Free to Run, $410 Million to Build: Recursive's Flint Re-Prices Cloud Inference

Recursive Superintelligence, founded in 2025, published Flint's weights on Hugging Face on October 21, 2026 and a lightweight inference engine called rec on GitHub. The model is a mixture of experts with 39.8 billion total parameters and roughly 2.5 billion active per token, a documented 262,144-token context window, and an output head quantized to four bits with an offline distillation stage that nearly halved training steps against a quantization-aware baseline. Recursive says its automated research system, the Eureka Machine, worked on the data, the architecture search, the optimizers, the training recipes, the infrastructure and the inference kernels, including a Metal megakernel that fuses the decode forward pass on Apple M4 and M5 silicon. The whole build took under four months with a team of fewer than thirty people.

Frontier coding agents are priced per token, and those prices are already compressing toward the middle of the stack. Anthropic lists Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens, with cache reads at $0.20 and batch calls at half price, while Claude Sonnet 5.5 sits at $2 and $10 and Claude Haiku 5.5 starts at $0.10 and $0.50. Multiply that by what a laptop can do. At the 125 tokens per second Recursive reports on an M4 Pro, one hour of local generation is roughly 450,000 output tokens, a hosted-price equivalent of about $9 at Opus output rates, $4.50 at Sonnet, $2.70 at Grok 4.5 and $0.23 at Haiku. A developer with the hardware already owns the cheapest coding tokens in the market.

This is the tension that makes Flint a cost-structure event rather than a hobby release. Recursive emerged from stealth in May 2026 with $650 million of funding at a $4.65 billion valuation from GV, Greycroft, NVIDIA and AMD Ventures, then signed a $410 million compute deal with Amazon Web Services in June 2026, a commitment Recursive said has absorbed most of its funding. The lab that says its models are built by AI paid a cloud provider for the privilege. On-device inference changes the shape of cloud spend rather than its size: a recurring per-token bill becomes a lumpy per-training-run bill, paid up front and amortised into free weights.

The clearest transmission into a reported income statement so far is Microsoft. GitHub Copilot moved to usage-based pricing in June 2026, Copilot revenue accelerated more than 60% quarter over quarter on that change, and 50 million developers use the product. Under per-seat licensing, a software vendor was insulated from token deflation; usage-based pricing removes that insulation. Azure revenue passed $100 billion in fiscal 2026, up 41%, with $41 billion of capital expenditure in the June quarter alone and a plan to roughly double capacity in two years. Microsoft now sells consumption it can no longer price at its own convenience.

CoreWeave is the cleanest example of what deflation does to a fixed-cost layer. In the June 2026 quarter it booked $2.575 billion of revenue and $1.696 billion of gross profit, yet depreciation and amortisation alone was $1.393 billion, 54% of revenue, and the company still reported an operating loss of $49 million and a net loss of $626 million. Every one of those dollars was bought against contracts that assume compute keeps a scarcity value, and a free, unrestrictively licensed model running on a laptop is a direct argument against that assumption.