Home > Fuel Room > Battle of the bots: what the AI price war actually means for CMOs

Battle of the bots: what the AI price war actually means for CMOs

25 Sep 2026

Battle of the bots: what the AI price war actually means for CMOs

The short answer: On 22 September 2026, Anthropic launched Claude Opus 5.5 and OpenAI launched GPT-6 Sol and Luna, all with major price cuts. For CMOs, the lesson is not “switch models”. It is this: when AI gets cheap, generic content, token-counting and loyalty to one AI provider all lose value. Proof, expertise, governance and being easy for AI to quote gain it.

On Tuesday, the two biggest AI labs launched new models within about 90 minutes of each other. Anthropic released Claude Opus 5.5, and roughly 90 minutes later OpenAI responded with GPT-6 Sol and Luna.

Somewhere, a procurement team that had just signed an annual AI contract quietly closed its laptop.

The coverage has mostly been a sports report. Who topped which leaderboard. Who undercut whom. Fun, but not especially useful if your job is to grow a charity’s donor base or fill a B2B pipeline.

So here is the version for the people who have to explain this to a board.

What actually launched (for those who skipped the livestream)

Right. Time to get geeky for a minute. We promise there is a point at the end.

	Claude Opus 5.5	GPT-6 Sol	GPT-6 Luna
Maker	Anthropic	OpenAI	OpenAI
Price (per 1M tokens, input/output)	$4 / $20	$2 / $10	$0.10 / $0.50
Change vs previous version	20% cheaper per token. Anthropic claims ~40% cheaper on typical workloads	Half the price of GPT-5.6 Sol	Roughly half the price of GPT-5.6 Luna
Positioning	Top-end. Matches Fable 5.1 on most work	Mid-range. Coding and automated tasks	Budget. High-volume routine jobs
Independent ranking	#1 on Artificial Analysis Intelligence Index (58)	Not yet independently compared like for like	Not yet independently compared like for like

On Anthropic’s side, Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, which Anthropic says comes to 40% less than Opus 5 on a typical workload. Independent tester Artificial Analysis put it top. At its maximum effort setting, Opus 5.5 scores 58 on the index, five points ahead of GPT-6 Astra and Claude Fable 5.1, which are tied on 53.

On OpenAI’s side, Sol now costs $2 per million input tokens and $10 per million output tokens, down from $4 and $20, while Luna falls to $0.10 and $0.50, down from $0.20 and $1.20. The step up in ability is modest. Both models are slightly more capable than the versions they replace, but sit well behind Astra, which is still OpenAI’s most powerful model.

If “per million tokens” made your eyes glaze over, relax. It does that to most people quoting it on LinkedIn too.

And the pace? Opus 5.5 arrives just two months after Opus 5 launched on 24 July, and Sonnet 5.5 and Haiku 5.5 are due “in the coming weeks”. OpenAI only began rolling out GPT-6 Astra on 3 September.

Three top-tier AI launches in three weeks. Any AI plan with a slide titled “our chosen model” is already a historical document.

Discounts are not generosity. They are migration pricing.

Accountants at the ready. And a light dusting of cynicism.

Price cuts at this scale are about moving customers onto new models before a rival does. One detail from developer Simon Willison makes the point neatly: GPT-5.6 has a 25% price rise scheduled for November, which makes GPT-6 half the price of the promotional rates for those older models.

Translation: the old model gets pricier, the new one gets cheaper, and you are gently herded forward. This is not a scandal. It is how software has always worked. But it does mean three things for you:

Your AI costs are not fixed. They move every few weeks, in both directions, often without anyone approving it.
Switching costs are the real price. Every prompt library, process and approval step built around one model is a small tax on moving.
Test scores are marketing too. One review found that no outlet had yet published a full, independently verified comparison of GPT-6 Sol, GPT-6 Luna, Claude Opus 5.5 and DeepSeek V4.1 Flash on the same tests under the same conditions.

Take 1: Price per token is the cost-per-click of AI

We have seen this movie. Marketing spent a decade optimising cost per click, then discovered that cheap clicks were often the most expensive way to not make money.

Token pricing is the same trap. The number that matters is cost per useful outcome.

Artificial Analysis spotted exactly this with Opus 5.5. It costs roughly the same per task as Opus 5, despite producing 1.6 times as many output tokens. At max effort it uses around 119,000 output tokens per task, against about 27,000 for GPT-6 Astra. Cheaper per token. Chattier per job. The maths cuts both ways.

What to measure instead

Cost per approved asset. Not per draft. Per thing that survives your editors and compliance.
Revision cycles per asset. A model that is 50% cheaper but needs three extra rounds is not cheaper.
Time to publish. Speed is where most of the real saving hides.
Error rate reaching review. In regulated sectors, this is the only number your legal team cares about.

If your AI report leads with “tokens saved”, it is a vanity metric wearing a lanyard.

Take 2: Cheap AI makes average content worthless

Here is the uncomfortable economics. When anyone can produce competent content for fractions of a penny, competent content stops being an advantage. It becomes the floor.

And audiences are already suspicious. A Gartner survey found half of US consumers would rather buy from brands that do not use GenAI in their consumer-facing messages, advertising and content. It also found 61% often question whether the information behind their everyday decisions is reliable, and 68% often wonder whether the content they see is real.

Awkward, given the industry spent most of 2025 bolting a chatbot onto anything with a login screen.

So the price war creates a paradox. Production gets cheaper. Trust gets more expensive.

The things a cheaper model cannot copy:

First-hand expertise. Named people who have actually done the thing.
Original evidence. Your data, your outcomes, your case studies.
Point of view. A model can summarise the consensus. It cannot hold your opinion for you.
Accountability. Someone whose name is on it and who will answer for it.

We made this argument in AI content just made your experts priceless. This week’s discounts just made it more true.

Take 3: The model war is not your strategy

Every time a leaderboard changes, someone asks us whether they should switch. We have learned to answer with a cup of tea and a long pause. Occasionally the answer is yes. Usually the problem is not the software.

The data backs this up. Gartner’s 2026 CMO Spend Survey found CMOs put an average of 15.3% of their marketing budgets into AI. Seventy per cent call becoming an AI leader a critical goal for 2026, yet only 30% report mature or fully developed AI readiness. Meanwhile, 56% of CMOs say their marketing team lacks the budget to deliver its 2026 strategy.

BCG found a similar gap. Almost every CMO it surveyed (96%) said AI is driving end-to-end transformation of their function, but only around a third have actually done the work.

Cheaper engines do not help if you have not built the road. Price cuts remove the budget excuse. They do not remove the process problem.

How to avoid getting stuck with one AI provider

Own the process, rent the model. Your briefs, style rules, source libraries and approval steps are the asset. The model is a component.
Match the model to the job. Budget models for tagging and summaries. Top-end models for reasoning, strategy and anything client-facing.
Avoid long lock-ins. Two-month release cycles make three-year commitments look brave.
Test quarterly, switch rarely. Run the same real tasks through new models. Move only when outcomes improve.

This is the thinking behind our AI-powered content systems work, and why we keep saying you need a better content system, not better content.

Take 4: Your next reader is a cheap model

This is the take we think most CMOs are missing.

OpenAI has priced Luna for scale. Luna is aimed at high-volume routine jobs such as summarising and extracting information, while Sol handles recurring coding and automated tasks. AI is also getting good at using software the way a person would. On one such test, GPT-6 Sol almost matched Claude Opus 5’s medium-effort score, 60.5% to 60.3%, at around 80% less cost, according to OpenAI’s own figures.

Read that list again. Summarising. Pulling out key facts. Clicking around software on someone’s behalf. That is exactly what AI search tools and assistants do to your website.

When that work costs a tenth of a dollar per million tokens, more of your audience’s research will be done by AI on their behalf. The first reader of your content is increasingly not a person. It is a model deciding whether you are worth quoting.

Marketers are catching on. Duke’s CMO Survey found 4 in 10 companies now use Generative Engine Optimisation, a capability that did not even feature in earlier editions of the survey.

What makes content easy for AI to use well:

Answer-first structure. The point up top, not after four paragraphs of throat-clearing.
Facts that are easy to lift. Tables, specs, prices, eligibility criteria, dates.
A clear identity. Who you are, what you do, who vouches for you.
Consistent claims everywhere. Contradictions between your site, directories and press confuse machines as much as people.

This is where SEO, GEO and AEO work as one system, not three retainers. And no, SEO is still not dead. It just has more readers now, and some of them are robots.

Take 5: “Sounds like AI” is no longer a quality test

A small detail in Anthropic’s announcement has big implications for content teams. Anthropic says Opus 5.5 leads with the most important information, avoids jargon and odd phrasing, and sticks to the writing rules you set. Earlier Claude models were widely criticised for formulaic, convoluted writing, sometimes nicknamed “Claudish”.

For two years, a lot of editorial quality control has quietly been “does it sound like a robot?”. Hunting for tell-tale phrases. Squinting at suspicious punctuation. (We banned em dashes from our house style years ago. We remain smug about it.)

Rest in peace, the “delve” detector. You served us well.

That test is expiring. When the machine writes cleanly and follows your style guide, the giveaway is no longer how it sounds. It is what is missing:

No evidence you cannot find anywhere else
No named expert
No opinion that could get you into mild trouble
No detail only someone who was in the room would know

The new quality bar is not “does it sound human?”. It is “could only we have written this?”

So what does this mean for you, specifically?

The price war lands differently depending on who you are. Find your row. Skip the others. We won’t be offended.

Sector	The opportunity	The risk	What we would do now
B2B technology	Buyers use AI assistants to shortlist suppliers	Your comparison and pricing pages get summarised without context	Publish honest, structured specs, integrations and proof. Make them easy for AI tools to quote
Charities and NGOs	Cheap models make archive tagging, grant research and impact summaries affordable	AI-written appeals erode donor trust	Automate the back office. Keep donor-facing storytelling human-led
Healthcare and education	Faster evidence summaries for commissioners, procurement teams, referrers and learners	Speed outruns clinical and editorial governance	Treat supplier safety claims as inputs to risk assessment, not substitutes for it
Professional services	Faster research and drafting	Generic advice becomes free	Publish your own frameworks with named experts attached
Publishers	Cheaper production of new formats and spin-offs	More of your content consumed without a visit	Invest in distinctive formats, licensing strategy and brand recall
Platforms and marketplaces	AI assistants can compare and book at scale	Poor product data means you are skipped	Clean feeds, availability, pricing and schema. Treat AI assistants as a channel

A note for regulated and mission-led organisations

Both labs leaned hard on safety this week. Anthropic says that across an automated behavioural audit of nearly 2,000 scenarios, Opus 5.5 outperformed every recent Claude model on almost every measure of misaligned behaviour. OpenAI says its internal rate of deceptive behaviour in coding tasks dropped to 1.3% for Sol and 2.8% for Luna, well below the 10.4% recorded for GPT-5.6 Sol.

Compliance colleagues, you may exhale. Slightly.

Genuinely encouraging. Also, both are self-reported. Your governance still needs to answer: who approves outputs, what gets logged, what never goes near a model, and how you prove it. If you want a starting point, see E-E-A-T in healthcare: why trust signals matter most where getting it wrong costs most.

For charities specifically, we covered the wider shift in your charity’s digital presence was built for 2019.

Hype vs reality: what to stop and start

One for your next board meeting. Screenshot responsibly.

The hype says	The reality is	So stop	So start
“Pick the winning model”	Leaders change every few weeks	Annual model debates	Quarterly testing on real tasks
“AI costs are collapsing”	Per-token prices fell. Per-task costs vary	Reporting tokens saved	Reporting cost per approved asset
“Now we can make 10x the content”	Average content just lost its value	Volume targets	Evidence and expertise targets
“AI writing is finally human”	Sounding human is now the floor	“Does it sound like AI?” reviews	“Could only we have written this?” reviews
“Traffic is what matters”	More research happens through AI assistants	Treating clicks as the only proof	Tracking citations and brand presence in AI answers

What to do before the next model drops (so, roughly Thursday)

Enough opinions. Here is the homework.

The homework, in order

Audit your AI spend by outcome. List every AI tool and model in use. Tie each to a business result, not a usage number.
Separate the process from the model. Document your briefs, style rules, sources and approvals so they work on any model.
Match the model to the job. Budget models for tagging and summaries. Top-end models for reasoning and anything public-facing.
Raise the editorial bar. Every public piece needs one of: original data, a named expert, or a clear point of view.
Make your key pages easy for AI to read and quote. Answer-first, structured, consistent. Start with your top 20 commercial or mission-critical pages.
Start tracking AI visibility. Where are you cited, mentioned or missing in AI answers for your priority questions?
Write down your governance. Especially if you are regulated. What never goes near a model, and who signs off what does.

Quick checklist

Tick each statement that applies to your organisation.

We know our cost per approved asset, not just our token bill Our processes would survive a model switch next month Every public piece carries evidence, expertise or opinion Our top pages answer the question in the first 60 words We track AI citations alongside rankings and sessions Our AI governance is written down and someone owns it

FAQs

The questions we expect in our inbox by Friday.

Which is better for marketing: Claude Opus 5.5 or GPT-6 Sol?

It depends on the task, and the answer will change by Christmas. Opus 5.5 currently leads independent intelligence rankings. Sol and Luna are cheaper for high-volume routine work. Most organisations benefit from using more than one, matched to the job.

Should we switch AI providers because of the price cuts?

Not on price alone. Test your real tasks on the new models and compare cost per approved output, revision cycles and error rates. Switch only if outcomes improve, not just invoices.

Will cheaper AI reduce our marketing costs?

Some production costs, yes. But the savings rarely land where people expect. The bottleneck in most teams is strategy, review and governance, not writing the first draft. Cheaper models make that bottleneck more visible.

Does this change our SEO and GEO strategy?

It raises the stakes. Cheaper models mean more summaries, more AI answers and more research done by AI assistants. Content that is structured, evidenced and consistent is more likely to be quoted. SEO, GEO and AEO should be run as one system.

Is it safe for regulated organisations to use these models?

It can be, with the right controls. Both labs report improved safety results, but these are self-reported. Regulated organisations need their own approval processes, logging and clear rules on what data and decisions never go near a model.

The AI labs will keep racing. That is their job. Yours is to make sure the race works for you, not the other way round.

About the author

Ashley Salek

Ash is Agency Director at Seventh Element, where he leads SEO, GEO and AEO strategy for B2B, charity, healthcare and education clients, including the search rebuild for Practice Plus Group that delivered a 1,012% increase in organic leads. Connect on LinkedIn or read what fuels Ash.