Two Frontier AI Models Ship on the Same Day: Grok 4.5 and GPT-5.6

Two of the biggest names in AI shipped flagship models on the same day. xAI released Grok 4.5, and OpenAI opened up GPT-5.6. They arrive from different directions, one as a coding-first challenger and one as a heavily gated release cleared by Washington only this week, but they land on the same argument: the model that costs less to run well is the one that wins developers.

Grok 4.5 is the coding play

Grok 4.5 runs on V9, xAI's new 1.5 trillion parameter foundation. That is roughly three times the size of the v8-small architecture behind Grok 4.3, which makes this a real change of base model rather than another round of fine-tuning. It's the largest model the company has shipped.

The focus is coding. According to xAI, the model was trained on real developer session data from Cursor, the AI coding editor, so it learned from actual debugging traces and multi-file edits rather than static code alone. Musk is calling it an "Opus-class" model, meaning he claims it performs around the level of Anthropic's Claude Opus 4.8 while running faster, cheaper, and on fewer tokens. Reporting puts the pricing near $2 per million input tokens and $6 per million output.

Grok 4.5 is live now inside Grok Build and in Cursor, though it is not yet available in the European Union. It had been in private beta at SpaceX and Tesla since late June before this public release.

GPT-5.6 is three models, and a political story

OpenAI's GPT-5.6 isn't a single model. It's a family of three. Sol is the flagship for hard reasoning and agentic coding. Terra is the balanced tier, which OpenAI says matches GPT-5.5 at about half the cost. Luna is the small, fast, cheap option built for high volume. The idea is that you send each task to whichever tier fits, rather than paying flagship rates for everything.

The new features worth noting are a "max reasoning effort" setting that lets Sol think longer on a problem, and an "ultra" mode that splits a task across multiple subagents working in parallel. OpenAI is positioning the family around agentic coding, biology, and cybersecurity as the areas where it pulls ahead of GPT-5.5.

The more interesting part of this launch is how it got here. OpenAI first previewed GPT-5.6 on June 26 to a limited group of roughly 20 trusted partners, held there at the request of the U.S. government because of the model's cybersecurity capability. Per OpenAI, the U.S. Department of Commerce cleared a broad rollout after additional testing, and Sam Altman confirmed the public launch on X. Under its own preparedness framework, OpenAI rates the models high in cyber and biology capability, but below its "critical" threshold.

The common thread is cost

Strip away the branding and both launches are making the same bet. Neither lab is leading with "we have the smartest model." Both are leading with price and token efficiency, because for anyone running these models at scale, the number that matters is cost per solved task, not a few points on a benchmark. xAI is pitching Grok 4.5 as an Opus-level model that burns fewer tokens to get there. OpenAI is pitching a cheaper Terra and a very cheap Luna so teams stop overpaying for simple work.

That reflects where the market has moved in 2026. The frontier is crowded enough that raw capability is table stakes, and the competition has shifted to who can deliver that capability at the lowest effective cost.

Benchmarks: