DeepSeek Cut Prices 75%. Meta Undercut Everyone. Your AI Bill Still Went Up.

Three pricing stories landed in the same week. Meta opened the Muse Spark 1.1 API at $1.25 per million input tokens and $4.25 per million output, undercutting Anthropic and OpenAI's $25-$50 per million output rates by 5 to 12x. DeepSeek made its 75% price cut on V4-Pro permanent, pushing output under $1 per million tokens. And Palo Alto Networks CEO Nikesh Arora went on CNBC to say none of it is enough: he wants token efficiency to improve 20% in the next twelve months and 90% the year after. Read those three stories back to back and you'd expect enterprise AI bills to be in free fall. They are not, and the reason is worth sitting with before you pick a model off a price sheet.

This week's numbers

• Meta Muse Spark 1.1 API: $1.25/M input, $4.25/M output, $0.15/M cached input, launched July 9 in US public preview
• Anthropic and OpenAI flagship output pricing: roughly $25-$50 per million tokens, a 5-12x gap versus Meta's rate card
• DeepSeek V4-Pro: 75% price cut made permanent in May, output now under $1 per million tokens
• Sam Altman, to CNBC on GPT-5.6: the model is "54% more token-efficient" on agentic coding
• Nikesh Arora, Palo Alto Networks CEO: his company already spends about $1 million a day on AI tokens, a figure he expects to hit $2-3 million a day with broader adoption

That last number is the tell. Arora runs a company watching its own AI spend, in real time, while every vendor around him cuts sticker prices. If price cuts translated cleanly into lower bills, his forecast would be going down, not up.

The 100x problem

VentureBeat's framing of the DeepSeek cut nailed the mechanism: a 75% price cut does nothing for a company whose agents are running hundreds of times more tokens per user request than its pricing model assumed. Chatbots answer once. Agentic coding workflows plan, call tools, read back results, retry, and plan again, and every one of those steps bills as tokens. Cut the price per token by three quarters and an agent that now takes four times as many turns to finish the same task has erased the discount entirely.

I think 54% is a good start, but it is not yet enough. — Nikesh Arora, on Sam Altman's claimed token-efficiency gain for GPT-5.6 (CNBC, July 9, 2026)

Arora's response to Altman is the whole argument in one line. A vendor announces an efficiency win, a customer running that model at scale says the win doesn't show up on the invoice. Both can be telling the truth. Efficiency per token and total tokens consumed are different curves, and right now the second one is winning.

What a 5-12x discount actually buys

Meta's Muse Spark 1.1 rate card is real and it is aggressive: on paper, output tokens cost a fifth to a twelfth of what Anthropic or OpenAI charge for a comparable flagship. That is a genuine advantage if, and only if, Muse Spark reaches a correct answer in a similar number of tokens as the models it's undercutting. If it needs more back-and-forth, more retries, or a longer chain of thought to land the same result, the effective cost per finished task can close that gap fast, or flip it. Sticker price is a rate. What you actually pay is rate times volume, and volume is the part no press release quotes.

The number to actually watch

We've made this point before from a different direction: in our test of the same model at low versus max thinking effort, more output tokens did not track with better answers, it just tracked with a bigger bill. That's the same mechanic Arora is describing at enterprise scale, just visible one prompt at a time instead of one invoice at a time. Effort level, not sticker price, is the dial that actually moves what you pay, and it's a dial every model exposes differently.

This is why we run the arena on live output rather than a rate card. Every model in the coding arena answers the same one-shot prompt, and you can see the actual token count it took to get there before you ever look at what the vendor charges per million. We've also run the direct cost comparison for Claude against DeepSeek at $50 per million tokens, and the same rule held: the headline rate told you almost nothing about what the winning answer actually cost to produce.

None of this means price cuts are meaningless. DeepSeek's permanent 75% cut and Meta's undercut of Anthropic and OpenAI are real, and over enough volume they matter. But treating the per-token rate as the cost of doing business is how a CEO ends up watching his AI bill triple while every vendor he buys from announces a discount. Watch tokens per finished task, not dollars per million advertised. If you want the same beat in Polish, nowosci.ai is covering this price war from the European enterprise-adoption side.