Fortytwo Launches Engine API, Reporting 24.7% Fewer Tokens on Coding Tasks

Fortytwo has launched Fortytwo Engine, an inference API that uses cooperating AI models to improve reasoning while reducing the amount of generated text needed to complete a task. The company reports that its first supported model, Qwen 3.8 27B, used 24.7% fewer tokens across 450 DeepSWE software-engineering task runs than the same model running alone, while also solving a larger share of the tasks.

The announcement targets a growing cost problem for AI agents: a low price per token does not necessarily produce a low price per completed job. Long coding and research runs can accumulate repeated context, unnecessary tool calls, and retries. Fortytwo’s approach tries to reduce that wasted work inside the inference process, keeping its per-token price at the base model’s level, according to the company.

Steering Reasoning Before Mistakes Compound

Fortytwo calls the technology behind the new Engine its Swarm Decoder. Multiple model participants shape a shared reasoning trajectory while it is being generated, evaluating which direction is stronger before the response is complete. The goal is to catch weak assumptions early, avoiding the additional generation and recovery work that follows a wrong turn.

That is a change from Fortytwo’s earlier architecture. In its 2025 swarm inference research, separate nodes produced independent answers, compared them head-to-head, and combined the strongest contributions. The company describes a statistical ranking mechanism that gives more weight to proven nodes, with peer evaluation helping establish their reputations. That earlier approach evaluated completed answers; the new Engine moves peer ranking into the reasoning process itself.

The distinction matters for agents that work through many steps. Selecting the strongest finished answer can improve a response, but it does not recover the tokens already spent generating weaker alternatives. Continuous steering is intended to prevent an unproductive trajectory from growing in the first place. Fortytwo says its token accounting includes all participating models, rather than counting only the final response.

CEO and co-founder Ivan Nikitin described the benefit as twofold: “Collective reasoning pays off twice: the answers are better, and tasks take fewer tokens.”

What Fortytwo’s Tests Show

In the company’s DeepSWE evaluation, the Engine achieved a score of 47.8 compared with 42.2 for the base model, a gain of 5.6 percentage points. Average token usage fell from 16.3 million to 12.3 million per run. Across all 450 runs, Fortytwo reports totals of 7.4 billion tokens for the base model and 5.6 billion for the Engine.

The launch materials also report higher scores on five other benchmarks, spanning science, coding, and general reasoning:

Source: Fortytwo’s launch announcement. Scores compare Qwen 3.8 27B with and without the Engine under the company’s evaluation protocol.

These are company-reported results. The 24.7% token reduction comes specifically from the DeepSWE runs; it should not be read as a measured saving across every benchmark or every production workload. For developers, the practical test is whether a model completes their own tasks accurately with fewer total tokens, retries, and tool calls.

Familiar API, With Defined Compatibility Limits

The service is available through Fortytwo’s developer platform. Its API documentation describes compatibility with the OpenAI Chat Completions API, allowing an existing SDK or client to use a Fortytwo API key and base URL, together with a supported model ID.

Documented capabilities include streaming, function tools, parallel tool calls, and structured output. Developers can also select a reasoning-effort level. These features make the service relevant to agents that need to call software tools and return machine-readable results, rather than only generate conversational text.

Compatibility has boundaries. The documentation lists Chat Completions and model listing as supported endpoints, while APIs such as Responses, Embeddings, Audio, Files, and Assistants are not supported. Some request fields are only partially supported or depend on the selected model, and unsupported fields return an error instead of being silently ignored. Teams moving an existing application should therefore check its actual requests against the compatibility guide.

A Broader Bet on Collective AI

Engine sits alongside Fortytwo Harbor, the company’s offering for organizations that want swarm inference deployed on their own infrastructure. Fortytwo positions Harbor around keeping data inside the organization’s perimeter and using a fixed hardware cost rather than per-token API billing.

The company’s wider research program also includes generating training data through its model network. Its website highlights Strand-Rust-Coder-14B, an open model trained on a Fortytwo-generated dataset. That work addresses model training, while the new Engine focuses on improving how existing open-weight models reason at inference time.

Founded in 2024, Fortytwo says it has a team of 15 engineers, including in Mountain View and Abu Dhabi, raised $2.3 million in pre-seed funding in 2025, and joined Abu Dhabi’s Hub71 ecosystem in 2026. More open models are planned for the Engine after Qwen 3.8 27B.

The launch puts the emphasis on cost per successful task. If continuous peer evaluation can reliably prevent enough wasted reasoning to offset its own overhead, developers may have another way to improve agent economics alongside better models and faster hardware.