Meta says it has caught OpenAI in frontier benchmarks. OpenAI wants to hand the U.S. government a 5% stake worth $42.6 billion. And a Chinese startup just released an open-weight model that runs near frontier performance at a fraction of the cost. The race isn’t slowing down. It’s getting weirder.
Meta’s Watermelon Model Matches GPT-5.5, Internal Town Hall Reveals
Alexandr Wang, Meta’s superintelligence chief, told employees during an internal town hall that the company’s upcoming Watermelon model has matched OpenAI’s GPT-5.5 on closely watched AI benchmarks. The model is still in training, and Wang said it uses an order of magnitude more compute than Avocado, the internal codename for Meta’s Muse Spark model released in April.
If accurate, this would be the clearest sign yet that Mark Zuckerberg’s massive AI spending is starting to close the gap with OpenAI, Google, and Anthropic. Meta has trailed those three in frontier model performance despite investing billions in chips, data centers, and elite talent. The company raised its infrastructure spending forecast to $125 to $145 billion this year, up from an earlier $115 to $135 billion range. Wang also pointed to progress on X, saying an update to Muse Spark would arrive soon with stronger coding and agentic abilities. When asked when Meta would have a coding model on par with Anthropic’s Claude Opus, he replied “pretty soon.”
Here’s what matters: Meta hasn’t named which benchmarks Wang cited, and the model isn’t released yet. Claims made in internal town halls are marketing until proven otherwise. But the fact that Meta’s superintelligence lead is willing to tell staff they’ve caught GPT-5.5 suggests the gap has narrowed enough to be worth claiming publicly. That changes the competitive dynamic. OpenAI no longer stands alone at the top.
Source: Yellow
OpenAI Proposes 5% Equity Stake to U.S. Government Worth $42.6B
OpenAI has proposed transferring 5% of its equity to a government-linked public wealth fund, worth roughly $42.6 billion at the company’s $852 billion valuation. Sam Altman raised the concept directly with Donald Trump in early 2025, and OpenAI sketched out the fund in a policy paper published in April. Altman has discussed the stake with Trump, Commerce Secretary Howard Lutnick, Treasury Secretary Scott Bessent, and Senator Bernie Sanders in recent weeks.
The proposal models itself on the Alaska Permanent Fund, which pays Alaskans annual dividends from oil revenues. Altman wants other leading U.S. AI firms to contribute similar stakes. Whether any competitor would follow remains unclear. Trump called the idea “a beautiful thing” last month and has already taken equity positions in Intel, IBM, and quantum firms during his second term.
The pushback is sharp. Sanders is pushing a harsher bill that would impose a one-time 50% stock tax on OpenAI, Anthropic, and xAI, dismissing the 5% idea as modest profit-sharing rather than genuine public ownership. Policy researchers warn the arrangement turns Washington into both shareholder and regulator of the same company. David Sherman, an AI strategist at io.net, called it “state-sanctioned centralisation of the most transformative technology of our generation.” David Weinstein, CEO of KayOS, said foreign companies should worry: “If you are a UK company, a South American startup or a Korean research lab, your access to critical AI tooling now sits at the discretion of a foreign government’s strategic interests.”
This isn’t just a policy story. It’s a vendor risk story. If the U.S. government takes an equity stake in OpenAI, every business using OpenAI models just gained a new variable in their risk assessment. Regulated industries like finance and healthcare already face strict third-party vendor requirements. A government shareholder adds political exposure to that mix. Ash Govindia of FintechOS put it plainly: “In financial services, you can’t afford to find out your AI vendor is unavailable the same week your regulator starts asking how your decisioning works.”
Source: Yellow, Yellow (Expert Reactions)
GLM-5.2: China’s Cheaper Frontier Model Pressures U.S. Labs
Z.ai, a Beijing-based startup, released GLM-5.2 as an open-weight model that developers say approaches U.S. frontier systems at significantly lower cost. The model has gained attention for coding, reasoning, and agentic AI work, and has climbed developer rankings on platforms like OpenRouter, where pricing data shows it costs far less per token than leading U.S. models.
Reuters reported rising global interest in the model. Executives and researchers are calling it a potential “mini DeepSeek moment,” referencing the Chinese lab that previously proved Chinese developers could compete on cost and efficiency. Unlike earlier Chinese models often dismissed as budget tools, GLM-5.2 has won praise for software coding, reasoning, and agent workflows. OpenAI’s own GeneBench-Pro results acknowledge that the performance gap between GPT models and GLM 5.2 is larger on scientific reasoning than on coding benchmarks, suggesting open-source models are specializing hard in certain domains.
Western enterprise adoption will remain limited. Security, compliance, and vendor trust concerns slow adoption in finance, healthcare, defense, and government work. But for startups, software teams, and businesses in emerging markets, near-frontier performance at roughly one-sixth the cost of closed alternatives could shift purchasing decisions. The pricing pressure matters. OpenAI and Anthropic charge usage-based fees that scale with workload complexity. An open-weight model you can self-host changes the math entirely.
Source: Yellow
GPT-5.6 Sol vs Claude Fable 5: The Coding Benchmark Split
Fresh head-to-head comparisons show a divided picture. OpenAI’s GPT-5.6 Sol tops Terminal-Bench 2.1, a test of command-line coding agents, at 88.8%, with its Ultra mode pushing to 91.9%. Anthropic’s Claude Fable 5 scores lower on that same terminal test, landing between 83.4% and 84.3% according to different reviewers. But Fable 5 owns a much bigger lead on SWE-Bench Pro, the benchmark most reviewers treat as decisive for autonomous software work: 80.3% versus 58.6% for the older GPT-5.5. OpenAI has published no GPT-5.6 figure on SWE-Bench Pro.
Price cuts the other way. Sol is listed at $5 per million input tokens and $30 for output, half of Fable 5’s $10 and $50. Several reviewers argue the sensible setup routes terminal-driven agents toward Sol once it opens up, and repository-level fixes toward Fable 5. On ExploitBench, Sol reportedly matches Mythos-class performance while spending roughly one third of the output tokens, which matters for long agent runs.
Access remains the sharpest divide. Sol is still in a limited preview for roughly 20 government-cleared partners. Fable 5 returned to global availability on July 1 with a temporary usage bonus for paid subscribers through July 7. June turned frontier model access into a moving target: Washington forced Fable 5 and Mythos 5 offline on June 12, citing cybersecurity risks, before Commerce Secretary Lutnick confirmed the reversal on June 30. Nobody outside the Sol preview can independently verify OpenAI’s benchmark numbers yet, a caveat several reviewers flagged.
Quick Hits
Anthropic – Fable 5 returned to global availability on July 1 after a two-week government-mandated shutdown. The company also proposed an industry-wide framework for scoring jailbreak severity, partnering with Amazon, Microsoft, Google, and other Glasswing partners. Practical infrastructure work that doesn’t grab headlines but shapes how the industry handles safety incidents.
OpenAI – GeneBench-Pro launched on June 30, a research-level benchmark for computational biology with 129 problems across 10 domains. GPT-5.6 Sol scores 28.7% at highest reasoning, up from below 5% for GPT-5 when the original GeneBench was built. The company also unveiled Jalapeño, its first custom inference chip built with Broadcom, delivered from design to tape-out in nine months. Both stories broke June 24-30 and were covered in the July 1-2 newsletters.
Mistral AI – OCR 4 launched with bounding boxes, block classification, and inline confidence scores across 170 languages. Priced at $4 per 1,000 pages via API, with a 50% batch discount. Already covered in the July 1 newsletter but worth noting the model continues to receive positive developer feedback.
Hugging Face – Cerebras partnership bringing Gemma 4 to real-time voice AI was the main fresh story (July 1). The open speech-to-speech pipeline uses Nvidia’s Parakeet for recognition, Gemma 4 31B on Cerebras for inference, and Qwen3 TTS for output. Already powers 9,000+ Reachy Mini robots in the wild.
Google DeepMind – June updates remain the latest: Gemini 3.5 Flash with computer use, DiffusionGemma for 4x faster text generation, Gemma 4 12B unified multimodal model. Nothing new since the June batch covered in previous newsletters.
Rundown for July 3, 2026. Sources: Yellow, Anthropic, OpenAI, Mistral AI, Hugging Face, Google DeepMind.