Two Points Apart: When AI Price Beat Performance

So, What Actually Happened?

Thursday I was staring at a benchmark table and it took me a minute to work out why it bothered me. Grok 4.6 scored 61 on one intelligence index, tying GPT-5.6 Sol Max, sitting two points under the leaders. Two points. I have spent years telling people the model choice is the strategic decision. We scanned 190,000 articles this week so you don't have to. Same day, an analyst piece argued price is now the benchmark that counts, with Palantir's Alex Karp on record saying token bills are breaking corporate budgets. Google shipped a new Flash model pitched on being the cheap workhorse. The question changed under me and nobody announced it.

The Bottom Line: Capability stopped being the differentiator this week. The invoice took over, and most AI budgets still have no line for it.

The Tracks That Matter

1. The Top Three Models Finished Two Points Apart

xAI's Grok 4.6 landed at 61 on the Artificial Analysis Intelligence Index, tying GPT-5.6 Sol Max and trailing the leaders by two. The write-up is more interesting than the number: they taught an existing model to check its own work mid-task rather than making it bigger. That is what a maturing product looks like. Alongside it, the argument that price is the benchmark that matters put a name to what buyers have been muttering since spring. When the top of the table is separated by rounding error, the thing you are actually purchasing is completed work per euro, and almost nobody's evaluation rubric from last year has a column for that.

Here's what works: Rerun your model evaluation with cost per completed task as the first column instead of the last. The ranking usually moves.

2. Flock Cut Its Camera Data Retention From 30 Days To Seven

After months of investigative reporting and public pressure, Flock rewrote how its license-plate camera network operates: searches now require a case number, unusual query patterns get reviewed automatically, and the default retention on the footage drops from 30 days to seven. Read it as a data story rather than a surveillance story and it is the clearest retention climbdown of the year, forced by journalists rather than by a regulator. The regulators are moving anyway. Thailand opened a public consultation on a draft data-sharing law the same week. Retention windows are the cheapest control to change and the most expensive one to have gotten wrong, because the exposure is already sitting on disk.

Here's what works: Pull the default retention setting on your three largest data stores this week. If nobody can explain why the number is what it is, it was a default, not a decision.

3. Form Energy Took $750 Million While Insurers Priced The Boom

Form Energy closed $750 million led by T. Rowe Price, with Sequoia among the investors, for iron-air storage manufactured in West Virginia. Storage is an unglamorous bet, which is roughly the point. The same day Allianz published a report on construction risk in the data-center boom, which is what it looks like when insurers move from quietly covering something to actively pricing it. Underwriters are usually the last people invited to the party and the first to notice the floor sagging. The compute story has been about chips for three years. The constraint moved to power, land, water, and now premiums.

Here's what works: Ask your cloud vendor where the capacity you are contracting for 2027 gets its power. If the answer is a roadmap, you have a delivery risk, not a supply agreement.

Quick hits:

• Microsoft deleted the AI features nobody used. It killed off several unsuccessful AI features and merged its separate Copilot apps, making it the first big vendor to publicly subtract rather than ship.
• IBM and OpenAI signed a delivery pact. The two formed an AI consulting and delivery partnership, which tells you the bottleneck the model vendors now care about is implementation labour, not model quality.
• Broadcom expects $100 billion of AI chips by 2027. A market analysis puts Broadcom's AI-chip revenue past $100 billion on custom-processor demand, alongside an AMD and Meta agreement to deploy up to 6 GW of GPUs.

Signal vs. Noise

🟢 Signal: Data quality. The specialist writing about data quality got quieter this week and considerably more consequential: fewer pieces named it, far more of what actually moved depended on it. That is the shape of a topic turning from a talking point into a precondition. One trade piece said it plainly, arguing enterprise AI's next bottleneck sits beneath the model. Most coverage is still busy grading models.

🔴 Noise: ”AI governance.” The phrase pulled heavy volume again and lost ground on both counts, said more often while less and less actually hangs off it. UiPath named the reason out loud: the governance gap is an architecture problem, not a policy one. Policies are cheap to write, which is exactly why there are so many of them.

From the 190K

We scanned 190,000 articles this week. Here's what no one's talking about:

Grok tied the second-best model in the world, Google shipped a Flash model sold on being cheap, and an entire product category surfaced whose only job is metering AI tokens and charging them back to internal teams.

Three desks filed those separately. The model desk covered the benchmark. The cloud desk covered the launch. The platform desk covered the plumbing, where inference platforms now ship quotas, showback and chargeback as headline features, with operators setting their own rate per million input and output tokens. Read all three on one morning and the sequence is uncomfortable: the industry finished building the billing department before it finished building the productivity case.

That plumbing only gets funded after somebody senior asks a question nobody can answer. Which team spent what, on which model, for what result. Chargeback is not really a finance feature. It is the fingerprint of an argument that already happened somewhere expensive.

What changes on Monday is small and slightly awkward. Find out whether anyone in your organisation can produce AI spend broken down by team. If they cannot, the next person to ask will be the CFO, and that version of the meeting goes worse.

By The Numbers

• Grok 4.6 scored 61 on the Artificial Analysis Intelligence Index — tying GPT-5.6 Sol Max and landing within two points of the leaders, which is the entire story of where differentiation went.
• OpenAI previewed an Ultrafast mode running GPT-5.6 Sol up to 14 times faster — speed became a headline feature, which is what gets sold when capability has converged.
• Form Energy raised $750 million led by T. Rowe Price — long-duration storage is drawing growth-round money, which happens when serious people have modelled a power shortfall.
• Broadcom's AI-chip revenue is projected past $100 billion by 2027 — custom silicon demand at a scale that reprices everyone's assumed GPU bill.
• Cognition AI is in talks to raise over $1 billion at a $40 billion valuation — private capital still paying capability prices in the week the public scoreboard said capability converged.
• See what's rising across AI and data this week →

Deep Dive: Everyone Got The Same Record Box

There was a stretch around 2010 when every DJ I knew was carrying the same USB stick. Same pool, same promos, same week. You would walk past the next tent and hear your own set played back at you. The gear stopped being the difference, the crowd could not tell any of us apart, and the bookings drifted to whoever was cheaper, closer, or easier to deal with.

The chart went flat at the top
61, 62, 63. Those are the leading scores on one intelligence index this week, and the distance between them is narrower than the error bars on most enterprise evaluations. Grok got there by teaching an existing model to verify its own steps rather than by scaling it up. That is a maturity signal, and maturity is where margins go to die.

So the fight moved to the door price
Google pitched its newest Flash model as the cheap workhorse. OpenAI previewed a mode running up to fourteen times faster. Palantir's chief executive said token bills are breaking budgets. Not one of them led with a new capability, because there is not one available that the others will not have by October.

And somebody quietly built the till
The most telling shipping this week was not a model. It was inference platforms adding token metering, quotas, showback and chargeback, so an operator can set a price per million tokens and bill an internal team for it. That is what a market looks like the week after it stops being magic.

What Actually Works

1. Put cost per completed task first: rank models on what finished work costs, not on the index score. Two points of capability rarely survives contact with a 3x price gap.
2. Get spend visible by team before you get asked: token metering exists as a product now. If you cannot answer ”who spent what on which model”, buy or build that this quarter.
3. Renegotiate on convergence: your vendor knows three alternatives sit within rounding error of them. That is leverage, and it expires the moment somebody pulls ahead again.
4. Check the power under the contract: for any 2027 capacity commitment, ask who supplies the electricity and who carries the construction risk. Insurers already started asking.

Same record box, different door price. The ones who kept working were the ones who knew what the night actually cost them.

What's Coming

The AI Trade Lane Between Allies

The US launched an effort to speed trade in AI goods between allies. Export paperwork is about to become a procurement variable in the next hardware cycle, and ”which allies” is the sentence everyone will spend the autumn arguing about.

India Starts Writing Its Own Rulebook

India is weighing dedicated AI legislation. If it lands close to Europe's shape, the compliance work duplicates. If it lands somewhere else, every multinational adds a third regime to the map.

The Comparison Nobody Wants Made

Barron's flagged that AI funding plans carry echoes of earlier bubbles. When the financing structure becomes the story instead of the technology, the market has quietly moved its attention to the balance sheet.

For Your Team

Monday's meeting prompt: ”If our chosen model sits two points from three alternatives, what exactly are we paying the premium for? And can anyone in this room produce our AI spend broken down by team?”

Share-worthy stat: The top three AI models this week scored 61, 62 and 63 on the same intelligence index. The gap between best and third-best is smaller than the rounding on most enterprise evaluations, and the price gap between them is not.

Go deeper: Track where model pricing and AI spend are separating →

The Track of the Day

”They taught the existing model to think and work better instead of making it bigger.”
From the Grok 4.6 write-up

Every DJ eventually learns the set does not improve by adding tracks. It improves by cutting the ones that were never doing any work.

We scanned 190,000 articles this week so you don't have to. Data Pains → Business Gains.

Published: August 14, 2026 | Curated by Yves Mulkers @ Ins7ghts