Launches

Grok 4.7 Closes a Month Where Every Lab Shipped Something

A point release arriving three weeks after the September cluster — and a reminder to check whether your model alias has drifted.

Priscilla Hayes

Senior AI Correspondent

Published 3 min read
Vibrant orange lines and dots form an abstract network on a dark background, evoking technology and connectivity.
Jump to 7 sections

xAI released Grok 4.7 on September 21, 2026, closing out a month in which every major lab shipped something. It is a point release in an established line rather than a new family, which sets a reasonable expectation: incremental capability, changed behavior, same general shape.

Arriving three weeks after the early-September cluster puts Grok 4.7 in an awkward position. The comparisons everyone will run are against models that landed while the benchmark suites were still being updated.

A detailed view of colorful source code displayed on a computer screen, representing modern programming and technology.

What a .7 release usually means

Point releases inside a version line tend to change three things: post-training behavior, tool-calling reliability, and whatever the previous version was visibly worst at. They rarely change the underlying architecture or the context window.

That is not a criticism. Most of the practical improvement between a 4.0 and a 4.7 comes from post-training rather than from scale, and post-training is where tool-use reliability and instruction-following actually live. Those properties matter more to a working agent than another point on a reasoning benchmark.

The thing to watch is whether a point release is additive or substitutive. An additive one keeps the old version callable while adding a new identifier. A substitutive one silently repoints an alias, which means a prompt you tuned last quarter is now running against different weights without anything in your code changing.

What xAI published for Grok 4.7

xAI's Grok 4.7 announcement calls it the company's most powerful model for coding and knowledge work. It says the model was trained to work natively with the Grok Bot harness, and that it is available in Cursor, Grok Build, the Grok API, third-party coding harnesses, model routers and cloud platforms.

The announcement lists a starting price of $2 per million input tokens and $6 per million output tokens, the same as Grok 4.6, plus a fast variant at twice the output speed for twice the price. It does not state a context window. Secondary coverage from Let's Data Science reports 500,000 tokens, text and image input, and four reasoning settings from low to xhigh.

On the benchmark comparisons

Treat every cross-lab comparison published this month with extra caution, because the reference points moved while the tests were running. A chart comparing a late-September model against early-September competitors is comparing against whatever version the chart's author had access to, which may not be what shipped.

A complex network of cables in a data center with a monitor in the foreground.

Model Drop has not independently verified xAI's published figures for this release, and we would say the same about every launch chart in September. The scaffolding question — what harness, how many retries, which verifier — determines the number as much as the weights do, which is the argument in our guide to reading a model launch announcement.

Independent evaluations will post their own numbers on their own schedule, typically weeks behind a launch: coding benchmarks like SWE-bench and human-preference leaderboards such as LMArena. Those are the ones worth waiting for.

Where it sits on the price ladder

The relevant question for anyone shipping is which rung this occupies, not where it ranks. The market has settled into a flagship band around $10 per million input tokens, a workhorse band near $2 to $4, and a cheap band in cents — a structure we map in our breakdown of LLM API pricing across the 2026 tiers.

A diverse group of professionals engaged in a collaborative office meeting with laptops and a whiteboard.

A new entrant changes your architecture only if it beats the rung you are already on, on your own evaluation set, at a price you can defend. Ranking third on a public leaderboard while costing more than what you run today is not a reason to migrate.

The long-context price tier

The $2 and $6 figures apply only to shorter prompts. According to Let's Data Science, once a request reaches 200,000 input tokens the higher rates of $4 input and $12 output per million apply to the entire request, not only the tokens past the threshold. The mer.vin analysis notes that this tier is documented in xAI's API docs rather than in the announcement.

For teams that send long contexts, the headline price is therefore not the one that will appear on the invoice. Run your real prompt lengths through the pricing table before comparing Grok 4.7 to a cheaper rung.

What the benchmark table shows

On xAI's own table, Grok 4.7 at xhigh effort scores 46.3% on CursorBench 4.0, against 40.4% for Grok 4.6 at high effort. On Terminal-Bench 4.0 it scores 37.6%, against 20.3% for its predecessor. The same table shows Fable 5.1 Max ahead on both, at 51.8% and 57.9%.

The effort levels are the catch. xAI mostly compares Grok 4.7 at xhigh with Grok 4.6 at high, so the gains are not a like-for-like measurement. All of these figures are the vendor's own.

There is a cheaper comparison available too. Google's Gemini 3.8 Flash reached stable GA earlier in the same month at well under a dollar per million input tokens, and for a large share of production traffic the relevant question is whether anything at the flagship band is needed at all.

Independent evaluation infrastructure is worth knowing about here, because it is what eventually settles these comparisons. MLCommons publishes standardized MLPerf benchmark suites with defined submission rules, which is a different quality of evidence from a vendor chart precisely because the methodology is fixed in advance and the results are audited.

Check the availability terms too. Regional restrictions, rate limits by account tier, and whether the model is exposed through the same API surface as its predecessor all determine how much work a migration actually is. Those details are usually in the docs rather than the announcement.

The practical read

For most teams, September 2026 produced four new options and no obligation to act on any of them. The sensible response to a dense release month is to re-run your evaluation set once, at the end of it, rather than chasing each announcement as it lands.

At Model Drop we would flag one genuine risk from a month like this: alias drift. If you call a model by a floating identifier rather than a pinned version string, several of these releases may already be serving your production traffic. That is worth checking today regardless of what you think of any individual launch.

By Greg Halston, Staff Writer at Model Drop. Reported September 21, 2026. Capability claims are the vendor's and have not been independently verified by Model Drop.

Model Drop covers AI launches — new models, platforms, features, and tools — for the people who have to decide what to actually ship on.

What usually changes in a point release like Grok 4.7?
Post-training behavior, tool-calling reliability, and whatever the previous version handled worst. Architecture and context window rarely change. That is not a small update, since tool-use reliability and instruction-following matter more to a working agent than another point on a reasoning benchmark does.
Why are September 2026 benchmark comparisons unreliable?
The reference points moved while the tests were running. Four major labs shipped within weeks of each other, so a chart comparing a late-September model against early-September competitors may be measuring against whichever version the chart's author had access to rather than what actually shipped.
What is alias drift and why does it matter?
If your code calls a model by a floating identifier rather than a pinned version string, a vendor update can silently repoint it to new weights. Your prompts then run against a different model with no code change on your side. After a dense release month, this is worth auditing immediately.
Should a dense release month change my model choice?
Not request by request. The sensible response is to re-run your own evaluation set once at the end of the month rather than chasing each announcement. A new model justifies migration only if it beats your current rung on your own tests at a price you can defend.

Written by

Priscilla Hayes

Senior AI Correspondent

8+ years experience

Priya has covered model releases since the first wave of chatbot launches and has never met a benchmark leaderboard she didn't immediately try to break. Before Model Drop, she ran technical comms for a model-eval startup, which means she reads system cards for fun and gets suspicious of any release that ships without one. She keeps a running spreadsheet of every 'frontier' claim made in a press release and how long it survived contact with actual users.

Credentials

  • 8 years covering AI model releases and infrastructure
  • Former technical communications lead, an independent model-evaluation lab
  • Built and maintains a public tracker of context-window and pricing claims across 40+ model families