
Applied Digital’s AI Fit-Out Revenue Runs Nearly Three Times Base Rent
Applied Digital reported $183.5 million of tenant fit-out revenue against $65.8 million of base rent in its AI-hosting business.
Read More
Anthropic has released Sonnet 5.5 without cutting Sonnet 5's per-token prices; its lower-cost claim rests on using fewer tokens to finish work.
ZharfAI ResearchSeptember 29, 20267 minutesSource: Anthropic
Anthropic has released Sonnet 5.5 without cutting Sonnet 5's per-token prices. Its lower-cost claim rests on using fewer tokens to finish work, a distinction that matters when setting a budget. CodeRabbit's early tests offer evidence of savings, but the sample is limited and the reported dollar comparison excludes other parts of the review system. AWS availability, Meta's enterprise-platform plans and MongoDB's abrupt leadership change add context to the competition for business customers. Separately, the Dallas Fed reports stronger factory activity alongside price pressure. None of these announcements replaces measuring what an accepted, completed AI-assisted task actually costs.
Anthropic's September 28 release of Claude Sonnet 5.5 leaves the token tariff unchanged from Sonnet 5: $2 per million input tokens, $10 for output and $0.20 for cache reads. The company attributes savings of up to 30% per task to lower token consumption, alongside more than 30% faster output generation. These are Anthropic's claims, not measurements made by ZharfAI. The distinction matters for a purchaser: a tariff reduction applies to qualifying consumption, whereas an efficiency improvement depends on what the model does to complete the assigned work. A budget cannot safely treat the two as interchangeable.
A token measures text processing, not a finished business outcome. Consider an illustrative invoice-extraction system. A shorter response that misreads the supplier still requires correction; another attempt and an employee's time add to the bill. Conversely, fewer visits to the same document can lower expenditure without any change in unit prices if the answer remains correct. ZharfAI's interpretation is that the acceptance criterion must come before the cost comparison. Otherwise an organization can report cheaper responses while quietly paying more to resolve each invoice. Brevity alone is neither improved accuracy nor proof of a cheaper complete process.
CodeRabbit, a commercial code-review provider and early tester, offers a bounded check. On 13 difficult cases, Sonnet 5.5 caught six known issues through actionable comments against Sonnet 5's four, but missed two the older model caught. Across 44 pull requests, mean review time was 6:33 versus 13:31; Claude calls priced at list rates averaged $0.46 versus $1.16. Other models' costs are excluded, and quality judging for that larger set remains unfinished. These observations are useful evidence, not an independent laboratory certification or a complete operating-cost comparison. ZharfAI has not reproduced the tests.
Those results justify an internal experiment, not the removal of a reviewer. Fewer comments may reduce reading time, or they may conceal an important omission. A software team can distinguish those possibilities by selecting its own changes with known outcomes and identifying which missed defects would be expensive. A button-color change and a modification to account-balance calculations do not carry the same risk. Even a lower average review bill can be outweighed by a rare failure in the second category. The useful decision is therefore not simply whether a model finds more problems in aggregate, but where its mistakes occur and who must resolve them.
Effort settings are another source of comparison error. Anthropic specifies Medium as the default in Claude apps and Claude Code, while the developer platform defaults to High. Recording only the model name leaves out a consequential part of the experiment. An internal trial should hold the instructions, available tools and effort budget steady while recording the model version. The team can then examine file rereads, retries and waiting time. Changing several settings at once might improve the service, but it obscures how much of the improvement belongs to the new model. It also makes a later regression harder to explain.
AWS announced availability the same day, with its availability section naming Bedrock's global inference profile and Claude Platform on AWS in North America. That is more actionable than an unspecified promise that a model will come to the cloud. It still requires a buyer to check permitted processing locations, account access and billing arrangements. A region in an example API call does not itself establish that all processing stays in that region. Distribution creates an option to evaluate; it does not authorize sending every category of company document through whichever endpoint happens to be available.
For a Persian-speaking team, a practical trial could draw on accounting-software changes involving Solar Hijri dates, rial-versus-toman amounts and short, ambiguous descriptions. This is ZharfAI's proposed example, not a demonstrated Sonnet application. Evaluate results before revealing the model name to reduce the influence of expectations about a new release. Record execution charges, human handling time and rejected results separately. An easy-only sample says little about difficult cases; a sample consisting solely of exceptional failures misrepresents ordinary operating cost. The test should make both limitations visible instead of compressing them into one attractive average.
Meta's September 28 Enterprise Platform announcement addresses the same business audience from a different position. It names CJ Desai to lead the initiative and identifies an initial portfolio including the Muse agent, Meta Business Agent, Muse API and Muse Code. The announcement does not establish pricing or general availability for a complete enterprise package. ZharfAI sees a prospective competitor worth tracking, but not yet a substitute for a contract whose terms a buyer can inspect today. Support, responsibility for errors and the ability to leave a service belong in the comparison alongside model capability and the supplier's name.
MongoDB confirms Desai's immediate departure and Dev Ittycheria's appointment as interim chief executive, while reaffirming its existing third-quarter and full-year fiscal 2027 guidance. That is not a new earnings result or a guarantee of delivery. A search for a permanent chief executive has begun. The company's September 29 Investor Day is scheduled for 11 a.m. New York time and had not occurred at this edition's cutoff. The transition is a direct organizational consequence of Meta's plans. By itself, it establishes neither weakening database demand nor a Meta victory in enterprise sales; those require different evidence.
The Dallas Fed's September 28 manufacturing release provides a separate economic signal. Texas's production index reached 29.5 while the company-outlook index declined from 19.2 to 8.7; the raw-materials-price index was 52.2. These are balances of responses reporting increases and decreases, not percentage rates of output growth or inflation. The survey collected 63 responses during September 15–23. Stronger current activity can therefore coexist with concern about costs. This regional survey is neither an assessment of the entire US economy nor evidence of Sonnet's economic effects. A software efficiency claim cannot be used to explain these factory readings.
The next evidence to watch is concrete: whether judging the larger CodeRabbit set confirms quality alongside speed; whether customers record lower complete-task costs after retries and human review; what purchasable products and commitments Meta supplies; and how MongoDB's interim leadership explains continuity. ZharfAI's conclusion is narrower than a general claim that AI has become cheaper. An unchanged tariff is compatible with savings, but the saving should be counted after an acceptable result is delivered. Lower token consumption begins the financial question. It does not finish it, and it does not erase the cost of getting an important task wrong.
TagsClaude SonnetAI pricingCode reviewCloud deploymentEnterprise softwareManufacturing

Applied Digital reported $183.5 million of tenant fit-out revenue against $65.8 million of base rent in its AI-hosting business.
Read More
Anthropic launched Claude Opus 5.5 on September 22 with standard input and output rates 20% below Opus 5.
Read Morellama.cpp release b10687 changes how two Qualcomm Adreno GPU generations route matrix multiplication.
Read MoreIf this note maps to a real system in your organisation, start with the services page or a shipped case study.