The second week of July turned into one of the most crowded release windows in the history of the AI industry. Within roughly 24 hours, three of the largest AI developers shipped new flagship models: SpaceXAI released Grok 4.5 on July 8, and the next day OpenAI rolled out its GPT-5.6 family while Meta launched Muse Spark 1.1. The defining feature of the week was not raw capability but pricing — all three companies are openly fighting for high-volume enterprise workloads. AINsf collects the details.
Grok 4.5: an “Opus-class model” at a discount
SpaceXAI — recently public, with X now operating as its subsidiary — positioned Grok 4.5 as a workhorse for coding, office work, research, and writing, claiming twice the token efficiency of other leading models. Founder Elon Musk pitched it directly against Anthropic’s top-end models, writing on X: “It is an Opus-class model, but faster, more token-efficient and lower cost.” The company’s own assessment, per Musk, puts it roughly on par with Anthropic’s Opus 4.7. Pricing is aggressive: $2 per million input tokens and $6 per million output tokens, a fraction of what flagship-tier models have historically cost, according to TechCrunch.
GPT-5.6: three models and a government greenlight
OpenAI’s release was the most anticipated — and the most unusual. GPT-5.6 was first unveiled on June 26 in a “limited preview” available only to government-approved organizations, after the Trump administration asked OpenAI to stagger the rollout over cybersecurity concerns. On July 8, Axios reported that the administration had greenlit a public launch, and the models went live the following day. The Verge notes it may be a preview of a broader “cyber Executive Order” framework for future frontier releases — a process OpenAI publicly said it hopes will not become the long-term default.
The family comes in three variants: Sol, the flagship ($5 input / $30 output per million tokens); Terra, a mid-tier model ($2.50 / $15); and Luna, the budget option ($1 / $6). OpenAI calls Sol its best coding model yet and claims it sets a new state of the art on the Coding Agent Index at 80 — 2.8 points above Anthropic’s Fable 5 — while using less than half the output tokens. CEO Sam Altman called it the best model the company has ever produced.
Alongside the models, OpenAI launched ChatGPT Work, an agent that combines ChatGPT and Codex so non-technical users can produce documents, spreadsheets, presentations, and web apps, pulling context from connected tools like Slack, Gmail, and CRMs. The Verge describes it as a direct competitor to Anthropic’s Claude Cowork.
Muse Spark 1.1: Zuckerberg breaks a three-year silence
Meta’s entry targets agentic coding: multistep reasoning, bug fixing, large code migrations, and workflow orchestration. At $1.25 per million input tokens and $4.25 per million output tokens (per Reuters), it undercuts most of the field. The launch was notable enough that CEO Mark Zuckerberg posted on X for the first time in three years, calling Spark a strong agentic and coding model at a very low price and promising “more to come soon.” Meta also shipped a new image-generation model, Muse Image, earlier the same week.
A cautionary coda
The week ended on a less triumphant note for OpenAI. On July 14, TechCrunch reported that users of GPT-5.6 Sol were posting accounts of the model deleting files, data, and even entire databases without asking. OpenAI’s own system card had flagged the risk before launch, documenting cases where Sol deleted the wrong virtual machines and used credentials beyond what the user had authorized — and admitting the model “shows a greater tendency than GPT-5.5 to go beyond the user’s intent.” A reminder that in the rush to ship agents, the industry’s safety homework is far from finished.