Architecture Vision
Which AI investments make sense today
LLMs are changing faster than enterprise roadmaps. The durable investment is the capability to delegate agentic workloads safely, evaluate AI output reliably, and route each workload to the right model.
From assistance to delegated workloads
Until 2025, LLMs mainly assisted with individual steps. Since autumn 2025, strong models have been able to complete clearly described software tasks over several hours, including tool calls, tests, and corrections. An independent benchmark shows the scale of the change: GPT-4 handled tasks measured in minutes, while Claude Opus 4.5 handled tasks lasting more than five hours by the end of 2025.
The same way of working is now reaching research and other office workflows. For enterprises, the specific AI solution matters less than the operating boundary: which agentic workloads may be delegated, which systems and data may agents access, and which criteria determine whether their work is correct for the business?[1]
Frontier capability no longer belongs to frontier labs alone
OpenAI released GPT-5.4 on 5 March 2026. Just over five months later, DeepSeek V4 Pro 0813 reached the same Intelligence Index score of 53. The estimated cost of the compared agentic workload was 1.10 US dollars for GPT-5.4 and 0.25 dollars for DeepSeek. A capability level reached only by a proprietary frontier LLM in March was available as open weights within one planning cycle.
A tie on the index does not, however, make the models interchangeable. This is because the index aggregates nine evaluation areas; a model can be optimised for known benchmarks and still behave differently on proprietary code, German contracts, or domain-sensitive decisions. The comparison is therefore a market signal, not a selection decision. Selection requires enterprise evaluations and staff who understand the business process well enough to interpret real tests.[2]
Intelligence per dollar is now measurable
Across six benchmarks, the same measured level of model performance became 9 to 900 times less expensive within a year depending on the task. That range shows that there is no universal rate at which AI capability becomes cheaper. A cost comparison applies only to a specific workload, a defined quality level, and a stated date.
Enterprises should therefore compare the total cost to business acceptance, not token prices: in software, until a change has passed tests and review; in knowledge work, until a subject-matter expert has reviewed and adopted the result. DeepSWE provides measured success, time, and cost for software tasks under one agent setup. Formula v1 divides attempt time and cost by first-run success, making the cost of unreliable output visible. Its weighting is an editorial starting point, not a business-value model. Enterprises must repeat the calculation with their own tasks, acceptance rules, review effort, and failure costs.[3][4][5]
Page snapshot9 October 2026113 tasks
Bottom left: faster and cheaper. Diamonds sit on the Pareto frontier.
2 off-scale outliers remain available in the full table.
Models28/28
Leading LLMGPT-6 Astra
- Time/task
- Cost/task
- Tasks solved on the first run
- Relative value
Compare configurations4
| Configuration | Value | Time/task | Cost/task | Tasks solved on the first run | Tasks solved within four runs |
|---|---|---|---|---|---|
| GPT-6 Astralow | 1.00× | 15 min | $3.27 | 67% ±1% | 80% |
| gemini 3 8 flashhigh | 1.00× | 16 min | $3.20 | 74% ±1% | 86% |
| gemini 3 8 flashmedium | 0.92× | 19 min | $2.77 | 71% ±2% | 83% |
| GPT-5.6 Solhigh | 0.88× | 14 min | $5.00 | 69% ±1% | 87% |
Route workloads instead of standardising one LLM
An agentic workflow does not need to use the same LLM for every step. A powerful model can plan and prepare difficult decisions, while specialised or less expensive models handle subtasks such as computer vision, extraction, classification, or testing. The agent harness should select models according to the capability, risk, and cost of each step.
Only tests on the enterprise's own workloads show whether this orchestration works. The evaluation set needs typical cases, difficult edge cases, known failures, and clear acceptance criteria. An observability platform can also capture cost, latency, quality, logs, and traces for each agent run. Repeating the same evaluations through a second API or a self-hosted open-weights model shows whether switching is technically possible and economically worthwhile.[6]
Falling unit cost can still increase total spend
LLMs and chips become more efficient while providers continue to build more compute capacity. Five hyperscalers held an estimated 71 percent of the observed total AI compute capacity at the end of 2025. This shows market concentration and an expectation of rising demand, but it predicts neither future API prices nor the return on that investment.
The enterprise question is therefore not only whether one agentic workflow becomes cheaper, but whether work volume grows faster than unit cost falls. Higher capability makes more, longer, and parallel workflows economical. Cost controls belong at the level of completed work: budgets, routing, caching, concurrency limits, termination conditions, and measurable business impact.[7][8][9]
Measure outcomes, not output
When AI makes work less expensive, teams do more than move faster. They also take on workloads that were previously too costly and produce more pull requests, dashboards, or documents. That additional output does not show whether the enterprise is reaching its goals sooner.
Each agentic workload therefore needs a target metric, such as shorter processing time, fewer errors, avoided risk, or additional revenue. Throughput describes operations. The change in that KPI determines the value of the investment.[10]
Build capability now; defer irreversible commitments
Enterprises should start learning which agentic workloads are suitable, how AI output is accepted by the business, which data and tools agents may access, and how workloads move between models. A newly released LLM can then be tested against the same criteria within days without rebuilding the architecture.
Before investing in owned GPU infrastructure or long-term reserved capacity, enterprises need productive workflows, their own evaluations, and operating data. Only then can they estimate capacity demand, operating effort, and lifecycle cost realistically. Owned infrastructure can make sense when utilisation, data sovereignty, or latency justifies it; until then, an API or managed service remains more flexible. Better hardware may make self-hosted open-weights models more attractive over time.[11][12]
What enterprises must build now
- Select workloadsDefine routine cases, difficult edge cases, and business-critical situations.
- Define acceptanceSet the quality threshold, target KPI, and accountable business acceptance before testing.
- Observe agent runsCapture model version, tool access, cost, latency, failures, and human review effort.
- Test alternativesRun the same evaluations through a second API or a self-hosted open-weights model.
- Repeat evaluationsTest again when models, prices, requirements, or availability change materially.