Weighted log-odds pooling of 5 agent estimates (research 52% (w=1.5); quant 95% (w=1.0); market 45% (w=1.0); sentiment 58% (w=0.5); historian 44% (w=0.7)), shrunk toward the technology base rate, yields 62% — an edge of +17% versus the market at 45%.
Will OpenAI release GPT-5 before the end of the year?
Will OpenAI release a new frontier model this year?
Will Anthropic release a new frontier Claude model this year?
Will Google release a new Gemini flagship model this year?
How much the forecast would move if each signal were removed — leave-one-out over the real pipeline (baked at snapshot time).
The flagship line may be rebranded, resolving the question NO on a technicality.
Frontier model generations have shipped roughly every 18-24 months.
Multi-gigawatt training clusters are under construction.
Longer internal evaluation cycles have delayed recent launches.
Pretraining team postings reference next-generation model work.
Every forecast is an argument. Seven agents deliberate; the skeptic attacks; synthesis pools.
The evidence base (3 supporting, 2 opposing signals) supports the event. The single strongest signal: Frontier model generations have shipped roughly every 18-24 months.
Of 4 comparable resolved events, 4 resolved YES (similarity-weighted rate 100%). Monte Carlo over the posterior gives a 90% credible interval of 80%-100%, and 100% of simulations land above the market's 45%.
Prediction markets price YES at 45¢ on $900,000 volume with medium liquidity — an implied 45% probability. A reasonable but not decisive prior.
Public sentiment runs 65% positive with increasing momentum. Sentiment is a weak standalone predictor, so it enters the pool at low weight.
Historically, technology questions of this shape resolve YES about 44% of the time. With a 500-day horizon, the base rate pulls the current market read only marginally to 44%.
The pipeline is above the market (64% vs 45%) — before trusting that edge, weigh: The flagship line may be rebranded, resolving the question NO on a technicality. Longer internal evaluation cycles have delayed recent launches.
Weighted log-odds pooling of 5 agent estimates (research 52% (w=1.5); quant 95% (w=1.0); market 45% (w=1.0); sentiment 58% (w=0.5); historian 44% (w=0.7)), shrunk toward the technology base rate, yields 62% — an edge of +17% versus the market at 45%.
Signals ingested for this question, by impact.
Frontier model generations have shipped roughly every 18-24 months.
Release cadence · impact 0.7 · Nov 7
Multi-gigawatt training clusters are under construction.
Compute buildout · impact 0.6 · Nov 7
The flagship line may be rebranded, resolving the question NO on a technicality.
Naming risk · impact 0.5 · Oct 24
Pretraining team postings reference next-generation model work.
Hiring signals · impact 0.4 · Oct 19
Longer internal evaluation cycles have delayed recent launches.
Safety reviews · impact 0.4 · Oct 21