- Lead. July 10, 2026 is the first day in the history of large language models that three frontier labs — OpenAI, xAI, and Anthropic — released new flagship products simultaneously, compressing what once unfolded over months of staggered launches into a single 24-hour window.
- Fact. OpenAI’s top-tier Sol Ultra model priced at $5 per million input tokens scored 91.9% on Terminal-Bench 2.1; xAI’s Grok 4.5 undercuts both rivals at $2 per million but its measured hallucination rate jumped from 25% on Grok 4.3 to 54%, the highest of any frontier model.
- Stake. Simultaneous releases eliminate the first-mover pricing window that labs have historically depended on; the competitive clock now runs in days, not months, intensifying cost pressure and accelerating the commoditisation of raw model capability.
OpenAI launched GPT-5.6 in three distinct tiers. Sol Ultra, priced at $5 per million input tokens and $30 per million output tokens, leads Terminal-Bench 2.1 at 91.9% and scores 50.9% on Agent’s Last Exam. Terra, at $2.50/$15, aligns with the performance level of GPT-5.5. Luna, the entry tier at $1/$6, scores 84.3% on Terminal-Bench and is engineered to outperform Terra on high-volume routine tasks where cost efficiency matters more than peak capability. OpenAI simultaneously folded its standalone Codex desktop application into a unified ChatGPT interface featuring Work, an agentic productivity layer that connects across fifteen integrated platforms including Google Drive, Slack, Salesforce, and GitHub.
Grok 4.5: strong agentic benchmarks, alarming hallucinations
xAI’s Grok 4.5, priced at $2 per million input and $6 per million output tokens, ranks fourth on the Artificial Analysis Intelligence Index with a score of 54 and claims the best agentic tool-use performance of any currently available model. On Snorkel’s GDPval+ professional-task benchmark, Grok 4.5 scores 29% against GPT-5.5’s 22%, a meaningful gap in agentic real-world tasks. xAI claims the number-one position on the SWE Marathon benchmark for software engineering and cites a 4.2x token-efficiency advantage over Anthropic’s Opus 4.8.
However, the independent evaluation firm Artificial Analysis flagged a sharp deterioration in factual reliability. Grok 4.5’s hallucination rate measured at 54%, nearly double the 25% recorded for Grok 4.3 — “the model knows more,” Artificial Analysis wrote, “but it is also more confident when it is wrong.” The model is not yet available in the European Union, with a mid-July launch pending regulatory review.
Anthropic’s Cowork arrives on mobile
Anthropic’s contribution to the day was the general availability of Claude Cowork on iOS, Android, and web simultaneously. Cowork is designed for long-running agent sessions that persist and resume across devices without requiring a desktop connection — a direct response to enterprise demand for AI workflows that operate across distributed teams. The mobile launch had been telegraphed but the timing, landing on the same day as GPT-5.6 and Grok 4.5, was not.
The compression of releases into a single day has structural implications. Previously, a lab that shipped a strong model gained weeks of exclusive benchmark visibility and pricing power before rivals responded. That window appears to be closing. As noted earlier this month, OpenAI had already been navigating restrictions on GPT-5.6 access for certain user categories before this broader public launch, suggesting the product cycle is now being managed around both commercial and policy timelines simultaneously.