Claude 2 turns the frontier model layer into a choice

Anthropic shipped Claude 2 with a 100,000-token window at an unchanged price. A second top-tier supplier shifts pricing power toward builders.
Claude 2 reads up to 100,000 tokens in one prompt, enough for hundreds of pages of documents, and Anthropic charges the same API price it charged for Claude 1.3. Released on 11 July through the API and as a public beta chat for anyone in the US and the UK, the model takes in more than ten times the 8,000 tokens of the generally available version of GPT-4.
For most of this year, building on a top-tier model meant building on one supplier, whose premium tier set the margin of every application. Now two labs offer models of a similar class on broad API access, and the challenger competes on context length and price rather than raising it. When a second supplier holds price while improving quality, model access starts to behave like a procured input, and pricing power shifts toward the buyers.
The quality gains are measurable. Claude 2 scored 76.5% on the multiple-choice section of the US bar exam, up from 73.0% for Claude 1.3, and 71.2% on a standard Python coding test, up from 56.0%. OpenAI moved in the same week: five days earlier it opened the GPT-4 API to all existing paying developers and announced that its older completion models will be retired in January. The labs still need ever larger rounds to train the next generation while their product becomes easier to substitute, which ties their economics to scale and to distribution through cloud partners. Value moves toward the companies that own the workflow, the customer and the data.
Builders gain most by designing for more than one model from the start: the model behind their own interface, evaluation sets that compare outputs on real tasks, each task routed by cost and quality. Long context opens products that read a whole contract, codebase or set of filings in one pass, which matters most in legal, finance and software. OpenAI's retirement schedule shows the other side. A product tuned to one model version has a shelf life set by the supplier, and migrations cost engineering time and sometimes quality.
Two suppliers still make a duopoly, and a duopoly can hold prices. The next price moves from both labs will settle it, since a cut by either would confirm that competition has moved to price. Open-weight models reaching this class within the next year would add a self-hosted option alongside the two labs and loosen their grip further. Enterprise buyers decide the rest: if they sign contracts with more than one model provider instead of standardising on one, substitution becomes the normal state of the market.
Competition among model suppliers works as a subsidy to the layer above them. Fanaura looks for AI-native companies whose advantage sits in data, workflow and distribution, the profile our venture thesis favours, because for them a better or cheaper model next quarter adds to the margin instead of erasing the product.
Building an AI-native company? Apply in five minutes. →
More from Fanaura
Building AI-native products that switch between models and win on data and workflow? Send us the project.
Fanaura invests its own capital and the capital of its limited partners (LPs), usually through a separate SPV for each deal. If we see a fit, we get in touch.
Get the news by email
One short note each weekday. No marketing, and you can unsubscribe at any time.