News

OpenAI and xAI Launch Competing Models on the Same Day, Independent Benchmarks Diverge from Vendor Claims

• By John K. Waters
• 07/10/2026

OpenAI and xAI both released major new AI models on July 9, a coincidence of timing that several industry outlets called one of the most competitive single days in the sector's history. OpenAI brought its GPT-5.6 family, consisting of three tiers named Sol, Terra and Luna, to general availability, while xAI launched Grok 4.5. Independent benchmarking conducted in the days since has told a more complicated story than either company's launch materials suggested.

OpenAI began a limited, government-reviewed preview of GPT-5.6 on June 26 with roughly 20 approved partners before opening all three tiers to the public through the application programming interface, ChatGPT and Codex on July 9. In its release materials, OpenAI said Sol is intended as its flagship for the hardest problems, Terra as a balanced option priced roughly half of Sol, and Luna as its fastest and cheapest tier. All three share a 1.05-million-token context window and a 128,000-token maximum output. Sol also includes two heavier compute modes, Pro and Ultra, that run on the same underlying model rather than as separate priced tiers. Ultra coordinates four parallel reasoning agents by default and can scale to sixteen, a design OpenAI said trades higher token usage for stronger results on demanding tasks.

OpenAI's own published figures show Sol scoring 88.8 percent on Terminal-Bench 2.1, a benchmark for command-line workflows requiring planning and tool coordination, rising to 91.9 percent in Ultra mode. Independent analysis firm Artificial Analysis measured Sol at approximately 59 on its aggregate Intelligence Index, placing it just behind Claude Fable 5, and found Sol led its Coding Agent Index outright. On the ARC Prize Foundation's ARC-AGI benchmark, an independent evaluation, Sol scored 96.5 percent on ARC-AGI-1 and 92.5 percent on ARC-AGI-2. OpenAI's own comparison table acknowledged that Claude Fable 5 and Claude Mythos 5 rank above Sol on SWE-Bench Pro, a coding benchmark, a rare instance of a vendor publishing a result unfavorable to its own flagship model.

xAI's Grok 4.5 launched the same day with less public documentation. The company's official announcement page returned an error during initial reporting, leaving benchmark, pricing and technical specifications largely unverified from primary sources in the hours after launch. On July 9, Elon Musk wrote on the social media platform X that Grok 4.5 had reached the top spot on multiple benchmarks, citing in particular the SWE marathon, a benchmark testing extended, multi-step software engineering tasks. After 24 hours of independent benchmarking, Artificial Analysis ranked Grok 4.5 fourth on its Intelligence Index with a score of 54, behind Claude Fable 5, GPT-5.5 and Claude Opus 4.8. Separate benchmark data published by Dervity put Grok 4.5's pricing at $2 per million tokens for input and $6 for output, positioning it as a low-cost option that scores between OpenAI's Terra and Luna tiers in terms of intelligence while matching Luna's output price. Reviewers highlighted the model's token efficiency as a genuine strength, though the magnitude of that advantage varied by source, and noted a hallucination rate that multiple outlets flagged as high relative to competing frontier models.

The near-simultaneous launches were not entirely coincidental. Both companies have spent recent months racing to close gaps against Anthropic, whose Claude Fable 5 and Claude Opus 4.8 models continue to lead or place highly on several independent benchmarks cited by both OpenAI and third-party evaluators. OpenAI's release also arrived under unusually close government involvement. The company said in its launch materials that it previewed GPT-5.6's capabilities to the United States government ahead of release and, at the government's request, began with a restricted preview for a small group of vetted partners before broadening access. OpenAI's accompanying system card rated all three GPT-5.6 tiers as having high cybersecurity capability, meaning the models could identify vulnerabilities and assemble pieces of exploits, but rated them below the threshold the company considers critical, meaning they could not autonomously carry out complete attacks against hardened targets in OpenAI's testing.

For enterprise technology buyers, the split between vendor-reported and independently measured results is the more consequential story than either launch on its own. OpenAI's tiered pricing structure gives procurement teams a formal mechanism to match model cost to task complexity rather than defaulting to the most expensive option, a shift that mirrors similar tiering strategies already adopted by Anthropic and Google. At the same time, the gap between xAI's launch-day claims and Artificial Analysis's independent ranking is a reminder that headline benchmark claims from any vendor warrant independent verification before informing a purchasing or deployment decision, particularly for agentic workloads, where a model's hallucination rate carries direct operational risk.

OpenAI has announced that GPT-5.4 will be retired on July 23, prompting remaining users to migrate to the GPT-5.6 family, while GPT-5.5 will remain available. Google's Gemini 3.5 Pro remains in limited enterprise preview roughly six weeks past its original public release target, a delay that industry observers say raises the competitive stakes for its eventual launch. Independent benchmarking of Grok 4.5 was still in its early stages as of this writing, and xAI has not yet published the fuller technical documentation typically accompanying a frontier model release.