Singapore’s public transport fares rise 7% from 26 December, driven by energy prices. In the same week the 30-year US Treasury yield passed 5.61%, its highest since 2002, on the same input. Meanwhile the cost of AI work per task fell again. Rising discount rates and falling unit prices pull an automation business case in opposite directions.
On 29 September, Singapore’s Public Transport Council granted a 7% fare adjustment. Adult card fares rise by 12 to 13 cents per journey from 26 December — a record increase in cash terms.[1]
The council attributed the formula’s output to a “substantial increase” in energy prices between July 2025 and June 2026, during the Middle East conflict.[1]
A day later, the 30-year US Treasury yield surpassed 5.61%, a level last seen in 2002. It rose for a sixth straight session, and the coverage named inflationary angst and heavy corporate-debt supply.[2]
The mechanism was stated plainly: government debt has been selling off around the world as elevated oil prices, tied to the war in the Middle East, ripple through the global economy. Investors now expect central banks, including the US Federal Reserve, to raise rates further.[2]
Singapore’s manufacturing survey tells the same story from the factory floor. Rising input costs and prolonged supplier lead times are intensifying operational pressures, according to SIPMM, while petrochemical firms face feedstock shortages from the closure of the Strait of Hormuz.[3]
Three unrelated headlines, one input. That matters because the three prices land in different places: a commuter’s card, a factory’s cost base, and the discount rate applied to every capital project a finance team approves.
The 30-year rate surpassed 5.61%, returning to territory that had long been the norm before the low-rate era spanning the global financial crisis and the pandemic.[2]
The 10-year yield reached 5.25%, near its highest since 2007, and the two-year settled around 4.89%. Treasuries have lost 2.6% in 2026, against a 6.3% gain in 2025.[2]
Two named causes are structural rather than cyclical. One is supply: Paramount Skydance launched the largest portion of a US$52 billion debt package, raising about US$32 billion in what a rates strategist called the fifth-largest investment-grade deal on record.[2] The other is a buyer’s strike — Citigroup’s description of a Treasury market where the usual long-end buyers have not appeared.[2]
For a finance team this is not a bond-market story. The risk-free rate is the denominator of every payback calculation. An automation programme that returns a fixed amount each year earns a thinner premium over that rate today than it did in the low-rate era, for identical execution risk.
Market-implied pricing points the same way: at least one quarter-point increase by the end of 2026, and close to three more by mid-2027.[2] A case built on the assumption that funding costs normalise downward is a case built on a forecast, not on the curve.
Singapore’s manufacturing PMI edged up 0.2 point from August to 51.7 in September, a 14th consecutive month of expansion, SIPMM reported.[3]
The electronics sector gained 0.3 point to 52.9, a 16th straight month of growth. DBS senior economist Chua Han Teng noted that the headline and electronics readings were their highest since October 2018 and January 2018 respectively.[3]
The demand is AI-related and it is regional. Chua attributed the strength to AI-linked demand running through South Korea and Taiwan, the upstream beneficiaries, with Singapore benefiting through trade linkages.[3]
Inside the same release sits the cost pressure that the fare formula and the bond market are also registering. OCBC chief economist Selena Ling attributed falling finished-goods indices to supply-side disruptions and higher import costs rather than demand constraints, and SIPMM’s executive director described manufacturers meeting demand while managing persistent supply constraints.[3]
The growth is real, narrow and cost-sensitive. It rests on AI-related demand in one sector, and that sector’s own inputs are getting more expensive at the same time.
On 28 September, Anthropic released Claude Sonnet 5.5 at the same per-token prices as its predecessor: $2 per million input tokens and $10 per million output tokens.[4]
The saving comes from fewer tokens rather than a lower price. The company reports up to 30% less cost per task and output more than 30% faster, because the model needs fewer tokens to do the same work.[4]
A customer benchmark makes that concrete. Balyasny Asset Management ran a private suite of 2,441 finance tasks and reported about 121,000 tokens per answer, where the previous model used 497,000.[4]
On knowledge work the substitution is nearly complete. Sonnet 5.5 scores 1,844 on GDPval-AA against Opus 5.5’s 1,846, at $2 and $10 per million tokens rather than $4 and $20.[4]
Source: Anthropic — “Claude Sonnet 5.5” (28 Sep 2026). anthropic.com
Anthropic’s own cost curves carry the caveat. As effort rises, cost per task rises with it, and at higher settings the cheaper model can perform comparably at a similar cost.[4]
So the deflation is conditional on how work is scoped, not on the sticker price. That is the numerator of an automation business case falling while the denominator rises.
Four adjustments follow, and none of them requires a view on where rates or model prices go next.
1. Re-run payback at the current rate. A case approved when funding was cheaper was approved against a lower denominator. Recompute it before renewing a platform contract or funding another phase of a build.
2. Discount cost per task, not price per token. The saving depends on scope and effort. A model that removes three quarters of the tokens on one workload can cost more on another, so the benchmark has to be the reader’s own tasks.[4]
3. Substitute before substitution stops working. A mid-tier model within two points of the top tier on knowledge work removes the premium from a large share of ordinary tasks. Reserve the expensive model for work that needs sustained judgment.
4. Stage the build. When unit prices fall and the duration premium rises, a smaller programme that can be re-scoped beats a large commitment priced today. Waiting costs less on the numerator and more on the clock.
None of this argues against automation. It argues against approving it on arithmetic set in a lower-rate year.
The fare rise is a formula output, not a policy decision. The formula generated a 5.3% adjustment for 2026; the council granted 7%; and 7.7 percentage points were deferred to future reviews.[1]
The offset is real. Close to S$200 million in additional subsidy for 2027 will cover the deferred adjustment, on top of more than S$2 billion in annual operating subsidies. Vouchers rise to S$80 per eligible household from S$60, with the income ceiling raised to S$2,100 per person per month.[1]
Energy is the shared driver, and energy mean-reverts. The fare formula looked at a window shaped by one regional conflict, and the same conflict is what pushed oil prices through the global economy.[1][2]
Long yields can rise for reasons unrelated to automation. Heavy corporate-debt supply and an unwind of the yen-funded carry trade are both named as contributors to the sell-off, and neither says anything about the return on a document pipeline.[2]
If energy prices retreat, the fare formula, manufacturers’ input costs and the long end ease together, and the collision disappears. That is the cleanest test available.
1. The 30-year US Treasury yield. Whether it holds above 5.61% once the corporate issuance pipeline clears tells you whether the denominator has settled at a new level.[2]
2. The next fare review. Watch whether the deferred 7.7 percentage points return, and whether energy still dominates the formula’s output.[1]
3. Singapore’s electronics PMI. Whether the AI-linked leg broadens beyond electronics, or whether supply constraints keep the expansion concentrated.[3]
4. Cost per task, not price per token. The next model release matters only if it lowers what a finished task costs at the effort setting a finance team would actually run.[4]
One input raised Singapore’s fares, its factories’ input costs and the long-end yield in the same week, while the price of automated work kept falling. The project that survives both moves is the one whose payback was computed at today’s rate.
This analysis is based on publicly available data as of 2026-10-04. For related coverage, see When the Treasury Becomes a Market Maker and The AI Bust Warning.