Foresight by Lightning Rod

• 1. overview (593 chars)

  Foresight is a prediction API for developers launched by Lightning Rod Labs. Its core capability is not to generate text that "sounds correct", but to return calibrated probabilities for any future real events. It is fully compatible with the OpenAI API format, and developers can access it by changing a few lines of code. The Brier Skill Score on solved real prediction market problems exceeds cutting-edge models such as GPT-5, Gemini 3.1 Pro, and Opus 4.6, while the inference price is only 1/4 to 1/2 of its opponents. Its underlying methodology, Future-as-Label, was selected as a Spotlight by the ICML 2026 AI Forecasting Workshop and included in the purchase list by the US Defense Innovation Market.

• 2. Product background (891 chars)

  Lightning Rod Labs is founded by Ben Turtel, who serves as CEO. Before founding the company, he sold the video platform Kazm to Harvard University and served as the founder and CTO of Rivet, an internal project of Google Area 120. He has more than six years of experience as a software engineer at Google and holds a master's degree in scientific computing from New York University. The company's technical roadmap can be traced back to a paper published in TMLR in May 2025 - a 14B model tied OpenAI o1 in prediction accuracy and generated more than 10% returns in real-time trading simulations.   The company's positioning is clear: general-purpose large models are trained to imitate text, while Lightning Rod advocates using real-world results to train models to make predictions. This idea was verified in July 2025 - Foresight-32B defeated a cutting-edge model that was a hundred times larger in real-time Polymarket data. Since then, the team has successively won first place in public benchmarks such as ProphetArena, ForecastBench, and sports forecasting.

• 3. Product functions and usage experience (1718 chars)

  What Foresight provides externally is a set of forecast models, delivered in the form of API. The calling method is exactly the same as OpenAI: point the client base_url to the corresponding address, and then pass in the model name and question. The difference is that there are two more prediction-specific parameters in the request body - research controls whether to automatically connect to the Internet to retrieve the context, and answer_type controls the return format. The final output of the model contains both probabilities and inference processes.   Its training method, Future-as-Label, is a real moat for the product. Traditional large models use massive amounts of static text to make predictions; Foresight uses real results in historical data that appear later than the prediction moment as labels, allowing the model to learn from real cause and effect rather than imitating general text. The immediate benefit of this is calibration: if the model gives an event a 70% probability, over the long term it should occur about 70% of the time. The ICML 2026 workshop paper shows that this method increases the Brier score by about 27%, reduces the calibration error by half, and can outperform larger models with a 32B model.   In terms of ability extension, Foresight does not only make public market and political forecasts. The team moved the same assembly line to multiple vertical domains: using a model trained on 10-K financial reports to predict the risk materialization of listed companies, the accuracy is the same as GPT-5 and three times better calibrated; using the Federal Reserve Beige Book to predict regional economies, the accuracy is about 26% higher than GPT-5; using public news and trade indexes to predict supply chain disruptions covering 25 countries and 88 commodities, the accuracy and calibration are about 34% and 59% better than GPT-5 respectively; in MIMIC-III predicts clinical events on clinical records, the calibration error decreases by about 70%, and the Brier score is slightly better than GPT-5.

• 4. Pricing and business model (817 chars)

  The public model is priced by token, and Foresight is $6 per million output tokens. Compared with GPT-5, Gemini 3.1 Pro, GPT-5.4, and Opus 4.6, the prices are 1.7 times to 4.2 times lower than the latter respectively. The combined cost per thousand predictions is also significantly lower than the general-purpose model. New users can get $50 in free credit when they sign up, and you can use a discount code to get another $50 during the Product Hunt event.   A business model has two legs. One is a public inference API for developers and quantitative teams; the other is privatized customization for enterprises and governments - customers use their own messy data to automatically generate training sets through the SDK and fine-tune exclusive prediction models, which are deployed in their own cloud and the data does not leave the domain. The government has passed DARPA's relevant qualification review and can be directly purchased through the federal innovation market.

• 5. User reviews (973 chars)

  What supporters buy most is the entry point of “calibration probability”. Many developers on Product Hunt said that instead of asking a general large model for a probability without calibration basis, it is better to directly get a scored prediction; it is especially valuable for scenarios such as prediction market robots, market making, and risk monitoring that require direct action probabilities. OpenAI-compatible interface form has also been repeatedly praised. The product ranked third on Product Hunt that day when it was launched.   Negative feedback focuses on three points. First, heavy use must be paid, and there is no free version, which is a threshold for teams who want to verify first. Secondly, the probability output itself still requires downstream interpretation and secondary calibration, and cannot be directly used as truth. The third is the correlation risk raised by the community: when a large number of prediction agents all run the same underlying model, their predictions will tend to be correlated, which will change the nature of systems that rely on independent signals. There are also concerns about calibration drift.

• 6. Industry evaluation (795 chars)

  On an academic and benchmark level, Foresight has a pretty solid record. It ranks first in the ProphetArena overall list by Brier score, leaving behind GPT-5, Gemini 3 Pro and all cutting-edge models, and also ranks first in the sports score list; it enters the top five on ForecastBench, surpassing Gemini 3 Pro, Claude Sonnet 4.5 and o3; v3 falls into the middle range of super forecasters on ForecastBench, and is run on a single GPU. Several third-party tool sites list it as a more accurate and cheaper prediction API than general-purpose LLM.   From a horizontal perspective, Foresight’s differentiation is not just another large model, but modeling the prediction itself as an independent problem. It does not try to cover general capabilities such as chatting and writing, but specializes in the vertical capability of calibrating probability, and trades lower reasoning costs for the feasibility of high-frequency calls in the agent loop.

• 7. Controversies and Risks (674 chars)

  The risk mainly comes from the product positioning itself. First, the prediction market is highly sensitive to calibration. Once the real settlement results deviate from the long-term calibration curve of the model, trust will be quickly lost, and the official disclosure of the monitoring mechanism for calibration drift is limited. Second, correlation risk is as mentioned above. Third, high-consequence, low-probability events may require action even if the probability is low. How the model can give actionable tips under low probability has not been fully explained in official documents. Fourth, enterprise privatized training relies on the quality and time series structure of the customer's own data. If the historical data itself is biased, the prediction will also inherit the bias.

• 8. Applicable groups and usage suggestions (950 chars)

  This product is most suitable for teams that are already using the OpenAI API and need to answer what will happen next in their workflow: developers of prediction market making and arbitrage robots, agent builders who need forecasting tools, and risk and quantification teams who want to monitor supply chain shocks, policy actions, and financial surprises. It is also suitable for enterprises and government agencies that want to turn their own historical data into deployable predictive models.   The groups that are not suitable for it are: ordinary users who just want to chat about the future in the dialog box; individual players who rely heavily on free quota and are unwilling to pay the API key first; and decision-making scenarios that require the model to give deterministic conclusions rather than probability distributions. If your need is for general writing or Q&A, it is more cost-effective to directly use cutting-edge general models; if your need is for thousands of probabilistic predictions that can be directly consumed by the program, Foresight's cost-effectiveness and calibration quality are worth a try.

• 9. Summarize (551 chars)

  Foresight redefines prediction from a bonus capability of general-purpose large models to an independent engineering capability that can be verified by benchmarks - trained with real results, delivered with calibrated probabilities, and lowered the access threshold with OpenAI-compatible interfaces. In scenarios such as prediction markets, risk monitoring and agent workflows, it is currently one of the cheapest and most accurate options on public benchmarks; whether it can continue to fulfill its calibration commitments and resolve correlation risks after large-scale deployment will determine how far it goes from a tool to a platform.

No in-depth report yet

• 

  彭萍桂

  —

  The angle of calibration probability is really tricky, and it is much more reliable than asking a general large model to make an unfounded guess.

---

• 

  Shirley_Torres_2022

  —

  It’s only $6 per million tokens, more than four times cheaper than Opus, and it’s really delicious in large quantities.

---

• 

  JeffreyPowell_99

  —

  OpenAI compatibility is so trouble-free. Just change the base_url to connect to the existing agent pipeline, with zero reconstruction.

---

• 

  u0ns5

  —

  My biggest concern is calibration drift. It’s true that the model won on Polymarket, but once the incident is largely resolved, can it maintain long-term calibration? The official said that it will continue to compare with real-time benchmarks, but it is not clear whether it is streaming verification or periodic evaluation. This is particularly important for real offers.

---

• 

  PatrickHendersonQ

  —

  We used GPT-5 to run a prediction market robot, and the monthly token fee was terrifying. After switching to Foresight, the cost was directly cut to a quarter, the calibration is even better, and I am ready to make the full cut.

---

• 

  iBertaFleury_pro

  —

  Some people use it to predict supply chain interruptions, covering 88 commodities in 25 countries. The accuracy is more than 30% higher than GPT-5, which is much more pragmatic than the pure chat model.

---

• 

  Steven_Russell_77

  —

  Product Hunt third of the day, something.

---

• 

  Douglas_Hall_77

  —

  Correlation risk is a real issue. If a bunch of prediction agents all adjust the same underlying model, their probabilities will become more and more similar, and in the end the system will be full of related signals. Officials say that enterprises can privatize training, but how to break the general API layer? It is hoped that the ensemble interface will be opened in the future.

---

• 

  BWood_7703

  —

  Heavy use requires payment, and there is no free version. Individual players are blocked out for the first time.

---

• 

  Megan_Hicks

  —

  Our risk control team took over Foresight to monitor policy actions and financial reporting surprises. In the past, analysts relied on slapping their heads to give probabilities, but now they directly feed news and SEC documents to produce calibrated distributions. Decision-making meetings finally have numbers that can be questioned and traced. However, low-probability and high-consequence events still require people to keep an eye on them, and they cannot fully trust the model.

---

• 

  3611ENAW

  —

  Set answer_type to auto and directly spit out the calibration probability in the <answer> tag, which is very easy to parse. With research=True, it can also automatically pull up the context, saving you the need to search again.

---

• 

  buJEN

  —

  Compared to a verbal prediction like "70% will cut interest rates," I believe it is backed by a Brier score and falls into the superforecaster zone. The quantitative team follows this approach. After all, we have compared it on ProphetArena. It ranks first in the overall rankings and in the sports score rankings. It is much more reliable than directly asking GPT-5 to get an uncalibrated probability, and it will not be biased by falsehoods during backtesting.

---

• 

  黎伟丹

  —

  The document says that a single GPU can run v3, and private deployment data does not leave the domain, which is very important for financial customers.

---

• 

  顾龙强

  —

  Will the calibration interval be given in the return? Or is it just an estimate? Knowing when not to believe it in a real offer is more valuable than knowing the probability itself.

---

• 

  Frank_PerezZ

  —

  The idea is very clear to make predictions an independent capability rather than a bonus of a large model.

---

• 

  SOcol

  —

  I tried using it for market making: I pulled the Polymarket real-time market, used the model probability to price each contract, and took the price difference between the market price and the fair value. Thousands of calls were run in one night, the cost was an order of magnitude lower than using the cutting-edge model, and the speed could keep up. The only worry is that when there are too many similar agents, the signals will become homogeneous.

No reviews yet