OpenAI has released GPT 5.5 as a research preview, recording one of the highest scores ever seen on the company's internal benchmarks at 91.7. The model has performed strongly across transactional and litigation-focused legal work and code generation, with OpenAI describing it as its strongest agentic coding model to date.
The release signals that the performance gap between GPT and rival models from Anthropic and Google DeepMind is narrowing, although the model's wider impact will depend on when it becomes available through the API.
Quality per dollar becomes the metric that matters
For the past two years the AI industry has prioritised raw quality over latency and cost. That is starting to change. GPT 5.5 has shown a 50 per cent reduction in reasoning tokens in some experiments, a shift that reflects growing pressure on labs to deliver performance gains without spiralling compute bills.
As companies mature, the focus is moving towards quality per dollar spent and quality per token used. The efficiency gains in GPT 5.5 suggest OpenAI recognises that brute-force scaling alone will not sustain the economics of frontier AI.
The profitability question looms
The broader concern across the industry is what happens when companies such as OpenAI begin charging prices that reflect their true costs. The current generation of AI services is heavily subsidised, and when that changes, application-layer companies may struggle to absorb the expense.
This is particularly relevant for vertical AI firms such as Harvey, the legal technology company that has built its business on top of frontier models. Nico Grupen, head of applied research at Harvey, has acknowledged the challenge of differentiating against a company like Anthropic that is moving directly into verticals including legal services.
Watch what labs say, not just what they ship
There is a pattern worth tracking in how AI labs communicate. When a company sits at the top of the benchmark rankings, its press releases emphasise model capabilities. When it falls behind, the messaging shifts towards product features and applications.
This oscillation is visible across OpenAI, Anthropic and Google DeepMind. It offers a useful lens for reading corporate announcements, particularly during the six to eight week gaps between major model launches when labs are working through post-training cycles.
Competition drives the pace
The expectation remains that the major labs will continue pushing the boundaries of model intelligence. Consumer demand and competitive pressure leave little room for any of them to ease off, even as the economics of the business remain unresolved.
Google DeepMind's Mythos model is also being watched closely, though it remains unclear whether it will be made publicly available. Some observers have raised security concerns about a potential release, adding another layer of complexity to an industry that is simultaneously racing to build more powerful systems and grappling with how to deploy them responsibly.
The release of GPT 5.5 confirms that the frontier is still advancing. The question now is whether the business models underpinning that advance can keep pace with the technology.
