OPENAI · MODEL RELEASE TRACKER
GPT‑6 Cyber
GPT‑6 Cyber is a cybersecurity-focused variant of GPT‑6 from OpenAI, announced as a model trained to find security vulnerabilities, develop working exploits, and propose fixes while operating within assigned safety boundaries.CURRENT SNAPSHOT4/5 DIMENSIONS WITH DATA
The dimensions that change the decision.
Not established from the available sources.
RELEASE TIMELINE
Published, source-backed release events only.
Alibaba launches Qwen4‑Max: 2.4T-parameter MoE with 1M‑token context and lower costs
Alibaba introduced Qwen4‑Max, a mixture‑of‑experts model with a 2.4 trillion‑parameter configuration that activates about 71 billion parameters per token, a one‑million‑token context window, and a revamped architecture Alibaba says reduces compute on long inputs. The company highlights large efficiency and cost gains (e.g., Artificial Analysis reports $3.27 per Intelligence Index task vs $10.02 for Qwen3.8‑Max), new “Live Delivery” agent transparency features (Qwen‑Live‑Delivery‑Plugin open‑sourced), and mixed external benchmark results versus Opus 5.5, GPT‑6, Kimi K3.1 Pro and others.
OpenAI launches Decisions API (public beta) using gpt-6-luna
OpenAI introduced the Decisions API in public beta, a service that evaluates text, images, or both and returns structured outputs (yes/no probabilities, category picks, or scale ratings). It uses only gpt-6-luna, is roughly ten times faster than the Responses API, charges $0.10 per million input tokens (output tokens free), supports zero-data retention and HIPAA-compliant use in the US and Europe, and comes alongside a simplified paid tiering (Build, Launch, Grow) with $500/$5,000/$200,000 monthly thresholds.
OpenAI rolls out GPT-6 with Intelligent UI in ChatGPT
GPT-6 is rolling out globally inside ChatGPT and introduces an "Intelligent UI" that delivers faster responses along with visuals and interactive experiences you can explore and use directly.
GPT-6 improves prompt caching
GPT-6 introduces improved prompt caching with higher cache hit rates, new diagnostics, explicit breakpoints, and user controls intended to reduce latency and costs. These features target prompt-heavy workflows that can benefit from more efficient reuse of prompt context.
SpaceXAI releases Grok 4.7 — faster, cheaper model for coding and knowledge work
SpaceXAI released Grok 4.7, a new base-model upgrade positioned as its most capable model for coding and knowledge work; it reportedly uses a larger base model, longer RL training focused on long-running tasks, and an entirely new safeguard stack. The company says Grok 4.7 matches the price and speed of Grok 4.6 while improving verification, longer-context handling, document/presentation generation, benchmark performance (CursorBench 4.0, GDPval, AA Briefcase), and safety metrics (LatchBio biosafety 62.4%, HackerBench v0.3 allowing 3.3% risky prompts); it is available in Cursor, Grok Build, the Grok API, third-party harnesses and cloud platforms, priced from $2 per million input tokens and $6 per million output tokens, with an optional faster (2x) variant at double the price.
VERIFIABLE FACTS
Every value stays attached to a source and date.
Announced by OpenAI on 2026-11-05 as GPT‑6 Cyber, a cybersecurity model focused on finding exploitable vulnerabilities and developing fixes.
GPT-6 improves prompt caching with higher cache hit rates, new diagnostics, explicit breakpoints, and controls intended to reduce latency and costs.
Tech reporting indicates GPT‑6 Cyber will be made available to highest-tier users in OpenAI's Daybreak Red program (a program that gives offensive cybersecurity researchers access to powerful frontier models).
GPT-6 Sol is described as designed for demanding, recurring development and operations tasks—implementing features, debugging, refactoring and reviewing code, analyzing data, and completing multistep processes across tools.
On an internal factuality evaluation, OpenAI found that GPT-6 Sol made approximately half as many factual mistakes as GPT-5.6 Sol.
Reported prices: GPT-6 Sol at $2 per million input tokens and $10 per million output tokens; GPT-6 Luna at $0.10 per million input tokens and $0.50 per million output tokens.
OpenAI provides model variants: paying customers receive GPT‑6 Sol while free users receive GPT‑6 Luna.
According to OpenAI, on OSWorld 2.0 GPT-6 delivers results similar to Claude Opus 5 at roughly 80% lower cost; on AutomationBench, OpenAI reported GPT-6 Sol at highest effort beats Claude Opus 5 with estimated cost per task at about 9% of Opus's.
Tested GPT-6 configurations produced a usable, editable 3D scene (as executable Blender code) from a single photo almost every time in the LEGO-Anything/Image-to-Code evaluation.
In the evaluation, GPT-6 configurations took a single image as input and wrote executable Blender code (text/code) to construct the 3D scene.
GPT-6 Astra scored 53.4% on indoor scenes and 39.6% on outdoor scenes on the LEGO-Bench benchmark.
Weaker GPT-6 configurations scored around 15% on the LEGO-Bench benchmark.
The benchmark and results indicate key failure modes include poor self-assessment and limited geometric accuracy; accuracy declines with greater scene complexity and is lower for outdoor scenes.
GPT‑6 produces multimodal outputs that combine text with interactive visuals such as diagrams, charts, forms, tappable buttons, and other elements users can interact with directly.
GPT‑6 can generate interactive tools and mini-apps (for example, a savings calculator with sliders, editable graphs, or a seat-map explorer) that adapt automatically to the user’s request.
The new Decisions API supports only gpt-6-luna so far, priced at $0.10 per million input tokens; output tokens are free (public beta, with general availability planned).
OpenAI says GPT‑6 can respond while still "thinking," which it claims cuts wait times by 44%; in internal tests GPT‑6 scored higher than its predecessor (GPT‑5.6) on difficult web searches using this method.
Not every GPT‑6 answer will include visuals; OpenAI says visuals are added when they add value and users can request fewer visuals if they prefer.
Reported Gwet's AC2 = 0.38 on 118 pediatric answers (as a judge score in the referenced study).
According to OpenAI, GPT‑6 Cyber can find security flaws, build working exploits, and suggest or prepare patches/fixes for vulnerabilities.
OpenAI reports GPT‑6 Cyber scores 48.5% on ExploitGym (compare: GPT‑5.6 Cyber 32.1%, GPT‑6 Astra 42.4%).
OpenAI reports GPT‑6 Cyber found 81 previously unknown severe/critical vulnerabilities in Firefox (patched by Mozilla) and, in Linux kernel testing, reproduced 1,724 vulnerabilities and developed working exploits for 948 of them.
OpenAI reports GPT‑6 Cyber was trained to respect operational guardrails: it avoided boundary violations in 99.93% of cases without special safety instructions and stayed within bounds in 100% of tests with safety instructions enabled; overall out-of-bounds behavior was reported as under 1%, down from ~5% for GPT‑5.6 Cyber.
OpenAI classifies GPT‑6 Cyber as 'Critical' in its Preparedness Framework cybersecurity category.
WHAT CHANGED
Stored passport versions, without reconstructed history.
No public fact values changed.