On February 23, 2026, Anthropic published a blog post arguing that AI can take on the tasks that once made COBOL modernization cost-prohibitive, and that teams could modernize “in quarters instead of years.” IBM shares fell about 13% that day, their steepest drop since October 2000. On October 19, OpenAI reported that its newest model migrates COBOL programs with 86.6% accuracy at low reasoning effort, up from 5.4%, and 97.9% at higher effort. Two days later IBM beat its quarterly estimates and called application modernization a long-term cycle it is still early in. The market read each announcement as a threat to the company that sells modernization. But the three companies, read at source, describe the same hard part. Anthropic: “you’re reverse engineering business logic.” IBM: “Translation captures almost none of the actual complexity.” OpenAI: “important business rules may only exist in code,” and production use “still requires expert review and testing against the system’s business rules.” There is a further twist: IBM had partnered with Anthropic in October 2025 and with OpenAI in August 2026, putting OpenAI’s models inside its own consulting platform. The translation is getting cheap. The understanding is not.
COBOL is old, and it is everywhere. OpenAI’s post repeats the commonly cited figures: it supports more than 40% of online banking systems, 80% of in-person card transactions and 95% of ATM transactions.[1] The people who wrote it are leaving. Anthropic’s February post put it plainly: “The developers who built these systems retired years ago,” and modernization means “reverse engineering business logic from systems built when Nixon was president.” Its claim was that AI “excels at streamlining the tasks that once made COBOL modernization cost-prohibitive,” so that teams could modernize “in quarters instead of years.”[2]
The market heard a threat to IBM, whose mainframes run much of that code and whose consultants sell its modernization. IBM shares fell 13.2% on February 23, their biggest one-day drop since October 18, 2000, erasing about $30 billion of market value.[3][4] IBM answered within days. Rob Thomas, its senior vice president for software, wrote that “Translation captures almost none of the actual complexity.”[5] An Anthropic spokesperson later described the capability as not new and the post as closer to a marketing message than a product launch.[6]
On October 19, OpenAI reported the number the February post did not have. On its COBOL migration evaluation — move a COBOL program to C++, Java, Kotlin, Python or Rust without changing what it does — GPT-6.1 Sol scored 86.6% at low reasoning effort, against 5.4% for GPT-5.6 Sol at the same setting, at 61% of the cost; at higher effort, 97.9%, “close to saturation.”[1] The grading is strict: hidden tests require “byte-for-byte parity on observable outputs,” and a migration counts only if every test passes.[1]
Two days later IBM reported its third quarter: revenue of $17.1 billion and operating earnings of $2.95 a share, ahead of the $2.86 consensus, with software up 9% and consulting up 2% at constant currency.[7] Arvind Krishna said IBM remains “in the early stages of a long-term application modernization cycle.”[8] The shares closed up 3.96% at $251.49 and gave the gain back after hours, with reported declines between 3.7% and 5.1%.[7][8] The announcement that was supposed to be the threat landed on a company reporting growth in the very business it threatened.
Up from 5.4% for the previous model. A migration counts only if every hidden test passes byte for byte. The number, size and source of the programs are not stated, and the COBOL evaluation is not released.[1]
Read at the source, the three companies agree on more than the market reaction suggests. Anthropic’s post says that deciding what to modernize first, and which business scenarios need validation by subject-matter experts, “is where human judgment becomes essential.”[2] IBM says translation misses almost all of the complexity.[5] And OpenAI’s own post opens with the same problem: “documentation may be incomplete, original developers may have left the company, and important business rules may only exist in code.” It names the example that makes the point: COBOL’s fixed-point arithmetic “discards excess decimal digits unless rounding is explicitly requested,” so a rewrite that rounds “can run without errors while producing different financial results.”[1]
That example is the whole case in one line. A 97.9% score means the migrated programs matched the originals on the cases the hidden tests check. Whether a migration keeps a rule depends on whether a test exercises it — and in a real bank, nobody has those tests waiting. Someone first has to find the rules that exist only in code. OpenAI says so itself: production use “still requires expert review and testing against the system’s business rules and dependencies.”[1]
The partnerships complete the picture. IBM announced a partnership with Anthropic on October 7, 2025, to bring Claude into its software.[9] On August 13, 2026, it announced one with OpenAI, embedding OpenAI’s models in IBM Consulting Advantage, the platform its consultants use with clients.[10][11] The models the market treated as IBM’s replacement are, in part, IBM’s tools. Cheaper translation does not remove a migration; it can make more migrations affordable. What stays expensive is the understanding.
AI takes on the tasks that made modernization cost-prohibitive; “quarters instead of years.”[2]
Claim86.6% at low effort, up from 5.4%; 97.9% at higher effort, on hidden byte-for-byte tests.[1]
ClaimTranslation captures almost none of the actual complexity. — Rob Thomas, IBM, February 2026 [5]
| Dimension | Evidence |
|---|---|
| Quality (D5) Origin · 84 | Behaviour-graded COBOL migration: hidden byte-for-byte tests, all must pass; programs not disclosed.[1]5.4% to 97.9% |
| Revenue (D3) L1 · 80 | IBM −13.2% on Feb 23 (about $30 billion); Q3 revenue $17.1 billion, EPS $2.95 vs $2.86.[3][4][7]−13%, Then a Beat |
| Operational (D6) L1 · 74 | Production use “still requires expert review and testing against the system’s business rules.”[1]Translation Cheap, Validation Not |
| Employee (D2) L2 · 66 | “The developers who built these systems retired years ago.”[2]The Knowledge Retired |
| Customer (D1) L2 · 62 | COBOL supports more than 40% of online banking and 95% of ATM transactions.[1]Rules Nobody Wrote Down |
The cascade originates in D5 (Quality): a measured jump in what models can do with COBOL, from 5.4% to 97.9% on a behaviour-graded test.[1] It propagates to D3 (Revenue), where IBM lost about $30 billion of value in a day on the first claim and reported modernization growth after the second,[3][7] and to D6 (Operational), the migration work itself, where translation gets cheaper and validation does not.[1][2] It lands in D2 (Employee), the dwindling number of people who know why old programs behave as they do,[2] and in D1 (Customer), the banks and agencies whose systems carry rules nobody wrote down. D4 (Regulatory) is not scored.
The case shares a shape with UC-324 and UC-320: the same AI, read in opposite ways by people with different stakes. It also connects to UC-337, where AI described the past from whatever was free to read. COBOL is what OpenAI calls a low-resource language: less public code to learn from, which is why models struggled with it until now. The honest limits: the COBOL evaluation’s programs are not described or released; the evaluation is OpenAI’s own; after-hours price reports conflict; and a stock’s move on a single day has many causes, including, in July, IBM’s own weak preliminary quarter.[12]
-- UC-338: Translation Was Never the Hard Part: 6D Diagnostic Cascade
-- Anthropic, COBOL Modernization with AI, Feb 23 2026: quarters instead of years; reverse engineering business logic; human judgment essential. IBM fell 13.2 pct Feb 23, biggest drop since Oct 18 2000, about 30 billion dollars. IBM Rob Thomas: translation captures almost none of the actual complexity. OpenAI, Advancing coding in low-resource programming languages, Oct 19 2026: COBOL migration to C++, Java, Kotlin, Python, Rust; hidden byte-for-byte tests, all must pass; GPT-6.1 Sol 86.6 pct low effort vs 5.4 pct GPT-5.6 Sol at 61 pct cost; 97.9 pct higher effort; programs not described or released; production still requires expert review against business rules. IBM partnerships: Anthropic Oct 7 2025; OpenAI Aug 13 2026, models in IBM Consulting Advantage. IBM Q3 2026 Oct 21: revenue 17.1B, EPS 2.95 vs 2.86, software up 9 pct, consulting up 2 pct cc; early stages of a long-term application modernization cycle; close up 3.96 pct, after hours down 3.7 to 5.1 pct.
FORAGE translation_was_never_the_hard_part
WHERE ai_cobol_migration_score_jump = true
AND market_read_as_threat = true
AND business_rules_named_hard_part = true
ACROSS D5, D3, D6, D2, D1
DEPTH 3
SURFACE translation_was_never_the_hard_part
DIVE INTO priced_the_translation
WHEN evaluation_programs_public = false
AND vendor_partners_with_model_makers = true
TRACE capability_to_market_cascade
EMIT translation_was_never_the_hard_part_signal
DRIFT translation_was_never_the_hard_part
METHODOLOGY 88
PERFORMANCE 40
FETCH translation_was_never_the_hard_part
THRESHOLD 1000
ON WATCH CHIRP high 'Anthropic's February COBOL post took 13 percent off IBM in a day; OpenAI's October evaluation showed 5.4 to 97.9 percent on behaviour-graded COBOL migration, and IBM beat estimates two days later. All three companies name the same hard part: business rules that exist only in code. The translation is getting cheap; the understanding is not.'
SURFACE analysis AS json
Runtime: @stratiqx/cal-runtime · Spec: cal.semanticintent.dev · DOI: 10.5281/zenodo.18905193
A 13% fall on a blog post about translation. All three companies say the expensive step is understanding what the code is for.[2][5]
The score checks behaviour on hidden tests. A rule no test exercises cannot fail one — and the tests are not public.[1]
Truncate or round: a rewrite can run without errors and still change the money. That is a business rule living only in code.[1]
Primary sources are OpenAI’s October 19 post and Anthropic’s February 23 post, read in full, and IBM’s third-quarter earnings call. Market figures come from Reuters-based and financial press coverage; after-hours prices conflict across sources and are reported as a range. Source 13 is an interactive companion made for this case.
This case follows a market that priced the cheap part of modernization and skipped the expensive part. Every organisation with a system older than its staff has rules nobody wrote down. If you want to find yours before a migration does, write.