Grok 4 Quietly Closed the Coding Gap, and Nobody Noticed

Grok has spent two years being the model people mention as a punchline rather than a option. Grok 4, released June 11, is the first version where that stops being fair.

The Numbers

HumanEval went from 84.3 on Grok 3.5 to 90.6. SWE-bench Verified went from 44.2 to 62.8. The context window went from 128K to 1M. None of those individually would be remarkable in mid 2026. Together, in one release, they represent the largest single-version jump any provider has posted this year.

It still is not top of our leaderboard. Claude Fable 5 and the Opus 5 tier are comfortably ahead on agentic work, and GPT-5.6 Sol beats it badly on hard reasoning. Grok 4 sits ninth on TC Score at 88.7, which is respectable rather than exciting.

The Part That Is Actually Differentiated

Live grounding. Grok remains the only major model with direct, real-time access to a large social firehose plus web search, and that is not a feature the others can trivially copy because they do not own the firehose.

For most coding work this is irrelevant. For anything touching current events, market sentiment, breaking security disclosures or the state of a fast-moving open-source project, it is the difference between an answer and a guess. Ask a frontier model about a library that shipped a breaking change last Tuesday and you get confident nonsense. Ask Grok and you get the changelog.

Grok 4 Mini Is the Sleeper

The Mini tier runs at $0.40 per million input tokens and keeps the live grounding. That is genuinely unusual. Every other cheap model in that price band is working from a frozen training snapshot. If you are building anything that monitors, summarises or reacts to current information at volume, there is nothing else at that price with fresh data.

Check the current rates on our price comparison page, and estimate throughput for your workload with the speed calculator.

The Reasons People Still Will Not Use It

Three, and they are all non-technical.

1. Enterprise procurement. xAI does not have the compliance paperwork trail that OpenAI, Anthropic and Google have accumulated. For regulated buyers that is a hard stop regardless of benchmark scores.
2. Brand risk. A meaningful number of teams will not put the model in a customer-facing product for reasons that have nothing to do with the model.
3. Tooling gravity. The agentic coding ecosystem has standardised around a handful of providers. Grok works fine through OpenAI-compatible endpoints, but it is nobody's default.

Where It Fits

Internal tools that need current data. Research assistants. Anything where you are the customer and the compliance surface is your own laptop. For those, Grok 4 is now a real option and Grok 4 Mini is arguably the best value in its band.

For customer-facing production work, the honest answer is that the ecosystem reasons still outweigh the capability reasons. That gap is narrower than it was, though, and if the next version closes it further the calculus changes.

See the full spec on the Grok 4 model page.