Fable 5 Just Dropped. And yes, the CHR went up again.
In April, OpenAI doubled frontier per-token pricing. Yesterday, Anthropic did the same. Here is what two doublings in seven weeks do to the Compute Heat Rate.
Anthropic released Claude Fable 5 yesterday. It is the first generally available model in the company’s new Mythos class, a capability tier that sits above Opus. The API price: $10 per million input tokens and $50 per million output tokens. Exactly double Claude Opus 4.8, the most capable model anyone could buy from Anthropic the day before.
In late April, OpenAI released GPT-5.5 at twice the per-token price of the generation before it. I ran the CHR economics then and called it a new ceiling for frontier inference.
So let’s do it again and see what it looks like, shall we?
The Compute Heat Rate measures the maximum electricity price an AI workload can sustain before the operator would rationally curtail. The formula is straightforward: take the revenue generated per MWh of electricity consumed, subtract non-electricity operating costs, and apply a margin requirement. What falls out is the price ceiling for that demand class.
For more on the details of the Compute Heat Rate research, formula, and pricing assumptions, visit www.computeheatrate.com.
So with the same reference configuration as the GPT-5.5 calculation, so the numbers are apples to apples: a current-generation NVIDIA B200 GPU running at roughly 2,500 output tokens per second (conservative, 50% utilization), drawing about 1 kilowatt, with a facility-level PUE of 1.3. At $50 per million output tokens, that GPU generates about $450 per hour in revenue. Subtract non-electricity costs (GPU amortization, cooling, networking, maintenance), apply the standard 30% margin requirement, and the Compute Heat Rate for Fable 5 inference comes out to approximately… drumroll….
$263,000 per MWh.
That is roughly 5,300 times the gas heat rate that has anchored wholesale electricity pricing for decades. It slots above GPT-5.5 standard ($156,000/MWh) and below GPT-5.5 Pro ($955,000/MWh). For context, the blended CHR across all AI workload types published in the Q2 CHR index is roughly $8,000/MWh, about 160 times the gas heat rate. The frontier tier has always been the high end but the high end keeps moving up.
A note on hardware: nobody outside Anthropic knows exactly what silicon serves Fable 5, and for this exercise it does not matter. The CHR holds the reference rig constant and lets the pricing do the talking. The pricing says each megawatt-hour of frontier inference is now worth two-thirds more than it was in April.
Two more signals in this release that energy people should be reading.
First, Anthropic is rationing. The launch post says demand for Fable 5 will be “very high, and difficult to predict.” Subscription plans get the model included only through June 22. After that it gets pulled and billed through usage credits, to be restored as a standard inclusion once capacity permits.
So the $10 and $50 pricing may not be the ceiling, it might be the floor. Every megawatt-hour serving Fable 5 today could carry more willingness-to-pay than the price sheet shows.
Second, the price sheet itself discriminates along the two axes electricity markets care about most: where and how fast. US-only inference routing carries a 1.1x multiplier on Anthropic’s API; a customer who needs tokens generated on American infrastructure pays $55 per million output instead of $50. Fast mode, the premium-speed option on Opus models, prices output at up to six times the standard rate.
In April I called GPT-5.5 the latest proof point for the Efficiency Trap: per-token costs collapse across generations, capability premiums reset the top of the pricing curve with every release, and total consumption grows faster than efficiency improves. Jevons’ paradox on a six-week release cycle.
Fable 5 is yet another proof point, and it adds a mechanism. Anthropic’s launch materials lean hard on token efficiency: one early tester reported Fable 5 getting in 36 hours nearly to where GPT-5.5 landed after four days, on a third of the reasoning tokens. More capability per token is precisely what justifies more dollars per token. The premium is the market pricing intelligence per electron.
The Compute Heat Rate is compounding and the value of electrons just doubled again.
Hans Royal is the originator of the Compute Heat Rate (CHR) framework. The full methodology is published as Royal (2026), The Compute Heat Rate: Quantifying AI-Driven Electricity Price Tolerance and Its Implications for Wholesale Market Repricing. Available at SSRN: http://dx.doi.org/10.2139/ssrn.6322318 and at computeheatrate.com.
Disclosure: Claude models (among others, including GPT and Gemini) run inside my research tools, and research for this piece was supported by the model it analyzes. I have no financial relationship with Anthropic beyond standard API fees, and Anthropic had no involvement in this analysis. The math is reproducible from the published price sheets.