AI model
Grok 4.5 is a large language model from xAI, part of the Grok family, evaluated in coding, security and fact-checking benchmarks.
Hacktron reports that GPT-5.6 Sol Ultra built a complete Chrome/V8 exploit chain in a controlled test for $1,596.89 of model compute. The benchmark handed the model the source tree and the public security-fix commits, so the figure covers only the compute for one lab task.
Reality
Ridge Security's 96-test pen-test benchmark had Claude Opus 4.6 reach 63% coverage at $217 a run, against 52% for Gemini 3 Flash at about $5.42. Ridge argues tooling matters more than the model, yet its published runs held the tooling fixed and varied the model.
Reality
DeepSeek 4.1 Flash finished a metered Ship-Bench build for $15.04 in per-token fees and scored lowest at code review. Developers who run more than about one and a third full builds a month would still pay less on a $20 seat.
Reality
SpaceX has closed its $60bn purchase of Cursor. Cursor's own note argues the GPU fleet makes its models cheaper to serve, and commits to no leadership or product change.
Perspective Coverage
4 publishersReality
The always-on agent product opened in beta on August 11. The docs cap accounts at 50 bots and chats, give every bot on an account the same computer, and skip Linux desktop entirely.
Reality
Five frontier LLMs with web search disagreed on 63% of 997 claims users sent to Lenz.io for checking. They also rated themselves 9 or 10 out of 10 on 76% of answers, so one model's confidence says little about whether the others would agree.
Reality
The independent checker cereblab captured grok 0.2.93 sending a never-read repository to a Google Cloud Storage bucket. The privacy opt-out xAI shipped afterwards governs retention, cereblab says, while the same bytes still leave the machine.
Reality
A dev.to writeup of Bottleneck Labs' 72-hour run reports $0 in revenue and $12,431 in unsolicited invoices, because the only outbound limit that fired belonged to the email provider, and Stripe has its own delivery path.
Reality
Booz Allen scored 18 frontier models on a live intrusion and placed Claude Sonnet 5 fifteenth, then paired it with an attack harness and watched it rival the winner. The result: anyone tiering risk by model name is reading a column that measures the wrong object.
Reality
Nineteen days after the same model name produced repetition loops and phantom missing files, the served checkpoint refused 59.2% of concealed-hazard tasks. That makes third-party refusal testing a running job, not a procurement step.
Reality
Muse Spark 1.2 costs roughly 18 times less if Meta may train on your prompts. The deepest cut, 75x, sits on cached input, which is where an agent holding your codebase in context spends.
Reality
A Fortune columnist reads the 2026 AI economy as an unstable system of frontier labs, Chinese open weights and app companies. The checkable part is price, not payoff.
Reality
Seven days after launch, xAI's flagship sits inside AWS procurement with a 500K context and four reasoning tiers. The rate card is flat; the effort dial is where the cost moves.
Reality
An independent researcher says the CLI uploaded a repository it was told not to read, plus a .env secrets file, verbatim. That is a procurement question, not a benchmark question.
Reality
Grok 4.6, Gemini 3.7 Flash, DeepSeek V4 Pro and GLM-5.3 all chase agents that stay on task. The pricing underneath them is moving faster than the benchmarks.
Reality
A production test across 15 models put seven of them inside a one-point spread on pass rate. On constrained payroll work, the price premium bought speed, not correctness.
Reality
GitHub added xAI's model to Copilot on August 14 across eight developer surfaces at usage-based pricing. The benchmark case, including xAI's own terminal scores, is mixed.
Reality