Foresight Review (2026): Pricing, Features & Honest Verdict

Foresight Review: The Verdict

Foresight by Lightning Rod is an AI forecasting model and API that predicts real-world outcomes across sports, politics, and markets. We found it surprisingly strong for its size, beating much larger frontier models on live prediction benchmarks. Best for developers and analysts building prediction features who care about cost. Price: free to try, paid and custom pricing on request. Rating: 7.5/10.

What is Foresight?

Foresight is a family of AI prediction models from Lightning Rod Labs. The public models are trained on real-world news and unstructured data to forecast outcomes in sports, politics, and markets, and you access them through an OpenAI-compatible API with research built in. The flagship is Foresight-32B, a compact 32-billion-parameter model that punches well above its weight.

It launched as generally available at foresight.lightningrod.ai. The company’s whole thesis is that you do not need a giant frontier model to forecast well. You need a smaller model trained specifically for prediction. They back this up. On a benchmark of 251 Polymarket questions, Foresight-32B beat o3, Claude Opus, Gemini-2.5-Pro, and Grok-4, models the team estimates are 10 to 100 times larger. Lightning Rod also sells custom models trained on a client’s own data, and counts Shore Capital, AirHelp, and Brunswick Group among its partners.

What Are Foresight’s Key Features?

Forecasting that beats bigger models

This is the part that matters. On 251 Polymarket questions, Foresight-32B scored a Brier score of 0.199, edging out o3 at roughly 0.206, while the untuned base model sat at 0.253. Lower is better. It was also better calibrated, with a 6.0 percent expected calibration error versus o3’s 7.5 percent. In a betting test, only Foresight-32B and o3 turned a profit. A 32B model doing that against frontier giants is the headline.

OpenAI-compatible API

You do not rewrite your stack to use it. Foresight ships a drop-in, OpenAI-compatible API, so if you already call GPT or Claude, swapping in Foresight for prediction tasks is a small change. The research step is built into the model, which means it gathers and reasons over current information rather than relying on a frozen training cutoff.

Cheaper inference at scale

Because the model is compact, inference costs far less than running a 100x-larger frontier model for the same forecasting job. For anyone making thousands of predictions, sports props, market moves, risk scores, that cost gap is the difference between a feature that ships and one that dies in the budget meeting. Smaller and better is a rare combination.

Custom models on your data

Beyond the public models, Lightning Rod trains custom prediction models grounded in a client’s proprietary data, with secure cloud deployment. This is the enterprise path. If your edge is your own dataset, a model tuned on it will beat a general forecaster, and that is what this tier is for.

How Much Does Foresight Cost?

Foresight is free to try through the public models, which is the right way to start. Lightning Rod does not publish per-token API pricing on the site, and custom model training is handled through a direct conversation with their team rather than a checkout page.

That opacity is a real drawback. The whole selling point is cheaper inference than frontier models, and frontier APIs publish their prices to the cent. To prove the cost claim, Foresight needs to show its numbers. For now you can test it for free, but budgeting a production deployment means an email rather than a pricing page.

Who is Foresight Best For?

Use Foresight if you are a developer, quant, or analyst building a product that needs probabilistic forecasts and you do not want to pay frontier-model rates for every call. Prediction-market traders, sports analytics teams, and risk desks are the obvious fit. The OpenAI-compatible API means you can A/B test it against your current model in an afternoon.

Skip Foresight if you need a general-purpose LLM for writing, coding, or chat. This is a specialist, not a Claude or GPT replacement. And note the honest caveat from the benchmark itself: the Polymarket crowd, the market baseline at 0.170 Brier, still beat every model tested. The crowd is hard to outpredict. Foresight is the best of the models, not better than the market.

Best Foresight Alternatives

OpenAI o-series API is the obvious comparison and the model Foresight beats most directly on these benchmarks. It is more general and more expensive for forecasting, but it is battle-tested and easy to adopt. Use it if you want one model for everything.

Anthropic Claude API is another frontier option with strong reasoning. It underperformed Foresight on the Polymarket test, but for blended workloads where forecasting is one of many tasks, a single capable model can be simpler than running a specialist alongside it.

Polymarket and Metaculus are the non-AI route. If you just want a good probability and not a model in your product, crowd forecasts are still the benchmark to beat, as Foresight’s own results show. Free to read, but you cannot embed them as an API the way you can with Foresight.

Final Verdict: Is Foresight Worth It?

Foresight is a sharp, focused product with the benchmarks to justify the hype. A 32B model beating o3, Claude Opus, Gemini, and Grok on live Polymarket predictions, while being far cheaper to run, is a genuinely impressive result, and the OpenAI-compatible API makes it painless to test against whatever you use today. If you are building anything that forecasts outcomes at scale, this belongs on your shortlist.

The caveats are real but manageable. Pricing is not public, so you cannot fully validate the cost story without contacting them, and the market crowd still beats the model, so set expectations accordingly. But it is free to try, the integration is trivial, and the core claim holds up under its own published numbers. For prediction-heavy products, Foresight is worth an afternoon of testing this week.