AI solves
the hardest coding problems.

AI now sits on the side of the table that solves the world's hardest coding problems.

AI's native tongue is Python and Markdown-shaped text. This series goes one step further. The question is not language but capability level. The code AI writes has crossed a line. Past that line, the structure of software development itself rearranges. This chapter establishes where the line sits.

Coding ability can be compared as a number, on competitive-programming ratings

There is exactly one mechanism in the world that assigns objective numbers to coding ability. It is the public rating of competitive programming. Codeforces, AtCoder, and ICPC all accumulate, over years, whether you can solve set problems in time, how many you solve, and how correct your solutions are. Each participant ends up with a number.

Codeforces rating bands distribute roughly like this.

The numbers have steps in them. The gap between 1500 and 1800 closes with study. The gap between 2400 and 2700 does not close with study alone. Past that point you need speed, algorithmic design, and a nose for the hardest problems. The world's top sits between roughly 2700 and 3900, and contains around fifty people.

AI has entered the 2700 tier

Through late 2024 and into 2025, the situation changed. OpenAI's publicly reported estimated Codeforces rating for the o3-series models came in at around 2727, announced at the o3 launch. Google DeepMind's AlphaCode 2, a step before that, demonstrated top-15% Codeforces performance. Later research models have pushed further. Anthropic has reported continuous improvement in coding ability for the Claude family.

There is room to argue about where the numbers come from and how they are measured. But the fact that AI has entered the 2700 tier is confirmed by several independent announcements that move in the same direction. This is not "AI became a useful assistant." It means AI sits on the side that solves the hardest problems.

The reason this happened is that competitive programming is a domain with explicit rules. Grammar, the standard library, and the type system are all formally defined. Whether code compiles, and whether the output matches the expected values, can be judged mechanically. AI passes human levels in domains like this, where the rules are explicit and the answer can be checked. When this series argues that coders disappear, the claim is limited to domains with that property. Complete replacement at the same speed does not extend to other AI applications such as desk work, self-driving, or robotics. That boundary is treated in 3-09.

What matters here is not the rank. It is the structural change of crossing a threshold.

• Up to 2400 was a band a strong specialist could reach with enough drill
• 2700 is a band that holds a few dozen people worldwide
• AI entered that band by a different path than the one humans climb, one person at a time

For a human to reach this band takes thousands of hours of practice starting young. On top of that, it takes passing a talent filter. AI entered the same band without taking that path. There used to be an objection that the training data contained the same problems. But Codeforces keeps running live contests with fresh problems, and AI models have repeatedly been observed returning 2700-tier solutions there. A band humans reach one person at a time, over a decade and more, was entered by AI all at once, by several paths.

Because AI can design as well, it became the strongest SIer

Reaching the 2700 tier proves the ability to write code fast and correctly. That ability is coding. Design is a different ability. The core of design is understanding context. You read the target system or situation, grasp what it needs, and assemble a structure that fits. That is design. Writing fast and correctly is not the same as understanding context and deciding structure. So design ability needs its own evidence.

That evidence came from an unexpected place. In the Fable / Mythos generation, a publicly reported case had an attacker misuse Claude, and the AI ran most of the operational steps of an attack on the target autonomously. Anthropic reported it in 2025 as an automated cyber-espionage campaign. Driving an attack yourself requires scouting the target system, reading its weaknesses, and assembling the line of attack. That is, it requires understanding the target's context deeply. That is the same ability as designing a system.

The same ability is also the ability to verify. To attack is to find the hole and strike it. To verify is to find the hole and close it. Both are expressions of one power: understanding a structure more deeply than the person who built it. Attack, design, and verification are not separate abilities. They are three faces of one power. That is why an AI that can attack can also design and verify.

So AI reached the top tier in two separate abilities. One is writing code, shown by the 2700 tier. The other is understanding context to decide structure, which is design and verification, shown by the autonomous attack. *AI has begun to exceed senior engineers at both coding and design.* With both in hand, the conclusion is a single one. In the Mythos/Fable era, AI became the strongest SIer. It reads requirements, decides structure, implements, and runs the result. It does the whole of the SIer's work by itself. And it is not a scarce resource of a few dozen people worldwide. It is callable by anyone for $20 a month.

The world's top tier is reachable for $20 a month

This is where the series' argument starts.

The paths to top-tier coding ability used to be narrow. You got hired by Google, Meta, or Anthropic. You spent years climbing the competitive-programming ladder. Or you paid seven-figure salaries. Capability above the threshold was a scarce resource. Palantir's FDE (Forward Deployed Engineer) model is the extreme upper end of that legacy path. It embeds top-tier engineers inside the customer's organization on year-long, eight-figure contracts. The mechanics are covered in 3-05.

Access to AI models comes in tiers, depending on how hard you intend to use them.

• The base plan — Claude Pro, ChatGPT Plus, Google AI Pro, at around $20 a month. For light use — watching tools that are already running, fixing them now and then — this is enough. Start here.
• The building period — while you have AI writing code all day, day after day, use the tier above (Claude Max and the like, from $100 a month). Once the building is done and the tools are running, drop back to the base plan. The breakdown by period is in 2-02.
• API pay-as-you-go — you can pay for what you use, but *do not use it here*. Run the same volume through the API and it comes to roughly *ten times* the flat plan.

The gap has two causes.

One is that the volume stacks up. When the AI writes code while using tools, every call resends the whole conversation and every file it has read. The input side stacks up once per call.

The other is that the unit price differs. The flat plan is priced below what the same volume costs when bought by the meter. Add up the published per-token rates and you pass the monthly subscription by a wide margin — which means the discount sits on the subscription side. The usage limit is the boundary placed there in exchange.

Count it out. Sonnet 5 lists at $2 per million input tokens and $10 per million output tokens, with cache reads at a tenth of the input rate (as of September 2026). If one call carries 80,000 tokens of context and produces 2,000 tokens, and 90% of that context is served from cache, the call costs about $0.05. Two hundred calls a day, twenty days a month, and you are at roughly $200 — exactly ten times the $20 flat plan. Run the same volume through Opus 5 and it passes $500.

In other words, the world's top-tier coding ability is reachable for $20 a month. One credit card and one browser, and you can start the same day.

This is not a story about prices dropping. The axis of the price structure itself changed. Before, a scarce capability carried a large fixed cost. Now, a comparable capability carries something close to zero marginal cost. These are not the same spreadsheet at two prices. They are different supply curves.

This is where the IT revolution completes

Top-tier coding ability reaches anyone for $20 a month. What that fact means is not that AI got faster, or that AI got convenient. It means that what has long been called the IT revolution finally completes here.

Look at what the term "IT revolution" named, in structural terms.

• The industrial revolution — production of physical goods moved from human hands to machines
• The first wave of computing — calculation moved from human hands, the abacus and the human computer, to machines
• The IT revolution — business processing moved from paper and pen to software

In the first two, the core of the revolution, mechanization and automation, reached the object of the revolution completely. The third is different. Software itself was still written by human hands. The tool of the revolution, software, kept being produced by hand. That is a state in which the core of the revolution has not reached the production of its own tool. So the change called the IT revolution was only an incomplete form of revolution.

By the industrial-revolution parallel: the power loom exists, but the loom's own parts are still hammered out by hand at the blacksmith's. The revolution's loop does not close until it reaches the production of the tool itself.

Now that AI carries both code and design, the loop finally closes. The act of producing software is itself taken over by machines. The revolution's tool comes to be built by the revolution's own process. That is what the completion of the IT revolution actually means.

With that view, the changes this series covers — the coder role ending, the structural uneconomy of the SIer model, the rearrangement of employment and industry structure — stop being isolated phenomena. They read as one delayed revolution catching up and finishing at speed.

Looked back at from the AI revolution's side, this completion has a second reading. The decades called the IT revolution were the preparatory stage of the AI revolution. It is the same as the information revolution of movable-type printing, which had its own preparatory stage in the spread of paper and the accumulation of manuscripts. What the printing press detonated was the text those preparatory centuries had copied and stored. In the same way, what AI is detonating is the assets the IT decades stored up. Above all, that means OSS, the open commons of code and knowledge. The preparatory stage's greatest legacy was not the closed product lines. It was this open accumulation. That one point decides how tools are chosen (1-05) and how the whole independence part is arranged (2-01).

Summary

Every chapter that follows is deduced from this one point.

• 1-02 — once coding itself becomes cheap, where does the unit of maintenance move?
• 1-03 — what happens to the role whose center is writing code, the coder?
• 1-04 — what role remains in its place, the builder?
• 1-05 — when customers themselves pair with AI, what happens to the structure of outsourcing?
• 2-01 to 2-16 — if customers hold it themselves, how do they stand up the whole company IT foundation, from authentication and documents to mail and core systems, on their own side? That is the independence part.
• 3-04 — can the SIer commission model compete with AI sitting above the threshold?
• 3-05 — where do existing commission relationships act as lock-in?
• 3-06, 3-07, and 3-09 — hiring builders, the transition of the SIer industry, and the fact that the transition arrives in the near term and does not reverse

These questions are not independent observations. They all derive from one point: the strongest SIer is available for $20 a month. This chapter exists to plant that point.

And the story does not stop at writing code. Once customers can hold it themselves, authentication, documents, mail, and core systems all come to stand on their own side. This series runs continuously from software development to company-wide digital sovereignty, which is the independence part.

One more frame applies to everything that follows. This series covers structural change inside software development. It does not take up the extreme positions, neither "leave everything to AI and humans are not needed" nor "AI has no creativity, so the impact is bounded." Once AI above the threshold has been in the market for some years, how do the commissions, the outsourcing, the employment, and the prices of software development rearrange? This series answers that practical question, chapter by chapter.

One more thread runs underneath. If AI carries both coding and design, what remains on the human side is the broad work of building and operating a system, including hardware, people, operations, and responsibility. Its foundation is closer to the liberal arts than to software engineering. That thread runs through the whole series, and 1-04 takes it up directly.

The next chapter takes up the most overlooked consequence of cheap coding. That is the structural change of the maintenance phase.

Related articles

• 1-04: The Builder Role
• 3-09: The Structural Transition That Won't Reverse
• 2-04: Write the Logic — Own Your Tools with Python and Flet
• Structural analysis 08: Subtracting the enterprise-IT tax
• Structural analysis 12: AI and the sole proprietor