GPT-4
TrialPlatforms
OpenAI's large language model used as a benchmark reference point.
Why it's here
Placed in Trial: 24 article(s) of evidence from 6 source(s), led by research-stage coverage, with 4 in the last 30 days. Confidence 100%.
Evidence (18)
- 3Simon Willison·10/9/2026framework_updatettok 1.0 updates default tokenizer to GPT-5/GPT-6
Simon Willison released ttok 1.0 after discovering the tool still defaulted to the GPT-4 tokenizer. The update switches the default to GPT-5/GPT-6, based on evidence that GPT-6 likely shares the same tokenizer family as GPT-5, although OpenAI has not officially confirmed that yet.
- 4Simon Willison·10/4/2026researchQwen3.8-27B Tested on Word-Return Addition
Simon Willison reports an experiment running Qwen3.8-27B locally to compute sums while returning the answers in words, comparing results with and without reasoning enabled. The model was highly accurate overall, especially with reasoning on, achieving 167 correct results out of 169 attempts in the one-shot run.
- 6OpenAI Blog·9/23/2026product_launchRingg resolves up to 65% of customer calls with OpenAI agents
Ringg says its AI agents, built with GPT-5.6, can resolve up to 65% of customer calls across voice, chat, WhatsApp, and web. The company also claims the system delivers multilingual support at about 90% lower cost than GPT-4.1-based operations.
- 6Hacker News·9/14/2026framework_updateAI Leaders’ Doom Warnings Draw Criticism
The article criticizes repeated public warnings from AI leaders about existential risk, arguing that these statements are often strategically useful while crowding out discussion of present-day harms. It cites examples involving Anthropic, OpenAI, and the AI Safety letter, along with the debate over regulation and labor impacts.
- 7Hacker News·9/2/2026researchStudy says many Perplexity citations do not support cited figures
An audit of Perplexity’s search models found that 34.7% of citation markers attached to figure-based claims pointed to pages that either would not open or did not contain any matching number. Measured per claim instead of per citation, 14.4% of claims failed, with gated pages and unreadable sources making up much of the gap rather than dead links.
- 5Hacker News·8/26/2026researchRAG Can Start With Simple Full-Text Search
The article argues that many retrieval-augmented generation systems are over-engineered and should begin with simpler approaches such as full-text search before adding embeddings or reranking. It outlines several retrieval recipes, including full-text search, agentic query rewriting, and hybrid search, with the choice driven by freshness, corpus churn, query patterns, scale, and team expertise.
- 6The New Stack·8/17/2026product_launchClaude can manage ElevenLabs voice agents through a hosted MCP connector
ElevenLabs launched a hosted MCP connector that gives Claude read/write access to existing ElevenAgents, letting developers inspect, edit, test, and even delete production voice agents from chat. The connector uses OAuth instead of local API keys and adds access controls plus testing tools to reduce the risk of unsafe changes.
- 4Simon Willison·8/7/2026product_launchCodex Builds a Better Raccoon Heist Game with GPT-5.6 Sol Ultra
Simon Willison compared the same one-shot game prompt across different coding agents and found Codex Desktop running GPT-5.6 Sol Ultra produced a more compelling Raccoon Heist game than a previous Claude Fable 5 attempt. The post also notes a visible bug in the generated game, which was later fixed with a follow-up prompt, and includes the full Codex transcript plus repository assets generated during the session.
- 4Simon Willison·8/5/2026product_launchClaude Fable 5 builds a raccoon heist game
Simon Willison describes using Claude Fable 5 in Claude Code for web to generate a playable browser game based on an old raccoon-heist concept, with GitHub Pages used for rapid preview and iteration. The workflow combined the Claude agent with OpenAI's image generation API to fill in textures and visuals while the project was built and committed from mobile.
- 7OpenAI Blog·8/3/2026product_launchOpenAI Built a Low-Latency Realtime Voice AI System in Six Months
OpenAI says GPT-Live enables continuous voice interaction with AI through a turnless speech model and a low-latency architecture. The system is designed to make conversations faster and more natural by reducing pauses and handoffs between turns. The company built the realtime stack in six months.
- 5Hugging Face Blog·7/30/2026researchGPU Utilization Becomes the New AI Constraint
The article argues that, as aviation is judged by aircraft utilization, enterprise AI will increasingly be judged by GPU utilization. It says the industry has shifted from a model-quality bottleneck to a compute-capacity bottleneck, with expensive GPUs generating value only when actively used.
- 6Simon Willison·7/20/2026researchAltman Email Reveals OpenAI’s Local GPT-3-Scale Model Plan
An email from Sam Altman to OpenAI’s board in October 2022 said the company wanted to soon release a language model with roughly GPT-3-level capability that could run locally on consumer hardware. The email said the move was partly intended to discourage rivals from releasing similar models and to make it harder for new efforts to attract funding.
- 7Hacker News·7/17/2026researchOpen source AI gains ground, but frontier gaps remain
A Hacker News-linked CTO letter argues that open-weight AI has become competitive in many production workloads, with lower inference costs and broad adoption in tokens routed through open models. The report says closed models still lead on reasoning and multimodality, while open models are strong in coding, instruction-following, and general use cases.
- 7OpenAI Blog·7/15/2026researchGPT-Red: Automated Red Teaming for Safer AI
OpenAI introduces GPT-Red, an automated red teaming system that uses self-play to find weaknesses in AI systems. The approach is designed to improve robustness against prompt injection while supporting broader safety and alignment efforts.
- 3Simon Willison·7/14/2026open_sourceSimon Willison’s custom Codex Desktop pet, Pedalican
Simon Willison describes creating a custom Codex Desktop pet: a pelican riding a bicycle that animates on his desktop and reports on Codex tasks. He also details how GPT-5.6 Sol, along with multiple rounds of gpt-image-2, generated the sprite assets and notes that the implementation relies on open-source skills from OpenAI.
- 7Simon Willison·7/8/2026product_launchOpenAI upgrades ChatGPT voice mode with GPT-Live
OpenAI has updated ChatGPT voice mode to use a new model called GPT-Live, replacing the earlier GPT-4o-era voice system. The new mode can keep the conversation flowing while delegating harder tasks such as web search and deeper reasoning to GPT-5.5 in the background. The change is rolling out in the iPhone app and is intended to make voice conversations more capable and responsive.
- 4Simon Willison·7/2/2026researchUsing DSPy to improve Datasette Agent SQL prompts
Simon Willison describes using DSPy to evaluate and refine the system prompts behind Datasette Agent's read-only SQL feature. An initial pass found that prompt wording around schema guidance could lead to column-name guessing and error-retry loops, suggesting prompt changes or adding column names to the schema listing.
- 6OpenAI Blog·6/3/2026model_releaseOpenAI Expands GPT-Rosalind for Life Sciences
OpenAI says GPT-Rosalind now has improved biological reasoning, medicinal chemistry knowledge, genomics analysis, and support for experimental workflows. The update is aimed at helping life sciences researchers with more capable AI assistance across research tasks.