Cheaper AI changes the economics of work.

Executive Summary

The central argument: once an AI system is capable enough for a particular task, reducing the cost of delivering a correct result can matter more commercially than another increase in a general benchmark score. That is an economic threshold for each workflow, not proof that intelligence has stopped improving.

• The price change is real, but uneven. A fixed 4,000-input / 1,000-output-token request cost $0.035 on launch GPT-4o. At current standard rates, it is $0.018 on GPT-6.1 Sol, $0.0009 on GPT-6 Luna, and $0.09 on GPT-6 Astra. Budget GPT-4o mini already cost $0.0012 for this workload in July 2024, so mini-to-Luna is a 25% reduction. Starting from August GPT-4o’s $0.020 gives a 10% reduction to current Sol. These are within broad tiers, not equal-quality indexes. Historical price cut, current prices.
• Similar composite scores can have different costs. Artificial Analysis reports 52 for Sol (max) versus 53 for Astra (max), with benchmark task costs of $0.72 and $3.26. A one-point gap, without a reported uncertainty interval here, does not establish statistical equivalence or equivalence on every task. Sol, Astra.
• U.S. labor makes review time decisive. BLS reports customer-service median pay of $21.53/hour for May 2025. A benefits-adjusted illustration gives $30.76/hour, so two minutes of review costs about $1.03, compared with a $0.018 fixed-token Sol call. The benefits estimate uses an aggregate ratio, not occupation-specific employer costs. BLS wages, BLS compensation.
• Replacement timing is not established. BLS projections and task-exposure research cannot identify a year when all jobs disappear. The occupation tables below distinguish observed evidence, conditional projections and unresolved timing.

Start at GPT-4o: cheaper is a trend, not a rule

GPT-4o launched on May 13, 2024. The August revision cut uncached input pricing from $5 to $2.50 and output from $15 to $10 per million tokens. That is a documented price cut inside the same model family. GPT-4o launch, launch API prices, August revision.

The graphs below separate selected general-purpose releases from budget releases. The paths connect observed milestones; they are not monthly market averages, an exhaustive frontier, or a constant-quality index. October 2026 points are current-price observations, not launch dates. All rates are USD for standard text processing, without caching or batch discounts.

The general-purpose fixed-token example falls from 3.5 cents at GPT-4o’s launch to 1.8 cents on current Sol: 48.6% lower. The budget example falls from 0.12 cents on GPT-4o mini to 0.09 cents on Luna: 25% lower. GPT-5 nano’s 2025 launch rates were lower still at 0.06 cents for this workload. Model selection changes the story; “every new model is cheaper” is not supported. Mini launch, GPT-5 pricing.

A token is a unit of model input or output. Tokenizers differ, and reasoning models may bill many tokens before producing a short answer. A fixed-token calculation answers “what do these rates cost for this volume?” It does not answer “what does a successfully completed job cost?”

Price each selected model, then compare capability

This is a 10-model selection across four providers, checked on October 18, 2026, rather than a claim to list every model in the market. Input, cached-input and output columns are USD per million tokens. “Fixed request” assumes 4,000 uncached input and 1,000 total billed output tokens, including reasoning if applicable. Cache-write fees, tool calls, retries, hosting, taxes and review are excluded.

Pricing conditions change the comparison. The OpenAI rows use short-context rates; requests above 272K input tokens have surcharges, as the model documentation specifies. Haiku’s displayed rates apply through 100K prompt tokens. Argon’s $2/$10 promotion ends November 30, 2026; the published December rates are $4/$20. DeepSeek rows use peak pricing; off-peak rates are half. Its Flash alias now serves version 1017, so the table leaves its benchmark score unjoined. OpenAI, Claude, Gemini, DeepSeek.

“Smart score” here means the Artificial Analysis Intelligence Index v4.3.2, a composite benchmark, at the named reasoning setting. It is not IQ, a percentage of human capability, or the probability a model will do your job correctly. Its English-centric tasks and weights favor particular kinds of work; use your own acceptance tests to choose a model. Reasoning settings are not equal compute budgets across providers. “Max,” “high,” and “max with fallback” are the source’s configuration labels; the fallback policy was not independently reproduced here. We do not treat these configurations as equivalent. Methodology.

The scatter plot includes all nine verified score/cost pairs in this report’s ten-model selection. It uses the same benchmark’s reported cost per task, which includes its token usage and caching assumptions. It is not cost per successful task. Sol’s $0.72 is about 78% below Astra’s $3.26 for a score of 52 versus 53. Opus reaches 58 at $5.98; Luna reaches 38 at $0.07. The useful question is how much the higher score improves your actual workflow. Sol, Astra, Opus, Luna.

Models have also become more capable

OpenAI’s April 2025 comparison reports SWE-bench Verified results of 33.2% for November-2024 GPT-4o and 54.6% for GPT-4.1: a 21.4 percentage-point gain. This is a provider-reported comparison from one publication, with harness and task-selection limitations; it is not a score for launch-day GPT-4o. OpenAI omitted 23 of 500 tasks; counting those as failures gives GPT-4.1 52.1%. Evaluation and caveats.

I have not connected older leaderboard versions into one historical “intelligence” line. Benchmark revisions and retrospectively estimated scores can create a misleading growth curve. The price timeline begins at launch GPT-4o; this coding comparison and the current composite snapshot answer different capability questions.

Compare the economics with U.S. wages

Use wage evidence and employer cost separately. BLS’s May 2025 medians are $21.53/hour for customer-service representatives, $50,670/year for bookkeeping clerks, and $135,980/year for software developers. The latter two become $24.36 and $65.38 per hour when divided by 2,080 paid hours. These are wage equivalents, not billable consulting rates. Customer service, bookkeeping, software developers.

For June 2026, BLS puts average private-industry compensation at $46.89/hour, including $32.82 in wages. Applying that 1.429× aggregate ratio to each occupational median produces an illustrative benefits-adjusted cost, not a BLS occupational estimate. It excludes additional company overhead and does not convert every paid hour into productive output. Employer compensation.

The real comparison is dollars per completed workflow

The calculator models one AI attempt, review, and full human fallback for every detected failure. All workloads, review times, success rates and overhead inputs are editable assumptions; the model rates and wage anchors are sourced. The same task is assumed to reach the human baseline quality after fallback.

AI-assisted cost = model cost + tools + amortized setup + expected error loss + review labor + (1 − success rate) × full manual labor.

The default uses a 10-minute customer-service task, two minutes of review on every attempt, 90% success before fallback, $0.05 of tools and $0.10 of setup allocation. It yields about $1.71 per completed task versus $5.13 manually, a modeled 66.7% reduction. The token call alone is $0.018. Undetected errors can erase the saving; increase expected error loss to reflect that risk.

Cash savings require a practical change in staffing, hours or avoided hiring. Time released for other work is productive capacity, not automatically money removed from payroll. Ten minutes of a developer’s output and ten minutes of support work are not interchangeable.

GPU cost and transistor count are different measurements

A transistor is a physical switch in a chip. A model parameter is a learned numerical value stored in memory. There is no one-transistor-per-parameter rule. NVIDIA discloses 80 billion transistors for H100’s GH100 compute die, 208 billion for a Blackwell GPU package, and 336 billion for Rubin’s package. The newer packages combine two compute dies. Separate HBM memory is outside these counts. H100, Blackwell, Rubin.

The 4.2× H100-to-Rubin transistor ratio does not imply 4.2× intelligence or 4.2× useful throughput. Memory bandwidth, precision, interconnects, workload shape and utilization also matter. H100 predates GPT-4o and appears here as the installed hardware baseline, not as a 2024 chip release.

These are posted single-GPU Lambda instance rates, not chip purchase prices or a U.S. market average. The snapshot lists H100 PCIe at $3.29/hour, H100 SXM at $4.29/hour and B200 at $6.99/hour, with different hardware and instance allocations. Availability is not established by a price page. Lambda prices.

At those rates, B200’s throughput break-even is about 1.63× the useful output of H100 SXM on the same task, quality and latency target to be cheaper per output unit. That is a break-even condition, not a measured performance claim. For example, $4.29 / 100 tasks = $0.0429 per task; $6.99 / 200 tasks = $0.0350. The task throughputs in this example are assumptions.

What does a transistor cost?

Public product announcements do not disclose audited manufacturing cost per working transistor for these AI GPUs. A credible chip-cost estimate would need wafer price, die area, yields, packaging, HBM, testing and allocation of development costs. I have not filled that gap with an invented foundry price.

A limited retail illustration is possible: NVIDIA announced RTX 4090 at $1,599 with 76 billion transistors and RTX 5090 at $1,999 with 92 billion (rounded counts). Dividing gives about $21.04 versus $21.73 per billion transistors. The ratio rises about 3.3%; it does not demonstrate falling manufacturing cost. These historical U.S. launch MSRPs include much more than silicon and are not current shopping quotes. RTX 4090, RTX 5090.

Does adding GPUs make the same model smarter?

Training cost, serving cost and API price are separate bills

DeepSeek’s V3 technical report lists 2.788 million H800 GPU-hours for its official training stages. At its assumed $2 per GPU-hour, that is $5.576 million. The paper explicitly excludes prior research and ablation experiments. This is a disclosed run-cost calculation, not the total cost of building the company or a current frontier-model budget. Training cost table and scope.

Training is paid before a model serves users. Inference is recurring compute; API price is what a provider charges. A provider can spread training costs across many requests, but the unit economics still depend on serving expense, capacity, utilization and demand. Cheap API calls do not reveal a lab’s total training budget or profitability.

Running the same weights with the same inference procedure on more GPUs does not automatically improve the answers. It can make a model fit in memory, reduce latency or serve more people. Small numerical differences are possible, but that is not a general intelligence upgrade.

There are three distinct ways compute matters:

1. Training: more useful compute can support a different, more capable trained model when data, architecture and optimization also improve. Compute-optimal scaling research shows that the allocation between model size and training data matters; GPU count alone is insufficient. Training compute research.
2. Inference: spending more computation on the same question through reasoning, search, candidates and verification can improve some answers. Gains depend on the task and the method; extra generation can also waste money or repeat an error. Inference-time compute research.
3. Serving: batching, caching, quantization, distillation, routing and better kernels can reduce the cost of a target result. Some techniques change model quality and must be evaluated. Higher utilization can matter as much as peak chip speed.

Lower hardware serving cost also does not compel a provider to lower API prices. Retail price reflects competition, capacity scarcity, product strategy and margins. And a falling unit price can coexist with a rising bill if a product runs more tasks, longer contexts or more reasoning.

Your PS4 analogy is useful within a task boundary

A fixed console gives developers a stable performance budget. They can improve software and content without replacing the hardware every year; Sony’s PS4 announcement describes that fixed platform. Sony’s announcement.

The AI equivalent is a task that already clears its quality threshold. If the model can reliably classify a document, then making classification cheaper, faster and easier to integrate may matter more than raising its score on advanced mathematics. There can be many commercial “good enough” thresholds.

The analogy weakens for open-ended agents: context grows, tools fail, goals change and rare mistakes accumulate. There is no evidence that AI as a whole has reached one fixed hardware requirement or universal capability ceiling. Sufficient compute for a bounded workflow is different from sufficient intelligence for every role.

Reliability is the bridge between a demo and a job

The next graph is arithmetic, not measured model performance. If every one of 100 required steps succeeds independently with 99% probability, the chance that all 100 succeed is only 36.6%. At 99.9% per step it is about 90.5%. Real errors are often correlated, and recovery or verification changes the result; the calculation simply shows why a strong single-step demo does not settle end-to-end reliability.

METR’s time horizon measures the human-expert duration of tasks an agent can complete at a stated reliability level, mainly in software domains. A 50% success horizon is not an eight-hour employee substitute. METR explicitly warns that its cleaner tasks and low-context human baseline do not represent all everyday professional work. Definitions and limitations.

The 2025 METR developer trial found a 19% slowdown with early-2025 tools across 16 experienced developers and 246 tasks. That narrow result is not a verdict on current tools. METR’s February 2026 follow-up explains why selection effects complicate newer productivity estimates. Measure complete workflows in the environment where people actually work. Original trial, 2026 update.

From 2026: cost sensitivity, with no assumed probability

For a fixed level of quality on a hypothetical workload, assume annual inference-cost declines of 20%, 50% or 70%. Starting at an index of 100 on October 18, 2026, the 2029 values are 51.2, 12.5 and 2.7; the zero-decline reference stays at 100. These rates are arbitrary sensitivity inputs selected to show compounding, not estimated decline rates, fitted forecasts, probability intervals, or bounds on plausible outcomes. The case labels do not mean that they bracket the historical evidence. A zero-decline path is included as an additional reference. Future premium models can still cost more, and promotional rates can expire.

Historical efficiency gains make further reductions plausible, but do not establish their pace. For context, Stanford’s 2025 AI Index reported a greater-than-280× reduction in GPT-3.5-level inference cost from November 2022 to October 2024. That predates this report’s main window and is not a 2026 forecast. Historical evidence.

My central expectation: routine inference keeps commoditizing, while premium reasoning remains a differentiated service. More spending shifts toward integration, verification and managing exceptions. Total AI infrastructure demand can grow even while a particular quality level becomes cheaper, if usage grows faster than unit costs fall. That is an economic inference, not an observed demand-elasticity estimate.

When will AI replace all jobs? The evidence does not identify a date

A task can be automated while its occupation continues to employ people. We distinguish task automation, workflow autonomy, role replacement, and economy-wide displacement. The evidence does not support a calendar date for eliminating all jobs, or an elimination year for any of the broad groups below.

Readiness milestones: measured acceptance quality, low review burden, reliable recovery, complete role coverage and sustained economics. Passing a milestone in one deployment is not a prediction for a whole occupation.

BLS’s AI exposure categories explicitly do not estimate worker replacement or adoption probability. Its 2025–2035 employment projections are a conditional structural scenario: they assume a full-employment endpoint and technological progress broadly consistent with historical patterns. Net growth is therefore not proof that a faster, more disruptive AI scenario is impossible. Exposure limits, projection assumptions.

Work where cheaper tokens leave major barriers

• Unstructured physical work: field repairs, personal care and changing real-world environments require capable robotics and safe deployment, not just text generation.
• Trusted relationships and accountability: negotiation, leadership, conflict resolution and responsibility for consequences involve acceptance and institutional choices as well as technical competence.
• Ambiguous or high-stakes decisions: unclear goals, missing evidence and expensive mistakes can make verification more costly than generation.
• Whole professional roles: a role bundles routine production with coordination, judgment and exception handling. A system can automate much of its output while still needing a person to own the result.

These are current barriers, not a claim that a category can never be automated. There is no defensible universal replacement date. The ILO’s task-based framework also distinguishes occupational exposure from actual job elimination; its global exposure percentages are not used as U.S. forecasts here. ILO framework.

The practical test in 2026 is to measure accepted outcomes per dollar on representative work, route easy cases to cheaper models, escalate difficult cases, and track review time, undetected errors and recovery. A higher benchmark score earns its price when it improves those outcomes enough.

U.S. occupations: a complete major-group overview, plus trades and professions

No supported replacement year is available for any row below. Unknown does not mean never. This overview covers all 22 civilian BLS major groups and ten selected detailed occupations; it does not claim to examine every one of the 831 detailed occupations individually. Military occupations are excluded.

The tables pair official May 2025 median annual wages with conditional 2025–2035 net employment projections. The numbers are not AI-attributed job-loss estimates. BLS’s overall projection is +3.5%, but its full-employment endpoint and historical-technology assumptions limit what it can say about rapid AI disruption. BLS projections, method and limits.

The task examples and remaining barriers are author interpretation, not official AI-risk classifications. A BLS exposure-table download was unavailable; no missing exposure score has been replaced with a guess.

Specific trades and professions can move in different directions

A programmer is not interchangeable with a software developer in BLS classifications. Growth does not rule out displacement in individual firms, and projected decline is not complete elimination.

Employment effects are already being studied; the evidence does not supply an elimination date

Stanford’s August 2026 revision, using ADP payrolls through June 2026, reports employment of exposed workers aged 22–25 19% below a counterfactual tracking less-exposed peers, largely through reduced hiring. The authors describe this as an observational pattern, not a causal estimate, and find no widespread economy-wide displacement. Pre-existing trends, education controls and sample differences matter. Revised Stanford study.

Anthropic’s March study found no systematic unemployment increase for highly exposed occupations, with tentative evidence of weaker younger-worker hiring. Its product-usage mapping, period and unemployment outcome differ from Stanford’s payroll measure, so the studies do not measure the same thing. Anthropic study.

What can a forward-looking source say about 2030?

Anthropic’s September 2026 economic model presents conditional futures. Its substantial case assumes AI can do half of knowledge work by 2030, with adoption lagging capability; its extreme case assumes much broader autonomous knowledge work and almost no new knowledge tasks for humans. These are assumptions that generate scenarios, not probabilities or predictions. The model excludes hyper-capable robots and cannot date the replacement of physical trades. Scenario framework and limitations.

For electricians and plumbers, hands-on work in varied sites remains a barrier to full automation. Growing data-center and electricity demand can also create trade work. Neither growing employment nor low current language-model exposure guarantees permanent protection from future robotics. Electrician duties, plumber duties.

The defensible forecast is conditional: work with digital inputs, clear acceptance tests and inexpensive recovery can move toward automation sooner. Broader role replacement requires the remaining tasks and responsibilities to pass those tests too. The evidence does not calibrate that sequence to occupation-elimination years.

Sources and reading notes

Evidence cutoff: October 18, 2026. Official provider price pages establish list prices; BLS establishes the cited U.S. wages and compensation averages; Artificial Analysis supplies the dated benchmark snapshot; primary research supports the mechanism and evaluation discussion. Source pages can change. Each chart’s source menu retains its reviewed rows and relevant scope notes.

Historical model rows are launch-price milestones through 2025. October 2026 rows are a current snapshot. They do not form a constant-quality index. Missing benchmark values remain unavailable. No proprietary GPU production-cost data, live GPU throughput tests, employer staffing records, or measured automation success rates were available for this report. No model API calls were used to test capability. Model benchmark results remain third-party measurements, not independently rerun experiments. The launch rates are stated directly in the May 13, 2024 API announcement and corroborated by the August price-cut announcement.

The calculator, reliability graph and cost paths are modeled illustrations. No occupation-elimination dates are estimated. The calculator assumes failures can be detected and completed manually; its error-loss input allows an expected charge for undetected harm but does not prove detection performance. Benefits-adjusted occupational costs apply one aggregate ratio to medians and are labeled as estimates. All dollar values are nominal USD.

The added occupation overview covers all 22 civilian BLS major groups and ten selected detailed occupations, not all 831 detailed occupations. The BLS exposure-classification download was unavailable, so no official AI-risk categories or numeric replacement probabilities are assigned. Group wages exclude benefits and have different coverage from employment counts.