AI Models on a Budget: Frontier Gets Only the Hard Cases

In early November, Amazon and Revolut swapped models and Apple trimmed AI budgets. Few cut frontier models off; most shrank what those models are asked to do.

Revolut gave AI coding tools to roughly 2,000 engineers, watched the bills arrive, and then changed the default. According to The Register's report on November 3, the British fintech now sends most of that coding work to GLM-5.3, an open-weight model from China's Z.ai that it runs itself on 56 Nvidia B200 GPUs. Revolut says the switch cut its AI coding inference costs by roughly 90 percent. It is one of the clearest cases so far of a company moving real work between AI models because of price.
Revolut was not alone that week. In early November several companies published numbers pointing the same way: routine, high-volume work is being handed to smaller or open-weight models, and the reason given is the token bill. The shift is real, but it is narrower than "companies are leaving frontier models." What they are doing is shrinking the frontier model's job to the hard and ambiguous cases. Meanwhile, revenue at the top labs keeps climbing.
The workloads that moved first
The Revolut story is about a default, not a divorce. The Register reports that self-hosting GLM 5.1 cost about a tenth of using OpenAI's Codex or Anthropic's Claude Code, which is what started the move. Claude Code and Codex are still in use. GLM is text-only, so a model router sends vision queries out to models like Gemini or GPT. The frontier models became the exception handlers. And the 90 percent figure is Revolut's own.
Amazon's case is smaller and cleaner. Two engineers wrote on the Strands Agents blog that they replaced Claude Sonnet 4.6, used to label tickets in a large internal ticketing system, with a decision model called Strands Decider. The job was closed-choice: each ticket needed to fit within 20 category types. The decision model came out about 44 times cheaper, a 98 percent reduction, and on a separate set of tickets it beat the production Sonnet setup. The authors expect the change to save millions of dollars in token costs a year.
Hightouch, a data company, showed what the split looks like when you keep the frontier model in the loop. In a November 6 post on grading identity resolution (deciding whether a merged profile actually describes more than one person), a small decision model named Jev looked at every case first and passed only the uncertain ones to Claude Opus 5.5. Cost fell to about a quarter of running Opus alone, and accuracy stayed close to Opus on its own.
The pattern is in the shape of the work. These are jobs with a small answer space and enormous volume, where a per-token price multiplied by millions of calls turns into a line item that finance starts asking about.
AI models priced like a commodity
Companies that have not switched models are cutting how much they use. TechCrunch, citing Bloomberg, reported that Apple lowered the daily AI spending cap for its engineering teams in October from $300 to $200, with some allowances set to fall to $100 a day. Outside engineering, a limit as low as $25 a day is reportedly on the table. TechCrunch framed it as the end of "tokenmaxxing": the push to use as much AI as possible lasted until the bills came due.
The market-wide data says the cheap tier is already doing most of the lifting. In a November 2 piece for finance leaders, OnlyCFO pulled numbers from OpenRouter, a service that resells access to many models through one API. Since January 2025, the average price paid per million tokens there fell from $0.77 to $0.14. Open-weight models now carry about 60 percent of weekly tokens on the platform but only 26 percent of spend. Most of the work is already flowing to cheap models; most of the money is not.
The pressure reaches suppliers too. TechCrunch reported on November 2 that Elon Musk deliberately tightened Grok Bot usage limits, writing on X that the cost of serving it was "unsustainably high relative to revenue."
Meta Muse Ember and Gemini sell the cheap tier themselves
The more telling signal is that frontier labs are packaging the split as a product. On November 9, Meta released Muse Ember, a lightweight model for the work agents repeat all day, such as tool calls and routing. The design pairs it with the larger Muse Wonder, which plans while Ember executes at volume. At 10 cents per million input tokens, Meta pitches it as well below Anthropic's Haiku and OpenAI's Luna, and price-competitive with open models such as DeepSeek V4.1 Flash and GLM 5.3 Flash.
Google went the same direction twice in two days. On November 11 it launched the Gemini Decisions API, which scores up to 1,000 closed-set questions in a single pass instead of generating text, at $0.10 per million input tokens and nothing for output. On November 12 it added Orchestrator to its Antigravity coding tool, an option that lets Gemini 4 Carbon plan and repair while the lighter Hydrogen makes edits and runs tests. Google says the combination reached near-Carbon success rates while using 24 percent less quota on average. After a staged Gemini 4 rollout, one of the first follow-ups is a tool for using the flagship less.
IBM's Arvind Krishna made the buyer-side version of the argument on the company's October earnings call: most enterprise tasks, like summarizing email or reviewing contracts, don't need the top model, and open-weight models are now good and cheap enough to run on-prem. It is worth remembering that IBM sells the hardware that on-prem AI runs on.
Read together, these launches suggest the labs expect the bulk of agent traffic to move to a cheap tier and would rather own that tier than lose it. That is an inference from their product choices, not something any of them has said.
What doesn't fit: frontier revenue is still climbing
The strongest counterevidence is the money. Bloomberg reported that OpenAI's quarterly revenue crossed $10 billion for the first time, and Anthropic's amended S-1 filing on November 2 showed $17 billion in third-quarter revenue. Y Combinator's Garry Tan says he spends up to $200,000 a year on tokens. He also called it too expensive right now, but he was not describing a plan to use less.
The OpenRouter numbers cut both ways as well. If open-weight models carry 60 percent of tokens but 26 percent of spend, about three-quarters of the money still goes to everything else. Total AI usage is growing fast enough that individual workloads can leave the frontier while frontier revenue rises anyway.
Moving also brings costs that don't show up in a per-token comparison. In late October, Cloudflare disclosed that GLM-5.3 on Workers AI had leaked data from other customers' requests into 155 responses. That is the same model family Revolut runs, and the cause was a serving misconfiguration combined with a bug in an open-source inference engine, not the model itself. Self-hosters own that risk. Around the same time, Cohere claimed Alibaba's Qwen model tried to send a file to a Beijing server during a security test. Most researchers read it as a training artifact rather than espionage, but it gives any company running Chinese open weights one more thing to explain to its security team.
Frontier prices are also falling. Google cut Gemini Omni 1.2 Flash API pricing by 30 percent, and every cut like that shrinks the gap that justifies a migration. Even the Amazon engineers drew limits around their result: it fits high-volume, closed-choice labeling, and not tasks that need explanations or a frequently changing catalog.
What would prove this right or wrong
The first number to watch is open-weight share of spend on OpenRouter. Token share is already high; if the 26 percent spend share starts rising, the money is following the work.
The second is the growth rate at OpenAI and Anthropic over the next quarter. Revenue can keep rising while growth slows. A sharp slowdown would suggest the split is eating into frontier demand. Steady growth would mean this is cheap work leaving a pie that is still expanding.
The third is whether more companies follow Revolut in changing the default for a broad workload like coding. Most of the published cases so far, Amazon and Hightouch included, carve out a single narrow task. If broad defaults start flipping, the word "migration" will be earned.
Sources
- British fintech Revolut slashes AI coding costs 10x by cutting out Claude and going self-hosted (The Register)
- Amazon saves millions replacing Sonnet with Strands Decider (Strands Agents)
- LLM judges for identity resolution (Hightouch)
- Apple is reportedly cutting budgets for its internal AI usage (TechCrunch)
- AI tokenomics (OnlyCFO)
- Meta releases Muse Ember, a new lightweight model for agent execution (TechCrunch)
Comments
Post a Comment
Please read our comment policy before commenting.