GPT-6 release and Astra math/capability claims
What happened
On September 22, Simon Willison reported that OpenAI had released GPT-6 Sol at half the price of GPT-5.6 Sol.1 Willison also reported that GPT-6 Luna cost half as much as GPT-5.6 Luna.2 Latent Space reported on September 25 that the family also included Astra, which led benchmarks.3
Math claims and disputes
On September 22, Latent Space said reports claimed an internal OpenAI model had resolved over 100 open math problems.4 Latent Space described independent verification as limited.4 OpenAI said on September 23 that its model had solved the Navier-Stokes problem, one of the Millennium Prize Problems.5 MIT Technology Review reported on September 28 that critics questioned whether the solution addresses what mathematicians care about.6
Latent Space reported on October 2 that OpenAI used 10,000 agents over 88 hours, at an estimated cost of about $40 million.7 The Verge reported on October 5 that researchers alleged scooping, spying and violations of academic norms by OpenAI.8 Mathematicians feared a chilling effect on the field, according to The Verge.8
New models and ChatGPT changes
OpenAI released GPT-6.1 Sol on September 29, saying it matched GPT-6 Astra on coding, computer use and professional tasks at one-fifth the cost.9 TechCrunch described the performance as near that of GPT-6 Astra.10 On October 7, OpenAI said GPT-6 with Intelligent UI was rolling out to paid ChatGPT tiers, with Free and Go following a day later.11 The Verge reported that ChatGPT now answers with interactive diagrams, charts, forms and buttons.12
Test results for Astra
drivingbench.com reported on September 23 that GPT-6 Astra ranked first in a test of AI models driving a Toyota Corolla on a cone course.13 Latent Space reported on October 16 that GPT-6 Astra scored 49 of 50 on TheoPhysEval, and that the authors confirmed several reference solutions were wrong.14 Import AI said on October 19 that the score came after experts corrected their own errors that the model flagged.15
Latest reported
OpenAI announced GPT-6.1 Terra on October 26, describing it as a fast, lower-cost model with stronger visual reasoning and computer use.16 Transformer reported on October 30 that Terra's API prices were half those of GPT-6.1 Sol.17 Latent Space reported the same day that Terra ranked eighth on Agent Arena, at a median cost of $0.62 per task.18
Sources for the overview
- Simon Willison: OpenAI releases GPT-6 Sol
- Simon Willison: OpenAI releases GPT-6 Luna
- Latent Space: GPT-6 family releases
- Latent Space: OpenAI internal model math claims
- OpenAI: OpenAI model solves Navier-Stokes problem
- MIT Technology Review: OpenAI's math solution questioned
- Latent Space: OpenAI Navier-Stokes breakthrough via multi-agent systems
- The Verge: Academic integrity controversy emerges
- OpenAI: OpenAI releases GPT-6.1 Sol
- TechCrunch: OpenAI launches GPT-6.1 Sol
- OpenAI: OpenAI rolls out GPT-6 with Intelligent UI in ChatGPT
- The Verge: OpenAI launches Intelligent UI in ChatGPT
- drivingbench.com: DrivingBench tests frontier models on real-world driving
- Latent Space: TheoPhysEval results for GPT-6 Astra
- Import AI: TheoPhysEval theoretical physics benchmark
- OpenAI: OpenAI releases GPT-6.1 Terra
- Transformer: OpenAI launches GPT-6.1 Terra
- Latent Space: GPT-6.1 Terra and Sol Agent Arena costs
The stories
- · OpenAI GPT-6 Sol and Luna release
- · OpenAI model math claims (Astra)
- · ChatGPT Voice update with GPT-6
- · GPT-6 Astra has gained the ability to drive a car
- · GPT 6.1 Sol: Near-Astra intelligence for a fifth of the price
- · ChatGPT Intelligent UI with GPT-6
- · TheoPhysEval results for GPT-6 Astra
- · OpenAI GPT-6.1 Terra release
- · GPT-6 Astra covert safety underperformance
- · OpenAI releases GPT-6.2 Astra and Sol