Giving Anthropic’s Claude Opus 5 agent one shared in-context history of the tasks it had already done in the run cost $914.36 for a 398-task run, against $161.77 with no memory at all, according to the cost table in a FAIR at Meta paper posted on 31 October. That is about 5.7 times the bill, by our arithmetic, and the shared memory scored lower: a normalized reward of 0.169 against 0.178 with no memory.
Across the four models tested, the single shared history cost between about 5.1 and 5.7 times the no-memory set-up, by our arithmetic from the paper’s Table 6.
Short explainer · Shared Agent Memory Cost Opus 5 5.7x
A vertical news Short summarising this story. Figures come from the FAIR at Meta paper on arXiv; our arithmetic is labelled.
Business Pill · The Memory Wall
A short explainer on memory as a constraint in AI systems. It teaches the general idea only and says nothing about this paper.
The key insight: By our arithmetic from the FAIR at Meta paper’s Table 6, one shared agent history cost Opus 5 $914.36 per 398-task run against $161.77 with no memory, about 5.7 times, while its normalized reward was 0.169 against 0.178.
Our first report on the paper covered its main finding: routing an agent’s past runs into separate memory clusters, a method the authors call Routed Continual Learning, cut estimated API cost by 65–72% against one shared history. The appendix puts dollar figures on every set-up.
The paper estimates API cost for every model call “using a frozen, dated public list-price schedule that accounts for token caching”. Its Table 6 reports the mean cost in US dollars per 398-instance rollout across five matched seeds.
For Opus 5, no memory cost $161.77, routed memory $316.61 and one shared history $914.36. Per task, that is about $0.41, $0.80 and $2.30, by our arithmetic.
For OpenAI’s GPT-5.6 Sol the same three set-ups cost $89.24, $125.99 and $456.22. For Claude Opus 4.7 they cost $128.86, $222.72 and $708.78, and for Google’s Gemini 3.7 Flash $39.59, $61.08 and $200.14.
Mean estimated API cost per 398-task run for three memory set-ups on four models, from Table 6 of the FAIR at Meta paper (arXiv 2610.38796, 31 October 2026). Costs use a frozen public list-price schedule including caching; multiples against no memory are our arithmetic. Chart by FourWeekMBA.
What the Extra Spend Bought
On Opus 5, the shared history did worse than no memory: 0.169 against 0.178 in normalized reward, the paper’s Table 1 shows. On the other three models it did slightly better than no memory, and its mean stateless-reference gain was 0.015, against 0.280 for the routed method. The paper’s heading for that result: “Global ICL loses its advantage under heterogeneous interleaving”.
The routed method cost more than no memory on every model, between about 1.4 and 2.0 times, by our arithmetic, but it also scored highest of the four staged set-ups for every model, according to the paper. For Opus 5 it reached 0.401.
Where the Tokens Went
The token table shows how many more tokens the shared history used. On Opus 5 it wrote 117.0 million tokens to the cache per run, against 7.4 million with no memory, about 15.8 times as many, by our arithmetic. Its total token use was 353.5 million against 123.8 million.
Each history had a 100,000-token context budget, “applied stream-wide for Global ICL and independently per cluster for the routed conditions”, the paper says. The accounting prices uncached input, cache reads, cache writes and output separately.
The Business Lever: Cost per Task
As we read the tables, the figure that matters for anyone pricing an agent product is cost per task, and in these runs the memory design moved it sharply: on Opus 5, the gap between the shared history and the routed method was about $1.50 a task, by our arithmetic.
Summed across the four models, one shared history cost $2,279.50 in these runs and the routed method $726.40, by our arithmetic. These are the paper’s list-price estimates for one benchmark stream, not any company’s actual bill.
The Structural Read
As we read the tables, the headline 65–72% saving is measured against the most expensive set-up: one shared history cost about 5.1 to 5.7 times no memory on all four models, by our arithmetic.
As we read the token table, cache writes grew the most in relative terms on Opus 5: 117.0 million tokens per run against 7.4 million with no memory, about 15.8 times, by our arithmetic, while cache reads rose from 109.7 million to 231.1 million.
As we read the paper, the routed method was the only staged set-up that scored highest on every model, and it still cost about 1.4 to 2.0 times no memory, by our arithmetic.
“This policy captures the caching advantage of long ICL histories.”
Three Implications
FOR AGENT BUILDERS In these runs, Opus 5 cost about $0.41 a task with no memory, $0.80 with routed memory and $2.30 with one shared history, by our arithmetic.
FOR MODEL CHOICE The paper reports the same ordering of costs on Gemini 3.7 Flash, GPT-5.6 Sol, Opus 4.7 and Opus 5.
FOR READING THE RESULT The costs are list-price estimates on one 398-task stream, run sequentially; the paper does not evaluate simultaneous updates.
The framework puts it plainly: “The agentic surface — the harness — is becoming where value pools in the entire AI stack.”
As we read it, memory design is part of that harness, and in the FAIR at Meta tables it moved cost per task several-fold on the same model. The framework is the owner’s argument and does not discuss this paper.
The costs are estimates from a frozen, dated public list-price schedule, not invoices, and they cover one 398-task stream that mixes CL-Bench tasks with banking instances. The paper says its runs were sequential and that it does not evaluate simultaneous updates or stale reads, which “remain important tests for distributed deployment”. Per-task figures, multiples and sums are our arithmetic.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
In the FAIR at Meta paper’s own cost table, one shared in-context history cost Opus 5 $914.36 per 398-task run, about 5.7 times the no-memory run, while scoring 0.169 against 0.178, and across four models the shared history cost about 5.1 to 5.7 times no memory, by our arithmetic.
A note on sourcing. Every dollar, reward and token figure comes from the FAIR at Meta paper (arXiv 2610.38796, 31 October 2026), whose costs are estimates from a frozen public list-price schedule. Multiples, per-task figures and sums are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
