Citigroup Adobe Restrict AI Flagship Models to Control Costs

Citigroup disabled Claude Opus 4.6, 4.7 and GPT-5.5 for employees on June 24 after GitHub shifted to usage based billing. Adobe ended its unlimited Claude agreement on June 30. Both moves followed similar steps at Atlassian, Amazon and two other unnamed companies that now route staff requests to smaller models.

A 404 Media review of internal documents showed that at least one of the six firms saw its monthly AI spend reach 15 million dollars, three times the prior level. Atlassian alone moved from 5 million dollars to 15 million dollars a month and projects more than 120 million dollars for the current fiscal year. 404 Media reporting Bloomberg follow-up The Verge analysis Reuters summary

The pattern is simple. Companies approved broad access to the strongest available models earlier this year. Usage grew faster than budgets. Finance teams then cut access to the highest cost options and directed employees toward lower capability alternatives.

What exactly changed in June

Citigroup acted first. Internal messages on June 24 told staff that GitHub Copilot would switch to per token billing and that the company would therefore block the three most expensive models. Employees were pointed to lighter versions inside the same interface.

Adobe followed six days later. The design software maker canceled its flat rate agreement with Anthropic. Staff who had used Claude without limits now face the same per use charges that apply to other tools. The company did not replace the model with a specific substitute but told teams to choose the cheapest option that still meets task requirements.

Atlassian data offers the clearest numbers. Its AI line item rose from 5 million dollars to 15 million dollars in a single month. The firm now expects annual costs above 120 million dollars if no further controls are added. Internal dashboards show the jump came from repeated calls to large context windows and long chain of thought traces.

GitHub itself signaled the next step. The platform plans to test an open source model option and a per user metering mode later this year. The combination would let companies keep costs predictable while still supplying AI assistance.

Who now carries the pressure

Engineering and design teams feel the restriction first. They lose the models that handled complex reasoning or large code bases in one pass. For instance, an engineer refactoring a 5,000-line module that once succeeded in one prompt might now issue three shorter prompts and spend extra minutes stitching outputs plus manual verification. Tasks that once ran in a single prompt now require multiple shorter calls or manual cleanup.

Finance departments gain visibility they lacked before. Previously, AI spend sat inside broad cloud contracts. The new billing models surface each token or query, making it easier to spot departments that exceed targets.

Vendors face a new sales reality. Anthropic and OpenAI can no longer count on unlimited seat deals from large accounts. They must instead sell tiered access or convince customers that smaller models deliver enough value for routine work.

The core tradeoff between capability and cost

Every flagship model carries higher per token prices because it uses more compute during inference. When thousands of employees issue daily prompts the difference becomes material. A three times increase in spend is common once usage moves from experiments to daily production.

Lower capability models reduce that bill but also reduce answer quality on hard problems. Internal tests at one firm showed a 22 percent drop in correct code suggestions when switching from Claude Opus to the next model down. The gap narrows for simple refactoring but remains noticeable for architecture changes or novel algorithms.

Companies are therefore creating tiered policies. Critical projects keep access to strong models after manager approval. Routine work shifts to lighter models by default. The policy creates friction that discourages overuse without banning the tools outright.

How the billing change accelerated decisions

GitHubs move to usage based pricing removed the last fixed cost buffer. Under the old seat model, extra usage did not trigger immediate invoices. The new structure bills every token, so finance teams saw real numbers arrive each month.

The timing mattered. Most firms approved broad access in the first quarter when budgets still assumed modest growth. By June those budgets were already exceeded. The GitHub change simply made the overrun impossible to ignore.

Adobe faced an additional factor. Its unlimited agreement with Anthropic was negotiated when usage was low. Once designers began using Claude for image descriptions and layout suggestions the volume grew beyond the original forecast. Canceling the flat rate restored cost certainty.

What remains uncertain

It is not yet clear whether the restrictions will hold once employees learn workarounds. Some teams may route prompts through personal accounts or third party aggregators. Others may slow their output rather than accept lower quality suggestions.

Vendors are also adjusting. OpenAI and Anthropic now offer volume discounts and reserved capacity contracts. These deals could restore access for approved departments without returning to unlimited terms.

The open source alternative GitHub plans to test may change the equation again. If a smaller model matches 80 percent of current performance at 20 percent of the cost, many restrictions could be lifted. Early tests are not public, so the outcome stays unknown.

What to watch in the next three months

Watch Atlassian quarterly filings for the next AI spend figure. If the number stays above 15 million dollars a month the controls are holding. A drop would show either reduced usage or successful migration to cheaper models.

Watch GitHub enterprise announcements for the first open source model option. A concrete release date and performance numbers will tell vendors and customers whether the cost curve can be bent without losing capability.

Watch internal policy documents from other large banks and software firms. If two more companies announce similar blocks the practice becomes standard. If none follow, the Citigroup and Adobe moves may prove isolated reactions rather than a lasting shift.

Employees at the affected companies now weigh every prompt against both time and cost. The experiment that ran without limits for six months has ended. The next phase will show whether lighter models can carry the workload or whether firms will need to restore selective access to the strongest tools.