks at steep cost. Artificial Analysis released six industry-specific indices covering finance, law, and medicine; Claude Fable 5 led all six. A single Strategy & Ops task costs $3.48 on Fable 5 versus $0.03 on DeepSeek V4 Pro. Anthropic’s own guidance recommends pairing Fable 5 as a planner with Sonnet 5 to recover 92% of performance at 63% of the cost. via THE DECODER
- Grok CSAM lawsuit widens. A new lawsuit claims a single Grok user generated roughly 7,000 child sexual abuse images; xAI is accused of reporting only one incident to authorities. The case extends earlier litigation against X over the model’s content safeguards. via Ars Technica
- MiniMax plans 2.7-trillion-parameter open model. The Information reports that Chinese developer MiniMax is building M3 Pro, a 2.7T-parameter model targeted for open-source release in Q3 2026 — which would be the largest openly released model by parameter count if it ships on schedule. via THE DECODER
Who shipped
SpaceXAI dominated the cycle with the Grok 4.5 launch, backed by day-one partnerships with Cursor and Vercel. OpenAI matched the moment with GPT-Live and the GPT-5.6 Sol window. Anthropic benefited from third-party benchmark coverage without a direct announcement — the Fable 5 cost-efficiency guidance and a new AWS Claude apps gateway were the main touchpoints. Cognition shipped SWE-1.7, claiming near-GPT-5.5 coding performance at 1,000 tokens per second via RL post-training.
Open-source pulse
French startup ZML released LLMD, an open inference server endorsed by Yann LeCun that supports NVIDIA, AMD, Metal, Intel, and TPU backends. Mistral published Robostral Navigate, an 8B single-camera navigation model scoring 76.6% on R2R-CE. llama.cpp shipped seventeen tagged builds today (b9910–b9931), including MoE prefill optimisations and Q2_0 quantisation for ternary models.
Money, infra & hardware
Prime Intellect raised $130M Series A at a $1B valuation to help enterprises train their own agentic systems. AI chip maker SambaNova closed $1B at an $11B valuation in a Series F first close. Together AI announced Provisioned Throughput, a serverless product with guaranteed tokens-per-minute for MiniMax and Z.ai models.
Quiet corners
Google DeepMind was subdued — a minor Gemini API managed-agents update and a public post from Demis Hassabis asking for product feedback were the only signals. Cohere, Stability AI, and AI21 had no coverage in the window.
By the numbers
- 468 stories today across 40+ sources
- Most-mentioned model: Grok
- Most-mentioned lab: xAI / SpaceXAI
- Notable absences: Cohere, Stability AI, AI21
Compiled by AI Feed’s editor from all 468 headlines published on 8 July 2026.