Anthropic’s Cheaper Opus 5 Overtakes Fable 5 in Corporate Spending

Miles Clements, a partner at Accel, has watched the choice from inside an investment firm that put close to $1 billion into Anthropic. “Most people don’t need to operate at the frontier,” he told the Financial Times. The period when customers routinely chose the most advanced option “was not a durable era.”

Within a month of its launch, Anthropic’s cheaper Opus 5 had overtaken its flagship Fable 5 in corporate spending.

The reversal describes a market in which buyers can move between models with little technical friction and increasingly judge them by the cost of an acceptable result. It also sets Anthropic’s daily-use product against the flagship whose capabilities are meant for work that lasts for days.

The spending signal

Ramp chief economist Ara Kharazian saw the split in purchases. In July 2026, Fable supplied 6% of the Anthropic tokens identified by Ramp’s management product but accounted for 11.4% of model-attributed Anthropic spending. By August 23, spending on Fable had settled near 11% of identified spending on Anthropic models, more than two months after its release. The same August 23 report said Opus 5 had surpassed Fable 5 in business spending, but did not disclose Opus 5’s exact share.

Tokens are small pieces of text and other data that a model reads or produces. Many business customers pay for those units. A higher token price can still yield a lower bill if the system needs fewer steps, but it can also make every failed attempt and revision more expensive.

Ramp placed Fable at roughly $10 per million tokens for its comparison, about twice the price it assigned to GPT-5.6 Sol. In July, Fable produced about 75% as much model-attributed spending as Sol, even though Fable was positioned as Anthropic’s strongest generally available product. Those figures measure purchases Ramp could assign to a particular system. They are not Anthropic revenue.

Ramp’s evidence has a boundary. Its model-level sample comes from a token-spend product and leans toward technology businesses. It cannot show why a particular company selected Fable, Opus, or a rival, and it does not represent the entire market.

A conflicting market

Vercel’s production traffic cuts against the weak-Fable reading. Its August index, based on requests sent through the company’s AI Gateway in July 2026, put Fable at 13.2% of all gateway spending, second only to Opus 4.8. The comparison is with Opus 4.8 because the index covers July 2026; Opus 5 launched on July 24 and was available for only the final week of that monthly window. Nine in 10 teams using Fable that month were new to the product.

Anthropic collected 65.1% of Vercel gateway spending on 30% of tokens in July. The average price of an Anthropic token was 4.4 times the average across other labs, yet customers continued routing work to it.

Vercel sees another nonrandom slice of demand. Its spending figures value traffic at published list prices rather than customers’ negotiated bills, and the gateway does not cover the whole business market. Ramp and Vercel therefore offer conflicting readings without either one settling Fable’s overall position.

Their disagreement still reveals how quickly the mix can move. Changing a model on Vercel’s gateway requires one line of code. Across the gateway in July, teams increased token volume by 59% and spending by 37%, while the average price per token fell 13.6%. Open-weight systems reached 36% of token volume. Among teams that ran more than 10 million tokens in both June and July, three in four changed at least a tenth of their model mix.

The price of a finished job

Anthropic released Opus 5 on July 24 at $5 per million input tokens and $25 per million output tokens, half Fable’s rates. Input tokens cover the material sent to a system. Output tokens cover what it generates. The sticker prices do not capture the full cost of getting useful work back.

A cheaper system may need more attempts, longer prompts, or more human review. Cost per successful task adds those steps together and asks what the finished job required, not what one token cost. On Anthropic’s evaluations, Opus matched or beat Fable on bounded coding and office work at a lower cost per completed task. Fable remains intended for long autonomous projects that have to stay coherent across connected steps.

Model routing puts that calculation into software. A company can send routine work to a cheaper system and reserve a costly one for assignments that fail elsewhere or require sustained autonomy. Low switching costs make the choice more like a dispatch rule than a long-term software commitment.

Mantas Lukauskas, the AI tech lead at website host Hostinger, has used large language models since 2020. He described the recent pricing changes as the “first real test” of whether U.S. labs can preserve prices for their most advanced products.

The missing denominator

Anthropic gives API customers exact rates and token counts, while many enterprise customers can inspect usage through analytics. The value calculation is less exact for individual Pro and Max subscribers whose consumption is included in a plan.

Those customers can see progress bars, reset clocks, percentages, activity statistics, and a usage breakdown. Max sessions reset every five hours, and weekly limits reset at a fixed account time. Anthropic may also impose monthly, model, or feature caps.

The public plan documents do not state how many tokens equal 100% of the included five-hour or weekly allowance. That absence concerns the denominator for Pro and Max subscriptions, not API pricing or enterprise token reporting. A subscriber can see how much of an allowance remains without seeing its total token size, which makes the value of one model versus another harder to calculate before the bar moves.

Growth beyond the flagship

Weak Fable demand in one dataset has not meant weak demand for Anthropic. Its annualized revenue rate reached $65 billion in July 2026, up from $47 billion in May. The rate extrapolates recent performance over a year; it is not booked full-year revenue. Preliminary second-quarter revenue exceeded $11.5 billion, about 14 times the year-earlier quarter, and the company recorded positive adjusted operating income.

Ramp’s broader July sample found 43.5% of U.S. businesses paying for Anthropic products, compared with 39.7% for OpenAI. That is provider adoption, not Fable adoption. It can rise while customers inside Anthropic’s catalog move down from the flagship or divide work among several systems.

Dianne Penn, Anthropic’s product leader, said customers should choose Opus 5 for value and Fable 5 for “days-long, very autonomous projects.”

Meta Spends Hundreds of Millions a Year on AI Models Through Microsoft

Meta has become one of Microsoft’s largest customers for artificial intelligence models sold through its Azure cloud, spending hundreds of millions of dollars a year for access through Microsoft Found

The Implicator

Meta Opens Muse Glimmer as Zuckerberg Presses U.S. AI Policy Shift

On Aug. 10, 2026, Meta released Muse Glimmer, a 30-billion-parameter model whose weights developers can download. It was built for agents that run on a Mac or PC. The release carried a policy argumen

The Implicator

Moonshot Attaches $20 Million Revenue Clause to Kimi K3 Open Weights

Moonshot AI on Monday published the custom Kimi K3 License with the model's downloadable weights, setting a $20 million aggregate-revenue trigger for Model as a Service operators. The agreement is not

The Implicator

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai