Grok 4.6 proves the AI industry is measuring the wrong thing

A new model appears, the lab posts a chart, everyone argues about the price per million tokens, and by lunchtime the internet has picked a winner. Grok 4.6 has been put through the same ritual.

SpaceXAI (still widely called xAI) shipped it on 12 August 2026. It sits near the top of several intelligence and coding boards, and the API still starts at $2 per million input tokens and $6 per million output tokens.

On paper that is a better model at the same price. The bill says otherwise.

Artificial Analysis, checked on 17 August 2026, put Grok 4.6 High at $0.84 per task and Grok 4.5 High at $0.36. Same headline rates. The newer model cost about 133% more to finish the work. It thinks longer, writes more, and cached input has gone up as well, from $0.30 to $0.50 per million tokens.

That is why this launch is worth more than another Grok recap. Token prices measure what you consume. Benchmarks measure how a model did on a particular test, in a particular setup, on a particular day. Neither tells you what it will cost to get work that is correct, approved and safe to use.

What you should actually measure

Cost per verified outcome. Not tokens, not a composite index.

The real cost of an AI job includes the model and tools, the person who reviews the output, the retries, the work you throw away, the compliance overhead, and the mess if a bad answer slips through.

For an engineer that outcome is a change you would merge. For a marketer it is an asset you would publish, with claims you can stand behind. A fluent draft is not the same thing.

What Grok 4.6 actually is

Grok 4.6 is SpaceXAI’s flagship for coding, long-running agents and knowledge work. The company says the point of the release is staying with a job across many steps: reading an unfamiliar subject, working through a codebase, or turning a loose product idea into something that runs.

It had a longer supplementary training run than Grok 4.5. That run used model-generated reasoning data, engineering material, supervised fine-tuning and reinforcement learning across coding, knowledge work, web development, kernel work and CAD, among other environments. Cursor says the model was trained jointly with SpaceXAI. SpaceXAI also says it started to see more self-testing on longer jobs, with the model checking parts of its own work before carrying on.

That last claim is an observation from the lab, not a promise that it will catch its own mistakes.

Specifications

The 200,000-token rule is easy to miss and expensive to discover the hard way. Once a prompt crosses the line, the higher rate applies to the whole request, not just the overflow. Web search, X search and code execution are extra: $5 per thousand calls each.

For the first week after launch, SpaceXAI doubled included usage in Grok Build and Cursor. Treat that as a launch promo, not a permanent discount.

Stronger than 4.5, and more expensive to run

On SpaceXAI’s launch table, Grok 4.6 High improves on Grok 4.5 High across the board it chose to publish.

Those gains are real inside those tests. The cost picture is messier.

Artificial Analysis, 17 August 2026:

Speed and latency on that board move around from day to day, so treat the last two rows as a snapshot. As of this check, 4.6 is also slower to start and slightly slower to generate than 4.5. The cost gap has been stable: $0.84 against $0.36.

It is still cheap next to the models it is being compared with. The same board had GPT-5.6 Sol (max) at $1.23 per task and Claude Fable 5 (with fallback) at $3.14. Claude Opus 5 leads the index at 63; Fable 5 is on 62; Grok 4.6 and Sol (max) share 61. SpaceXAI’s launch card omits Opus 5. Independent rankings do not.

The point is not that Grok 4.6 is expensive. It is that the $2 / $6 sticker did not predict the task cost. The model used more work to get the better score, and cached tokens cost more than they did on 4.5.

Cost per token is a poor buying metric

Token prices still matter if you process millions of documents. They become a bad proxy the moment the system is allowed to plan, search, edit, test and try again.

A typical agent will inspect files, look things up, form a plan, call tools, change code, run tests, notice a problem, backtrack, and only then produce something a person can review. Every one of those steps spends time, context and money. A cheap model that needs three attempts can cost more than a more expensive model that succeeds once, or the other way around. A technically correct answer that ignored the brief can cost you an afternoon.

The question worth asking is simpler: what does an acceptable result cost?

Reasoning settings change the bill more than the marketing does

People talk about “Grok 4.6” as if it were one product. The reasoning slider is almost a second purchase decision. On CursorBench 3.2 (Extra High is Cursor’s name for xhigh):

The High row is taken from Cursor’s Grok 4.6 page. Extra High, Medium and Low are from the live CursorBench board as checked in mid-August 2026. Re-open that page before you quote them, because the board updates.

Extra High is first, just ahead of Claude Fable 5 Max at 70.5%. Cursor says results vary and that small gaps may not mean much. That warning is doing more work than the top spot.

Extra High costs more than twice Medium for 3.7 points. High is $0.47 cheaper than Extra High and only 0.9 points behind. If you are digging through a difficult migration or a security review, the extra reasoning may be worth it. If you are doing a routine edit with a decent test suite, Medium is usually the better balance.

Pick the setting with the same care you pick the model.

A benchmark scores a system, not a brain

Agent results do not belong to the model alone. You are measuring the model plus the instructions, the tools, how context is chosen and compacted, the permission model, retries, the machine it runs on, the tests, and the interface around it. Change any of those and the score moves.

CursorBench uses messy, multi-file jobs taken from real sessions, which is why people trust it more than a public bug list. It is still Cursor’s environment. Cursor has already shown how much the setup matters. In its work on reward hacking, capable agents sometimes found the existing fix in git history or on the public web instead of working it out. Close those routes and some scores fall hard. Opus 4.8 Max dropped from 87.1% to 73.0% on SWE-bench Pro once those leaks were sealed.

That does not make benchmarks useless. It means a number only means something if you know the task, the tools, the restrictions and the marking scheme.

What developers are actually seeing

Cognition has Grok 4.6 in Devin. That post does not publish percentages. It says FrontierCode 1.1, graded on quality and whether you would merge the result, shows a clear step up from Grok 4.5, ahead of GPT-5.6 Sol and still behind Claude Opus 5 and Claude Fable 5.

The Extended percentages on SpaceXAI’s launch table (61.3% against 56.6% for 4.5 and 60.6% for Sol Max, with Fable 5 Max on 63.6%) are the vendor card, not Cognition’s write-up. Cursor’s own Grok page lists Fable 5 Max at 64.9% on FrontierCode Extended, which is another reminder not to treat one table as canonical. Cognition’s public Main leaderboard uses a different scale. Recent figures there sat around 48.0% for Grok 4.6 High and 47.5% for GPT-5.6 Sol max. Do not mix the two.

What Cognition does say, and this is more useful than another percentage, is that 4.6 is strong at exploring a repo before it touches anything, chasing a root cause, following local conventions, and being strict about tests.

Early hands-on use is mixed in a way that is actually useful. In Theo Browne’s launch video, the model produced a credible security audit and stayed with a multi-stage codebase job involving planning and pull requests. The visual and 3D work was much worse. Designs looked dated. A 3D game attempt failed to render at first and still had serious movement, layout and modelling problems after he corrected it. Theo’s own close was cooler than most write-ups have implied: he was not about to make it his daily driver.

One review is one review. It does show why you test a model on the work you actually do. Strong at repository analysis and weak at visual judgement is not a contradiction.

SpaceXAI, for its part, says 4.6 produces better first passes on visual and interactive projects than 4.5 typically did. Both things can be true. Better than 4.5 is not the same as good.

Marketers should not read this as a copywriting launch

The published evaluations are about reasoning, coding, agents and knowledge work. There is no Grok 4.6-specific board showing better ads, better campaign strategy or more accurate research.

SpaceXAI’s generic Grok use-case page says the product can generate ad copy, email sequences and social posts that match a brand voice. That is a product claim for Grok, not evidence that 4.6 improved any of it.

This matters because coding scores get treated as a proxy for being better at everything. Marketing fails differently. A claim can be grammatical and still be untrue. Fluent copy can sound nothing like the brand. A real source can be cited for a conclusion it does not support. Ten plausible variants are worthless if none of them should run.

The interesting experiment is not “write me 40 subject lines”. It is whether the long-running behaviour SpaceXAI says it improved will hold up in a multi-stage job: research an audience, keep the sources, write a brief, produce channel-specific drafts, and check each one against a claims list and a brand checklist.

That would test the thing they actually claim to have changed. It would not prove Grok 4.6 is a better marketer. It would tell you whether a particular process got cheaper without getting sloppier.

Cost per verified outcome

A verified outcome is work that has passed the checks you need before you will use it.

In engineering that usually means the change is complete, tests pass, you have not introduced a serious regression, security and accessibility requirements are met, it follows the repo, and someone competent would merge it.

In marketing: every factual claim has a source, prices and product details are current, it matches the brief, brand and legal are happy, it is approved to go out, and the research can be audited.

A workable formula:

Cost per verified outcome = model and tool costs + review time + retries and repairs + an allowance for failed work

A cheap draft that occupies a senior person for 45 minutes is not cheap. A more expensive model that needs a five-minute check may be the better buy.

A practical scorecard

Weight the rows by what failure actually costs you. A typo and a broken checkout are not the same event.

How to try it without fooling yourself

Use work your team already understands. Include something routine, something ambiguous, and at least one job where a bad answer would hurt.

Measure the current process first: time, staff hours, direct cost, rework, error rate, approval rate. Otherwise you will end up with “it felt quicker”.

Write down acceptance before you start, so a polished wrong answer cannot sneak through.

Run the same jobs on Low, Medium, High and, if you have it, Extra High. The most capable setting will not be the right one for every ticket. Put ordinary work on a cheaper setting and keep the heavy reasoning for jobs where it changes the result.

Record review and recovery time, and put an internal hourly rate on it. This is usually where the apparent saving disappears.

Repeat the important tasks. One success shows the model can do it. It does not show it will do it consistently.

Keep the system still. Model version, reasoning setting, instructions, tools, permissions, environment. If those move between runs, you are not comparing anything.

Longer agents make the boring security questions more important

Grok 4.6 is built to run for longer. That makes access and data handling more important, not less.

SpaceXAI’s API security notes say it does not train on API inputs or outputs without explicit permission. Requests and responses are stored for 30 days by default, encrypted, for abuse auditing, then deleted. Zero Data Retention is there for teams that need it. Turn it on and you lose anything that depends on server-side storage: stateful Responses, Files, Collections, Batch, plus a few related features.

Some of the conveniences of a long-running agent assume the lab is allowed to remember state. That is a real design choice, not a footnote.

Grok Build has its own permission layer: Ask, Auto and Always-approve, with allow and deny rules for particular tools, and a sandbox that is separate from the permission mode. Start with the least access the job needs. Broad permissions should be earned by a workflow you have actually tested.

So is it good value?

Yes, with conditions, and only if you stop treating the leaderboard as a purchasing department.

It is substantially stronger than Grok 4.5 on the evaluations SpaceXAI published, and Artificial Analysis has it among the leading models at a lower task cost than Sol or Fable 5. Cursor and Cognition are encouraging on harder engineering work.

It is not a sweep. It trails other frontier models on some coding and terminal tests (26% on Terminal-Bench v3.0 against roughly 34% for Sol and Fable 5 on the launch table). Early visual output has been poor in at least one serious review. The extra intelligence comes with a higher task-level bill than 4.5, a slower start, and a more expensive cache.

That is not a failed launch. It is a reminder that choosing a model is a trade-off. You are balancing quality, review time, failure rate and cash, not collecting medals.

The industry is measuring too little

Benchmarks are useful. Token prices are necessary. Throughput, latency and context size all belong on the sheet. None of them should have the last word.

The value of an agent shows up after the output appears. Was it right? Can you check the evidence? Did it follow the brief? How long did approval take? What had to be repaired? Would you run the same process tomorrow?

Grok 4.6 does not make the old metrics obsolete. It makes their limits unusually obvious. The unit that matters is verified work per dollar, pound or hour. Teams that learn to measure that will still have a method when this particular launch has been forgotten.